• Pre-Market Highlights

• AI & Semiconductor Value Chain

• GPU / CPU / ASIC

• AI Cloud & Data Center Operators

• HBM / DRAM / NAND / SSD / HDD

• Foundry & Key Components

• Optical Communications & Photonics

• Space & Satellite

• Internet / Platforms

• Software / SaaS

• Consumer Electronics / Intelligent Vehicles

Overview

404K SEMI-AI | 2026-09-01

Pre-Market Highlights

AI capex continues to expand, but the critical delivery bottleneck has shifted from raw GPU availability to whether an entire AI factory can be deployed on schedule. Nvidia provided robust growth guidance as Vera Rubin enters multi-cloud rack deployments; meanwhile, optical interconnects, PCBs, liquid cooling, power infrastructure, and regulatory permitting increasingly dictate the timing of revenue recognition.

Enterprise demand is broadening as paid AI adoption accelerates across manufacturing, construction, and healthcare, though switching costs between foundation model providers remain low. Software vendors are rapidly transitioning AI monetization from per-seat add-ons to consumption-based pricing; while this expands total revenue potential, customer retention, gross margins, and underlying usage require close validation.

Near-term risks center on execution rather than underlying demand. A high proportion of flagship data center projects worldwide face delays, and hyperscalers' massive lease commitments remain contingent on power availability, permitting, and physical construction. As prices continue to rise across memory, wafers, IC substrates, and optical components, hardware supply chain revenue visibility is improving, but margins may not necessarily expand in tandem.

AI & Semiconductor Value Chain

GPU / CPU / ASIC

• AI Factory Platforms: Delivery bottlenecks have shifted from discrete GPU and HBM silicon to full-rack validation. As Vera Rubin and Spectrum-6 (SPECTRUM6) enter multi-cloud deployments, customers must concurrently secure switches, optical interconnects, PCBs, liquid cooling, and power delivery; bare chip shipments no longer equate to recognizable system revenue. Key watchpoints include full-rack yields, joint qualification, and power energization timelines.

• Nvidia
  1) Management guided next fiscal year sales growth to 70%, well above consensus of 45%; FY27 Q2 revenue reached $96.2 billion, with Data Center revenue at $89.0 billion (92% of total).
  2) Vera Rubin's revenue opportunity per GW expanded to $40 billion, up from $18 billion for Hopper and $25 billion for Blackwell.
  3) Baird maintained an Outperform rating with a $500 price target, projecting that inference architectures will sustain long-term growth.

"Nvidia remains unequivocally supply-constrained. That is the underlying message behind providing guidance not just for the next quarter, but for the entire upcoming fiscal year."

• Broadcom: OpenAI and Broadcom completed the initial RTL for the Jalapeño inference ASIC in approximately 9 months, optimizing the chip for latency, performance-per-watt, and KV cache memory slicing. While custom ASICs shorten design cycles, software complexity and model lock-in still limit their displacement of general-purpose GPUs. Key watchpoints include mass-production yields, software enablement, and actual procurement volumes.

• AMD: Next-generation thermal management is extending from cold plates down to the package interface. Reports indicate AMD is evaluating graphene thermal interface materials (TIM) between the package and the cold plate or lid to enhance GPU thermal conductivity; the critical test is whether commercial-grade materials can pass reliability and high-volume manufacturing qualification. Key watchpoints include sample validation, lifetime reliability data, rack-level integration, and supplier qualification progress.

• Cerebras: Partnering with Compute Nordic Finland to construct a 165 MW AI data center, with initial 50 MW construction underway and plans to expand capacity first to 80 MW and ultimately to 165 MW. Backed by a 7-year contract with an estimated investment of €1.0B to €1.7B, delivery timelines hinge on physical construction and grid energization.

AI Cloud & Data Center Operators

• Data Center Projects: Supply constraints extend well beyond silicon. Among tracked flagship data centers under construction globally, roughly 40% have been delayed or canceled over the past year, 15% are progressing slowly, and only 30% are tracking on schedule. Given the lack of full methodology disclosure, this dataset serves primarily as an indicator of execution risk rather than a broad industry benchmark, requiring cross-validation against more comprehensive project samples.

"Among the most high-profile data centers under construction globally, roughly 40% have been delayed or canceled over the past year, 15% are progressing relatively slowly, and only 30% are 'on schedule.'"

• CoreWeave: Vera Rubin racks have entered CoreWeave's deployment roadmap, signaling that neo-cloud providers are shifting from procuring bare GPUs to integrating full-rack systems. Revenue recognition remains contingent on networking, liquid cooling, power infrastructure, and joint customer validation—hardware delivery does not instantly translate into operational compute. Key watchpoints include rack energization rates, utilization levels, and customer acceptance.

• Nebius: Secured zoning approvals and power supply contracts for the entire 1.2 GW capacity across its Pennsylvania campus, following prior approval of the initial 600 MW. The company reaffirmed its deployment schedule: 260 MW by October 2027, 660 MW by January 2029, and over 1 GW by November 2029; regulatory approval remains distinct from actual grid energization and commercial operation.

• Lambda & Hut 8: Anthropic signed a $35 billion compute agreement with Lambda, with Hut 8 constructing a 700 MW campus in Texas while Nvidia anchors the underlying lease and supplies GPUs. By chaining silicon, cloud services, and real estate leases into a single structure, execution risk concentrates heavily on construction timelines and counterparty performance.

• Huaqin Technology: Supernode products completed canary deployments with two major tier-1 clients, with monthly run-rates stabilizing above 100 racks in late July. While mass-production milestones are visible, racks require continued co-validation with domestic GPUs, networking, and thermal solutions. Key watchpoints include monthly output sustainability, qualification conversion into scaled revenue, manufacturing yields, cash collection, and lead times.

• Honor Electronic: H1 data center power supply revenue reached RMB 1.136 billion (+18.1% YoY), with high-power server power supplies up 41.4% YoY to RMB 936 million (82.4% of segment revenue). Initial small-batch shipments for its first Google project are slated for Q4, while high-voltage direct current (HVDC) solutions remain in the sampling phase and should not be prematurely booked as firm orders. Key watchpoints include design wins, volume ramp, and revenue recognition.

HBM / DRAM / NAND / SSD / HDD

• Micron: Management highlighted acute memory shortages across all segments, noting limited visibility into when supply will balance demand. DDR4 spot prices carry a 43% premium over contract prices, compared to 16% for DDR5, reflecting severe legacy capacity crowding as HBM and AI servers consume standard DRAM wafer allocations. Key watchpoints include contract pricing trends, customer inventory levels, wafer capacity allocation, and incremental supply additions.

"We have entered an environment of severe memory shortages across all market segments. Despite our best efforts on the supply side, we cannot currently foresee when supply will catch up with demand."

• Samsung Electronics: HBM5 targets 2x the performance of HBM4E, a 20% gain in performance-per-watt, and a 20% reduction in thermal resistance, with even more aggressive targets for subsequent zHBM. Meanwhile, zNAND-O is slated to sample in 2028 featuring 10x DRAM's bit density and 7x NAND's read bandwidth alongside improved power efficiency; both technologies remain preliminary roadmap targets rather than committed production specs.

• SK Hynix: Q2 wafer procurement jumped 104.0% QoQ to KRW 670.6 billion against just 10.5% output growth, driving raw material inventory up 61.9% in H1. Procurement significantly outpaced production, likely reflecting advance locking of wafer supply ahead of price hikes and preparation for upside HBM orders; realization remains contingent on downstream conversion. Key watchpoints include inventory turnover and HBM4E qualification.

• SanDisk: Positioned High-Bandwidth Flash (HBF) as a new memory tier bridging HBM and NAND, with Grade 3 targeting 512GB capacity and 3.072 TB/s bandwidth—providing 8x to 16x the capacity of HBM at comparable cost. While ideally suited for MoE weights and massive KV caches, inadequate software-level data placement remains the primary hurdle. Key watchpoints include controller readiness, runtime software support, and customer validation.

Foundry & Key Components

• TSMC
  1) Supply chain checks indicate that sold-out AI capacity is driving price hike discussions of 10% to 15% across process nodes, with N2, N3, and N5 potentially facing an additional 10% hike by late 2027; margin realization hinges on broad customer acceptance.
  2) Management projects 2029 AI/HPC system computing power to reach 50x 2024 levels, stressing that HBM, packaging, substrates, testing, and optical interconnects must scale in tandem.

"The critical signal is that TSMC no longer frames silicon photonics scaling merely as a foundry or CoWoS capacity issue."

• Samsung Electro-Mechanics: Disclosed a 1-year MLCC supply contract for 2027 valued at KRW 1.0722 trillion (9.5% of annual revenue). Combined with two prior high-reliability AI rack capacitor orders totaling KRW 750 billion, rising rack power densities and capacitor content are translating into tangible backlog, though client identities and per-system dollar content remain undisclosed. Key watchpoints include delivery execution and realized margins.

• Victory Giant Technology: H1 revenue reached RMB 11.6 billion (+28.8% YoY) with net profit attributable to the parent at approximately RMB 2.8 billion (+33.3% YoY); Q2 gross margin fell 6.7 percentage points YoY to 32.1%. Capex rose 228% YoY to approximately RMB 7.1 billion, as concurrent capacity expansion and raw material cost inflation compress the profit elasticity of incremental revenue.

• Envicool: H1 revenue rose 17.2% YoY to RMB 3.017 billion, while net profit attributable to the parent fell 14.3% YoY to RMB 185 million; data center room-level and cabinet-level thermal management revenues grew 25.8% and 15.3%, respectively. Despite growing liquid cooling demand, domestic price competition, prolonged qualification cycles, and expansion costs have caused profit realization to lag revenue growth. Key watchpoints include gross margin trajectory and operating cash flows.

• Hengtong Optic-Electric: H1 optical communications revenue surged approximately 130% YoY, with gross profit multiplying 5x from the prior-year period; Q2 optical fiber ASP exceeded RMB 100, with A1 and A2 shipments each accounting for roughly 40%. The 800-metric-ton Phase II optical preform facility in Inner Mongolia is scheduled for commissioning by late 2026; capacity ramp and North American customer qualification represent key growth drivers. Key watchpoints include ASP sustainability and utilization rates.

Read more

• I. The Most Important Shift: Customer Demand Now Far Exceeds Deliverable Supply

• II. Why AI Servers Consume So Much Memory

• III. Why the Shortfall Persists Despite Rapid Supply Growth

• IV. Prices Will Rise, but Shortages Will Not Translate One-for-One into Profits

• V. Both DRAM and NAND Are Tight, but Their Shortage Intensity and Earnings Leverage Differ

• VI. Which Companies Can Most Readily Convert Tightness into Cash Flow?

• VII. The Three Most Likely Sources of Overestimation

• VIII. Six Numbers to Track Next

本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读

AI servers are pushing memory chips beyond a conventional price-upcycle into a structural shortage lasting through 2027. UBS expects demand to remain well above deliverable supply, although long-term agreements and inventory reductions will limit the pace of price realization.

I. The Most Important Shift: Customer Demand Now Far Exceeds Deliverable Supply

The most valuable insight in this UBS monthly report is its separation of “demand” into two fundamentally different figures. The first is unconstrained demand—what customers request without regard to supply limitations. The second is constrained demand—the volume memory manufacturers can actually deliver given existing capacity and product allocation. The gap between the two is the starting point for understanding the 2027 memory market.

The current overall DRAM demand-fulfillment rate is only about 60%. In other words, for every 100 units customers request, manufacturers can currently supply only around 60 units. Server DDR bit demand is approaching 80% growth in 2027, while unconstrained bit demand for server SSDs and storage SSDs could exceed 100%. UBS estimates that deliverable constrained demand may grow by only about 40%. The gap between customer requests and actual delivery capacity continues to widen.

This gap should not be treated directly as orders or revenue. Customers may submit duplicate requests, secure capacity in advance, or cancel part of their purchases after prices change. Volumes covered by long-term agreements must also be distinguished from incremental customer requests. The report’s unconstrained-demand figures are better viewed as a stress test: they show how much customers want, not how much will necessarily transact.

Even after applying a discount, a DRAM fulfillment rate of about 60% remains extremely low. Memory manufacturers are no longer deciding whether to expand capacity, but which customers should receive their limited wafer and packaging capacity first. AI servers, HBM, server DDR, and enterprise SSDs offer higher margins, and customers are willing to pay for delivery certainty, naturally moving these products to the front of the queue. Memory products for conventional PCs and smartphones must accept slower supply growth and higher procurement costs.

UBS has therefore reduced its confidence in end-demand resilience. It previously expected both PC and smartphone unit sales to grow 2% in 2027, but now sees downside risk to both forecasts. Demand may not disappear, but higher memory costs and insufficient supply could force device manufacturers to delay product launches, reduce configurations, or offer fewer promotional models. As shortages spread from servers to consumer electronics, the industry’s beneficiaries and pressured players will increasingly diverge.

II. Why AI Servers Consume So Much Memory

AI servers drive memory demand through both a rising device count and greater memory content per device. Looking only at GPU or server shipments would materially underestimate demand. UBS multiplies the GPU and ASIC volumes of NVIDIA, AMD, Google, Amazon, Microsoft, Meta, and other vendors by the HBM capacity per accelerator and DDR capacity per server, producing bit-demand estimates by customer and product.

The model projects total GPU and accelerator volumes to increase from 16.933mn units in 2026 to 28.493mn units in 2027, an increase of about 68%. Over the same period, total HBM demand rises from 32.17bn Gb to 61.52bn Gb, up 91%. HBM demand is growing faster than accelerator volumes, indicating that memory capacity per chip continues to increase. As compute chips become more powerful, they must process more model parameters, context, and intermediate results, while the cost of processors waiting for data also rises. Vendors therefore need larger, faster memory configurations to maintain throughput.

The increase in memory content per system is even more pronounced in UBS’s server DRAM model. Average server DRAM capacity is expected to rise 57%, from 927GB in 2026 to 1,455GB in 2027. Average DRAM capacity in AI servers increases from 8,961GB to 14,754GB. With both system volumes and per-system capacity rising sharply, total server DRAM demand reaches 324,124mn Gb in 2027, up 83.1% year over year.

These figures explain why servers are reshaping the entire DRAM market. Server DDR is expected to represent 48% of DRAM bit demand in 2027, with server HBM accounting for another 11%, for a combined 59%. Smartphones, PCs, televisions, and game consoles remain large markets, but they can no longer determine industry pricing on their own. Server-related demand has become both the largest source of incremental growth and the highest-priority category in manufacturers’ capacity-allocation decisions.

HBM’s customer mix is also changing. NVIDIA remains the largest customer, but UBS expects its demand share to decline from 58% in 2026 to 41% in 2027. Google rises to 31% and AMD to 12%. NVIDIA’s lower share does not imply declining demand; it reflects faster investment growth in Google’s internally developed ASICs, AMD accelerators, and other cloud platforms. The HBM market is gradually shifting from dominance by a single major customer toward simultaneous expansion across multiple computing platforms.

Changes in product specifications can also be misinterpreted. The report expects NVIDIA’s VR300 initially to use HBM4 8-Hi rather than HBM4E 12-Hi, while a previously considered 4-Hi configuration is constrained by the development timeline. A specification reduction for one product changes that model’s capacity but does not automatically reduce total demand. Higher accelerator shipments, richer configurations among other customers, and transitions to next-generation products can offset lower capacity in a single model. UBS maintains its forecast for 91% HBM demand growth in 2027.

III. Why the Shortfall Persists Despite Rapid Supply Growth

Suppliers have not stopped expanding capacity. UBS expects total HBM supply to increase from 36.616bn 1Gb units in 2026 to 53.728bn 1Gb units in 2027, growth of about 47%. Total DRAM bit supply is also expected to grow 23.3% in 2027. The problem is that demand is growing faster, while products also compete for the same wafers, equipment, engineering resources, and advanced-packaging capacity.

Expanding HBM capacity crowds out conventional DDR. The report expects monthly HBM wafer capacity to rise from 390k wpm in 2025 to 690k wpm in 2027. Memory manufacturers cannot expand total wafer input without limit at the same pace, and process transitions take time. Allocating more wafers to HBM leaves less incremental capacity for conventional DDR. With server DDR demand also approaching 80% growth, supply pressure is emerging simultaneously in both high-bandwidth memory and conventional server memory.

Technology upgrades cannot resolve the shortage immediately. DRAM and NAND can increase output per wafer through more advanced process nodes, additional layers, and higher bit density, but yield ramp-ups delay the release of effective supply. HBM also requires stacking, testing, and advanced packaging; a shortfall at any stage constrains final deliveries. Financially, capital expenditure first appears as equipment and depreciation, while qualified products convert into revenue and cash flow only after volume production ramps. Higher prices strengthen the incentive to expand, but they cannot bring next quarter’s wafers into the present.

Read more

• I. Behind RMB954 Million in Revenue, Zhipu AI Has Become a Different Business

• II. Cloud Gross Margin Turns Positive, but the Revenue Mix Weighs on Consolidated Margin

• III. No Inflection in the Income Statement Yet, but Loss Margins Have Narrowed Sharply

• IV. US$1.6 Billion of ARR Is Impressive—but Demands the Strictest Definition

• V. Unit Token Costs Fell 80%, but Total Compute Spending Is Still Rising

• VI. The HK$31.375 Billion Placement Shifts the Constraint from “Does It Have the Money?” to “Can the Money Generate Gross Profit?”

• VII. Coding Is the Proven Entry Point, but Co-worker Determines the Long-Term Revenue Ceiling

• VIII. Valuation Will Shift from the Model Narrative to Three Forms of Financial Evidence

• IX. Eight Metrics to Watch in the Next Earnings Report

本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读

Zhipu AI has moved beyond proving that demand exists, but has yet to demonstrate that revenue can cover R&D; and compute costs. API revenue surged 27-fold and cloud gross margin turned positive. US$1.6 billion in ARR suggests growth is still accelerating, but this remains a management-defined annualized run-rate figure.

I. Behind RMB954 Million in Revenue, Zhipu AI Has Become a Different Business

The biggest change in the first half of 2026 was Zhipu AI’s shift from on-premise projects to cloud usage. Revenue reached RMB954 million, up 399.7% year over year. Open-platform and API revenue rose 2735.7% to RMB825 million, accounting for 86.5% of total revenue, versus just 15.2% a year earlier.

These figures show that Zhipu AI is no longer primarily dependent on one-off large-model project deliveries. Customers are increasingly paying for usage, subscriptions, and recurring workloads. Cloud revenue is no longer merely a proxy for developer activity; it is now the company’s primary reported revenue stream.

The company’s official segment reporting confirms the shift. Cloud-deployment revenue was RMB825 million, while on-premise deployment revenue was RMB129 million. The latter fell 20.5% year over year, with its revenue contribution dropping from 84.8% to 13.5%. Cloud growth more than offset the contraction in on-premise business and lifted total revenue to nearly five times the prior-year level.

By product line, enterprise-agent revenue increased 304.4% to RMB55.56 million; enterprise general-purpose large-model revenue declined 54.6% to RMB67.04 million; and technical services and other revenue totaled RMB6.12 million. Open-platform and API revenue exactly matched cloud-deployment revenue, while the other three product lines comprised on-premise revenue.

The growth in enterprise agents and contraction in general-purpose large models also point to a change in purchasing behavior. Traditional on-premise projects were closer to “delivering a model to the customer,” making revenue highly dependent on individual project acceptance. Agents embed models into enterprise workflows, so customers are paying for task execution, system integration, and continuous optimization.

This transition remains at an early stage. Enterprise-agent revenue was only RMB55.56 million, or 5.8% of the total—too small to support a second growth engine on its own. Its significance lies in showing that enterprise customers are moving beyond model-capability testing toward workflow deployment. If agent revenue sustains rapid growth, the on-premise business could evolve from general-model delivery toward systems more closely tied to business outcomes.

The 54.6% decline in enterprise general-purpose large-model revenue may also reflect deliberate reprioritization. As the company shifted resources toward cloud APIs, Coding, and Agent offerings, on-premise general models became less important. If customers merely migrated existing on-premise deployments to the cloud, this would represent a change in revenue format. Only a reduction in existing projects without migration to the platform would indicate genuine demand loss. The announcement did not disclose customer migration data, so the two cannot yet be distinguished.

This mix shift is critical to the investment case. On-premise projects typically carry larger contract values, longer delivery cycles, and heavier customization requirements. API and subscription revenue can generate recurring usage, but places greater demands on inference costs, platform reliability, and customer retention. Whether Zhipu AI’s revenue quality improves will depend on whether usage persists—and how much gross profit the company retains from each yuan of usage revenue.

II. Cloud Gross Margin Turns Positive, but the Revenue Mix Weighs on Consolidated Margin

Cloud unit economics have clearly improved. In the first half of 2026, cloud-deployment gross margin reached 24.6%, versus -0.4% a year earlier. Cloud gross profit rose to RMB203 million, accounting for the vast majority of companywide gross profit. The fact that rapid scaling no longer produces negative gross profit is the most important positive signal in these results.

However, consolidated gross margin fell from 50.0% to 26.4% because of the revenue mix. On-premise deployment accounted for 84.8% of revenue a year earlier and carried a 59.1% gross margin. Its revenue contribution fell to 13.5% in the current period, while its gross margin declined to 37.9%. The contraction of the higher-margin on-premise business and rapid expansion of the lower-margin cloud business naturally diluted the consolidated margin.

This does not diminish the significance of cloud margin turning positive, but the two issues should be assessed separately. The first is whether the API business can generate profit; a 24.6% gross margin provides a provisional affirmative answer. The second is whether the company itself can become profitable; current gross profit remains far too small to cover R&D;, sales, administrative, and financing costs.

Product-line margins show a similar divergence. Open-platform and API gross margin matched cloud deployment at 24.6%; enterprise agents generated a gross margin of approximately 35.6%; enterprise general-purpose large models, 42.3%; and technical services and other, just 9.4%. Gross margins for all three latter categories declined year over year.

Cloud gross margin still needs to rise further. Greater inference scale can spread platform and engineering costs, while model optimization can reduce the compute required for a given task. Conversely, price competition, discounted traffic, and long-context usage for complex tasks could push margins down again. The 24.6% margin shows that the business model is gaining a foothold, but does not yet demonstrate that it can consistently reach mature software-company profitability.

III. No Inflection in the Income Statement Yet, but Loss Margins Have Narrowed Sharply

Zhipu AI generated only RMB252 million in gross profit in the first half, against RMB2.131 billion in R&D; expenses. R&D; spending was approximately 8.5 times gross profit, meaning gross profit covered only about 11.8% of R&D; investment. These two figures alone explain why the company continues to report substantial losses.

R&D; expenses rose 33.6% year over year—well below revenue growth—but increased by RMB537 million in absolute terms. Cost of sales surged 635.4% to RMB702 million, outpacing revenue growth. Operating loss widened from RMB1.899 billion to RMB2.147 billion, indicating that higher gross profit has yet to offset the expansion in R&D; and other costs.

The rise in R&D; spending was not simply the result of headcount growth. Full-time employees totaled 981 at period-end, versus 883 a year earlier—an increase of approximately 11.1%, well below the 33.6% growth in R&D; expenses. Investment intensity is increasing across training compute, data environments, infrastructure engineering, and senior talent.

General and administrative expenses fell 44.2% year over year, while share-based compensation declined from RMB159 million to RMB85.76 million, suggesting that listing-related and organizational overhead is receding. R&D; and compute remain the main sources of operating leverage, while back-office costs are showing some discipline. Rising financing costs and financial-asset impairments offset part of these efficiency gains.

Lower share-based compensation also makes the reported loss appear to improve more quickly. Adjusted net loss adds back share-based compensation and is therefore more useful for assessing the underlying operating trend. Even so, it remains sensitive to the timing of R&D;, compute procurement, and revenue seasonality, and should not be treated as a stable forward loss level.

Read more

• The Real Signal from Q2: Headline Profit Grew, but Underlying Operations Remained Weak

• The Revenue Mix Is Changing, with Margin Improvement Preceding Revenue Growth

• Among the Four Growth Drivers, Foldables and Liquid Cooling Have the Clearest Timelines

• Why Citi Cut Revenue Forecasts but Raised Medium-Term Earnings

• Different A-Share and H-Share Ratings Reflect Valuation, Not Fundamentals

• The Strongest Counterargument: Recovery Requires Multiple Projects to Deliver on Schedule

• Five Sets of Numbers to Watch Next—The Fourth Quarter Is the Critical Window

本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读

Lingyi iTech has a foundation for a fourth-quarter earnings recovery, driven primarily by a better product mix and new-business ramp-ups. However, the decline in second-quarter recurring profit means this thesis still needs to be validated through shipments, expense ratios, and gross margins from September through December.

The Real Signal from Q2: Headline Profit Grew, but Underlying Operations Remained Weak

The most important second-quarter signal was the combination of weaker-than-expected revenue and a better-than-expected gross margin. Lingyi iTech reported revenue of RMB12.506 billion, up 3% year on year but down 1% quarter on quarter. Revenue was 12% below Citi’s forecast and 8% below consensus. Demand timing, foreign-exchange movements, and new-product ramp-ups have yet to lift revenue to the level previously anticipated by the market.

Gross margin sent the opposite signal. It reached 16.9% in the second quarter, up 1.9 percentage points year on year and 0.4 percentage points quarter on quarter, and 1.5 percentage points above Citi’s forecast. The company has scaled back its low-margin imaging and display Top Module business and redirected resources toward thermal management, batteries, and power products. This reallocation has already improved profitability per unit of revenue.

Reported net profit alone does not reflect operating quality. Second-quarter net profit attributable to shareholders was RMB372 million, up 2% year on year and superficially ahead of expectations. However, this included RMB123 million in fair-value gains and RMB97 million in investment income, while RMB119 million in asset impairments or write-offs weighed on earnings. Citi calculated recurring net profit at only RMB150 million, down 40% year on year. The report did not provide a complete reconciliation for nonrecurring items, taxes, and other differences, so recurring profit cannot be derived by mechanically netting the disclosed items.

Operating profit also indicates that the recovery remains incomplete. Second-quarter operating profit was RMB590 million, down 8% year on year, while the operating margin fell 0.6 percentage points to 4.7%. The operating expense ratio rose to 12.2%, up 2.5 percentage points year on year and 1.5 percentage points quarter on quarter. Much of the benefit from the higher gross margin was absorbed by rising expenses and insufficient revenue.

The company also provided figures excluding foreign-exchange effects: on this basis, second-quarter revenue would have reached RMB13.1 billion and net profit RMB1 billion. These figures suggest substantial currency pressure, but they represent a counterfactual management-adjusted view. Reported revenue of RMB12.506 billion and attributable net profit of RMB372 million remain the starting point for assessing the period’s operating performance; the two sets of figures are not interchangeable.

The second quarter therefore leaves three clear conclusions: revenue recovered more slowly than expected; the product-mix shift did lift gross margin; and recurring and operating profit have yet to improve in tandem. For the material fourth-quarter recovery forecast by Citi to emerge, at least two of the three variables—revenue, gross margin, and the expense ratio—must continue to improve.

The Revenue Mix Is Changing, with Margin Improvement Preceding Revenue Growth

Lingyi iTech’s core capability is integrating materials, precision functional components, modules, and assembly services on a single manufacturing platform. Traditional electronic-device operations still account for the overwhelming majority of the business, while incremental growth is coming from higher content per device and expansion in automotive, liquid cooling, power, and robotics. The second-quarter gross-margin improvement is an early indication that the changing revenue mix has begun to flow through the income statement.

In the first half of 2026, the electronic-device business generated RMB20.5 billion in revenue, down 2% year on year and representing 82% of total revenue. Gross margin rose 0.6 percentage points year on year to 17.5%. Revenue from the traditional business did not grow, but profitability improved modestly. The company’s deliberate withdrawal from low-margin processes helps explain how weak revenue and a higher gross margin could occur simultaneously.

The automotive and advanced air mobility business delivered a different growth profile. First-half revenue reached RMB3.5 billion, up 198% year on year and accounting for 14% of total revenue, while gross margin increased 6.3 percentage points to 14.1%. Its gross margin remains below that of the electronic-device business, but the gap has narrowed materially. If scale and yields continue to improve, automotive’s profit contribution should rise faster than its share of revenue.

The segment data also require careful treatment. RMB12.506 billion is revenue for the second quarter alone, whereas RMB20.5 billion and RMB3.5 billion are first-half segment figures; the latter two cannot be treated as the second-quarter revenue mix. The report did not provide complete segment profit forecasts, customer order volumes, or capacity-utilization data. At this stage, the direction of the mix shift can be confirmed, but the figures are insufficient to build a precise segment valuation model.

The value of the revenue-mix shift is that the same RMB100 million of revenue can generate different levels of gross profit. As low-margin assembly revenue contracts and higher-margin thermal management, liquid cooling, and precision components increase their share, the overall gross margin can rise even if total revenue falls short of the original forecast. Citi’s outlook for 2027–2028 relies primarily on this structural shift continuing.

This strategy will also create near-term friction. New-project launches require upfront R&D;, engineering, and sales investment, while the associated revenue and profit often emerge several quarters later. The higher second-quarter expense ratio and lower operating margin show that the company remains in a transition phase in which the gross-margin mix is improving but the expense burden remains heavy. The genuine inflection point will come when incremental gross profit consistently covers expense growth.

Among the Four Growth Drivers, Foldables and Liquid Cooling Have the Clearest Timelines

The first driver of the fourth-quarter recovery is foldable and premium smartphones. The company supplies components including precision hinge structures, display support plates, ultra-thin vapor chambers, and steel battery-shell modules. According to management discussions cited by Citi, content value per device for these products is several times that of other handset models. Mass-production ramp-up is scheduled for the third and fourth quarters of 2026, with volumes expected to increase month by month from September through December.

Read more

• Pre-Market Highlights

• Full AI & Semiconductor Value Chain

• AI Models, Applications & CapEx

• Hyperscale/Cloud CapEx & Data Centers

• GPUs, CPUs & ASICs

• HBM, DRAM, NAND, SSD & HDD

• Semiconductor Equipment, Advanced Packaging & Materials

• Optical Communications & Optics Supply Chain

• Internet & Platforms

• Software & SaaS

• Consumer Electronics & Intelligent Vehicles

Executive Summary

404K SEMI-AI | August 31, 2026

Pre-Market Highlights

Pre-market tech leadership continues to broaden deeper into infrastructure. While faster GPUs remain critical, the actual return on AI capital expenditures is increasingly dictated by whether data can reach compute cores in time, whether racks can secure sufficient memory and optical interconnects, and whether data centers can access stable power grids.

Memory represents the most acute constraint. HBM continues to crowd out leading-edge DRAM wafer capacity, driving persistent contract price gains across conventional DRAM and NAND. Meanwhile, competition among SK Hynix, Samsung Electronics, and Micron has expanded beyond stacking yields into custom base dies, advanced packaging, and customer co-design.

The application layer has entered a key validation window. As MiniMax accelerates commercialization, OpenAI and Anthropic continue expanding their training toolchains and hardware architectures. Concurrently, developments in autonomous agents, factual recall, and coding agents suggest that software orchestration, reasoning quality, and dataflow efficiency may become bottlenecks sooner than raw parameter scaling.

Full AI & Semiconductor Value Chain

AI Models, Applications & CapEx

• MiniMax
  1) Morgan Stanley notes that token consumption surged 20x between January and July 2026, while customer count grew from 200,000 at year-end 2025 to 2.0 million, with enterprise clients generating 80% of ARR.
  2) The M31 (M3.1), M3 Pro, and H31 (H3.1) lineups constitute the 2H2026 product portfolio; whether gross margin rebounds as inference unit economics improve will serve as the primary test of commercialization quality.

"Token consumption grew 20x, while customer count expanded from 200,000 at year-end 2025 to 2.0 million today (a 10x increase)."

• OpenAI
  1) Management stated that Astra achieved its "AI Research Intern" milestone ahead of its September schedule, marking a transition from question-answering to executing end-to-end research workflows.
  2) OpenAI also procured tens of thousands of Mac minis for reinforcement learning workloads and used AI-assisted hardware description languages to design over half of its custom chip cores, expanding toolchains across model training and silicon design in tandem.

• Anthropic
  Anthropic is leasing Mac minis via AWS to handle select reinforcement learning tasks while using Sonnet 5 to post-train early Opus 4.8 checkpoints: testing over 50 approaches in roughly 60 hours, it achieved near-production alignment scores using just 2,400 samples. The key question is whether this few-shot post-training methodology can be consistently replicated at scale rather than remaining an isolated experimental result.

• AI Infrastructure Architecture
  The AI performance bottleneck has shifted from single-chip compute throughput to dataflow efficiency. Over the past 20 years, server peak floating-point performance has grown roughly 3x every two years, whereas DRAM bandwidth has increased by only 1.6x and interconnect bandwidth by just 1.4x. As compute speeds accelerate, the idle overhead incurred while waiting for data becomes increasingly visible.

"Actual real-world performance of AI systems is no longer determined by any single chip or compute metric alone, but by a holistic dataflow architecture orchestrating compute, memory, storage, networking, power, and thermal management."

Hyperscale/Cloud CapEx & Data Centers

• Google
  1) Google's planned $17 billion data center campus is slated to pay $40 million annually to the host county over seven years, surpassing the jurisdiction's current annual residential property tax revenue of $25 million.
  2) The development agreement requires Google to absorb energy costs and mandates closed-loop cooling, residential buffer zones, and noise caps, tying project returns directly to localized community commitments.

• Microsoft
  Architecture disclosures at Hot Chips indicate that Microsoft's local accelerator links feature 6 or 7 x 400G connections, whereas inter-compute-tray and inter-rack links provide only 2 x 400G. Performance bottlenecks are consequently shifting from per-chassis GPU density to cross-rack bandwidth and latency, making network topology a direct determinant of cluster utilization.

• Amazon
  AWS is leasing Mac minis to Anthropic for reinforcement learning workloads, reflecting a task-specific segmentation of cloud AI hardware: general-purpose GPU clusters continue handling large-scale training, while select OS-level and RL tasks migrate to dedicated bare-metal machines. Whether operational cost savings outweigh cluster management complexity remains to be verified.

• Dell Technologies
  Dell previously reported AI server revenue of $16.1 billion, AI orders of $24.4 billion, and an order backlog of $51.3 billion, alongside FY2027 AI server revenue guidance of approximately $60 billion. Management noted that demand continues to outstrip supply, with memory remaining the primary bottleneck. Upcoming earnings should focus on backlog conversion rates and gross margin resilience rather than headline order intake.

• PJM Large-Load Tariff Mechanism
  PJM filed to subject new data centers exceeding 50 MW to priority curtailment rules. The prior two capacity auctions recorded capacity deficits of 6,623 MW and 6,831 MW, respectively, with regional peak load projected to expand by 32 GW by 2030. Data center competitiveness has evolved from securing chips to securing dedicated new power interconnection capacity.

"Ultimately, user-perceived responsiveness depends not only on raw model performance, but equally on how fast the required data can be ingested and moved."

GPUs, CPUs & ASICs

• NVIDIA
  1) The carrying value of NVIDIA's private company equity investments rose from $380 million a year ago to $4.79 billion, roughly 3.4x its net property, plant, and equipment (PP&E;); $1.59 billion in other non-operating income for the April quarter reflects these mark-to-market revaluations flowing through the P&L.;
  2) NVHBM integrates memory controllers and PHY interfaces into a custom base die, claiming up to a 30% bandwidth increase, a 15% power reduction, and a 25% increase in compute silicon area—extending its platform dominance directly into memory specifications.

• Intel
  1) Intel secured approximately $23 billion in external financing ahead of its capital spending cycle, raising its 2026 CapEx budget from $18 billion to $20 billion.
  2) Its EMIB-T advanced packaging technology reportedly targets a 40% gross margin and has attracted interest from major hyperscalers; SK Hynix is also evaluating Intel Foundry for HBM4E base dies, though formal purchase orders have not yet materialized.

• AMD
  AMD is establishing AgentX as its primary software benchmark for autonomous agent workloads. While its hardware maintains an attractive performance-per-dollar profile, software stack maturity, internal R&D; GPU allocation, and automated QA validation clusters remain near-term constraints. Competitive success will hinge on software converting theoretical compute into effective throughput rather than silicon specs alone.

• Broadcom
  Samsung Electronics' lead in HBM4 speeds is converting what was once a pure supply dependency into a competitive advantage for Broadcom's custom ASIC programs. Concurrently, PTI reported that its 2027 panel-level packaging capacity is fully booked by AMD and Broadcom, tethering Broadcom's execution capabilities to both HBM4 and advanced packaging supply chains.

"This relationship is critical because HBM R&D; cycles can no longer be decoupled from accelerator design cycles."

HBM, DRAM, NAND, SSD & HDD

• Memory Pricing
  In 3Q2026, conventional DRAM contract prices are projected to increase 13%–18% QoQ, NAND by 10%–15%, and server DRAM by 13%–18%; 4Q2026 server DRAM price gains are expected to moderate to 3%–8% QoQ. While pricing remains upwardly mobile, consumer-side price resistance and long-term procurement agreements (LTAs) are beginning to cap further acceleration.

• Micron
  Morgan Stanley estimates Micron's 2026 DRAM bit share at approximately 24%. The company has commenced volume shipments of HBM4 to lead customers and targets mass production of HBM4E in 2027. Crucial tracking items center on whether its HBM ramp can simultaneously resolve base die integration, packaging yields, and customer qualifications, rather than wafer input growth alone.

• Samsung Electronics
  1) Industry sources indicate Samsung's current lead in HBM4 speed may enhance its ASIC collaboration with Broadcom.
  2) Monthly HBM wafer starts reportedly expanded from 150k to 260k wafers while conventional DDR output remained flat. Incremental capacity is prioritized for high-margin HBM, sustaining supply pressure on mobile and PC DRAM pools.

• SK Hynix
  1) Capital investment in its Indiana advanced packaging facility exceeds $4 billion, with cleanroom readiness scheduled for October 2028 and volume production of next-generation HBM slated for 2H2029.
  2) The company is evaluating Intel Foundry for HBM4E base die fabrication; TSMC quotes are reportedly 3x to 4x the cost of SK Hynix's in-house 1 billion-class DRAM base dies, though Intel's production yields and delivery reliability require validation.
  3) Site selection feasibility studies for a Japan joint-venture facility are underway.

"The facility adopts a structural division of labor: 'leading-edge DRAM wafer fabrication in Korea + advanced packaging and test on US soil.' This is a localized layout targeting packaging and test rather than full-supply-chain reshoring."

• Kioxia & SanDisk
  The joint venture plans to invest approximately ¥5 trillion in Japan through FY2033, including ¥1.8 trillion for the new K3 fab targeting an FY2030 ramp to expand monthly output from 350k to 550k wafers. UBS views the expansion primarily as strategic leverage for long-term contract negotiations; should 2H2027 capacity release coincide with soft consumer demand, NAND pricing faces downside cyclical risk.

• CXMT
  Morgan Stanley forecasts CXMT monthly capacity to expand from 300k wafers in 2026 to 500k in 2028 and 800k in 2031, representing 13% of global DRAM wafer capacity and 11% of bit shipments in 2026. Process yields on its ~16nm GEN4B node are estimated to exceed 90%; however, multi-patterning overhead increases mask counts, cycle times, and unit costs, meaning scale impact will precede true technological parity.

• YMTC
  Morgan Stanley modeled stress tests based on 30%–60% YoY AI SSD demand growth in 2028, outlining YMTC monthly capacity scenarios of 310k to 470k wafers. Full ramp across all five fabs yields a potential 24% global NAND share. If AI CapEx holds and industry supply additions remain disciplined, market balances will stay tight; conversely, uncoordinated capacity additions would amplify oversupply risk.

"Explosive demand for compute and High Bandwidth Memory (HBM) in AI data centers has triggered severe capacity crowding-out effects."

Semiconductor Equipment, Advanced Packaging & Materials

• Applied Materials
  1) Applied Materials identifies reducing data movement power consumption as the central architectural bottleneck in AI semiconductors, extending its process roadmaps from GAA and HBM to 2.5D/3D packaging and hybrid bonding.
  2) Channel checks indicate customers have locked in orders across a two-year horizon under terms featuring cancellation and expedite penalties. Long lead times reflect tight equipment balances, though revenue conversion depends on customer fab buildout schedules.

• PTI
  PTI plans to allocate NT$70 billion to build panel-level packaging lines for AI silicon targeting 2027 volume production, noting that capacity for that year is fully booked by AMD and Broadcom. Its initial automated fan-out panel-level packaging (FOPLP) line for HPC adopts a 515x510 mm panel format; commercial yields and customer qualification remain key execution milestones.

• ASE Technology Holding
  Google TPU demand, rising outsourcing from Intel, and packaging price adjustments are driving advanced packaging and testing orders at ASE. As Google expands TPU design engagements across Broadcom, MediaTek, Marvell, and AMD, ASE's revenue conversion will depend on whether upstream TSMC wafer starts and CoWoS allocations keep pace.

• China Jushi
  UBS raised its price target from RMB 43 to RMB 58 with a Buy rating, upgrading 2026–2028 net profit estimates by 22%, 84%, and 107%, respectively. The price of 7628 electronic-grade fabric reached RMB 10.20/meter in August, with air-jet loom lead times extending to 12–18 months and projected supply deficits of 13%, 10%, and 6% over the forecast period. AI electronic fabric qualification and accelerated capacity additions represent the primary risks to this thesis.

"Constrained by bottlenecks including 12–18 month lead times for Toyota looms, a lack of qualified domestic replacement equipment, and capacity conversions from standard fabric to high-margin thin/ultra-thin fabric for AI servers."

Optical Communications & Optics Supply Chain

• TSMC
  TSMC is leveraging its COUPE platform and electronic-photonic design automation tools to extend advanced packaging capabilities into Co-Packaged Optics (CPO); NVIDIA Spectrum-X has entered this architecture, targeting volume ramp in 2H2026. Broader customer adoption depends on whether optoelectronic integration yields, packaging cost curves, and design ecosystems can lower commercial production barriers.

• Lumentum
  Evercore initiated coverage with an Outperform rating. The company projects the AI optical market to expand from roughly $18 billion to over $90 billion by 2030; despite an 8x increase in EML output relative to FY2023, shipment deficits exceed 30%. Reliability and high-volume manufacturability data for 1060nm VCSEL arrays will be disclosed at an upcoming September conference.

• Coherent
  Coherent indicated that Near-Packaged Optics (NPO) and CPO offer comparable dollar-content value, with active engagements underway across nearly all key strategic hyperscalers. The company expects initial CPO revenue in the December quarter and plans to launch its PhotonLink integration platform on September 21; the validation focal point is whether customer pilot readiness converts into sustained purchase commitments.

• VIAVI Solutions
  CPO testing equipment has begun generating revenue, with management guiding for expanded contributions in the December quarter. Because a single channel defect can compromise the yield of an entire integrated package, customers must execute design verification, yield optimization, and test protocols ahead of volume assembly—positioning test equipment revenue as a leading indicator for CPO deployments.

• Marvell Technology
  Consensus models project that Marvell's Celestial acquisition could reach an annualized revenue run-rate of $500 million by 4Q2028 and $1.0 billion in 2029, compared to negligible contributions today. While CPO qualification runways provide long-term optionality, they elevate execution risk; investors should focus on hyperscaler design wins, initial revenue milestones, and operating loss narrowing rather than underwriting distant terminal valuations.

Internet & Platforms

• Meta
  Meta’s multistate settlement totals up to $18 billion, comprising up to $16.7 billion in core litigation payouts, $459 million in privacy claims, and $75 million in legal fees. Facebook and Instagram must cap teen screen time, mute notifications by default during school hours, and restrict disabling certain safety settings without parental consent; the settlement involves no admission of liability.

"The agreement requires Facebook and Instagram to implement additional safeguards, including capping teen screen time, muting notifications by default during school hours, and prohibiting teens from disabling certain safety settings without parental consent."

• TikTok, YouTube, and Snap
  Of Meta's settlement, $5.3 billion in payouts over the next 10 years is contingent on competitors adopting similar safeguards and contributing funds. This structure converts a one-time litigation penalty into industry-wide coordination pressure: if rivals follow suit, compliance costs are shared across the sector; if they refuse, Meta can market its remediation as a child-safety differentiator.

• Apple
  Age verification may shift down to the App Store infrastructure layer. Authenticating age at the device or app-store level reduces redundant identity collection across individual apps, but centralizes verification liability and privacy exposure on Apple. The key implementation challenge is whether platforms can transmit only the necessary age bracket without expanding broader personal data collection.

• Google
  Google’s $17 billion data center project locked community benefits into a long-term agreement: paying $40 million annually over 7 years while covering power costs and complying with closed-loop cooling, buffer zones, and noise limits. The feasibility of mega-platform expansions increasingly depends on clearly addressing local tax contributions, water, power, and noise burdens.

• Amazon
  Amazon holds preferred shares in Anthropic, making valuation changes an influence on non-operating income. While investing in frontier AI labs locks in AWS cloud demand, it also introduces private-equity valuation volatility to the balance sheet; sustainable returns must ultimately come from cloud revenue, long-term contracts, and operating cash flow rather than unrealized revaluations.

• AI Equity Accounting
  Mark-to-market revaluations of private AI investments reportedly generated substantial paper gains for Google and Amazon. If future IPOs price below these private valuations, write-down risks will flow directly into the income statement. Financial analysis must decouple core operating income from non-cash equity revaluations.

• Teen Default Settings
  Default safeguards reduce parental enforcement friction: automatically suppressing notifications during school hours and defaulting to screen-time limits is far more effective than relying on manual user configuration. Platforms must demonstrate safeguard efficacy through post-implementation retention, session duration, and complaint data, rather than merely offering settings menus.

• Boundaries of Algorithmic Distribution
  Regulatory scrutiny is shifting from "what users post" to "what platforms actively amplify." Passive hosting entails different liabilities than algorithmic amplification; if recommendation engines continuously push harmful content to teens to optimize engagement, product architecture and distribution mechanics become standalone targets of legal liability.

"Core infrastructure must remain open, and passive hosting platforms should generally not bear liability for user content. Instead, regulatory scrutiny should concentrate on ad-monetized platforms that lack price constraints due to engagement-driven incentives."

• Product Design Liability
  Litigation is bypassing traditional content liability shields by framing addictive user interfaces, push notifications, and recommendation algorithms as product safety defects. For platforms, this broadens exposure from individual content moderation to feature engineering, default settings, and experimentation logs, transforming compliance into an ongoing product R&D; expense.

• Section 230 Protections
  While Section 230 shields platforms from liability for user-generated content, this defense weakens when plaintiffs focus on proprietary product design choices. If legal precedent adopts this distinction, social platforms must maintain rigorous audit trails detailing recommendation algorithms, youth experimentation, and safety governance decisions.

• Economics of Ad-Supported Platforms
  Ad-funded platforms monetize engagement, and users do not directly pay for safer or less addictive experiences, resulting in weak pricing constraints. Regulatory intervention targeting market failures should focus on incentive alignment and algorithmic amplification, rather than conflating cloud infrastructure, passive hosting, and ad platforms under a uniform regime.

• Empirical Evidence on Youth Mental Health
  An NBER study of New Jersey data from 2008 to 2019 revealed that the 2012 rise in reported youth mental health issues coincided with revised screening guidelines, and the 2016–2017 spike aligned with diagnostic coding changes, while rates of self-harm, suicide attempts, and mortality remained largely flat. Reported diagnosis rates should not be conflated with causal evidence of deteriorating behavior.

"Meanwhile, actual occurrences of self-harm, suicide attempts, and mortality remained largely flat, indicating that reporting trends cannot simply be equated with an underlying behavioral crisis."

• Parental and Household Governance
  Platform default guardrails support parental enforcement but cannot replace household boundaries. A more rigorous approach involves tracking nighttime usage, school-hour alerts, parental control adoption, and objective harm metrics in tandem, avoiding the tendency to attribute every behavioral shift entirely to social media or specific feature sets.

• Age Verification and Privacy Trade-Offs
  While age verification bolsters youth protection, it also legitimizes expanded platform collection of government IDs, facial biometrics, and device data. Implementations must enforce data minimization, strict purpose limitations, and retention caps to prevent a safety safeguard from mutating into new privacy and breach liabilities.

"While age verification introduces privacy risks, it provides Meta with a formal justification to expand data collection and could push Apple and the App Store into serving as the foundational layer for age authentication."

• Competitive Linkage Provisions
  The $5.3 billion in contingent payouts over the next 10 years tethers Meta’s net liability to industry behavior. If peer platforms implement comparable safeguards, Meta’s relative compliance disadvantage narrows; if they refuse, competitors face heightened litigation and brand risk. This structural design turns the settlement into both a loss-mitigation agreement and a competitive lever.

"Legally and commercially, the core dispute is pivoting from 'whether platforms are liable for user-generated content' to 'whether platforms harm children through addictive product design,' effectively bypassing traditional Section 230 immunity."

Software & SaaS

• Coding Agent Orchestration
  Across 169 SWE-bench Verified tasks within a 20,480-token context window, solely tuning tool output compression, stall detection, and command guardrails boosted Qwen3.6’s average per-task benchmark from 28% to 49%, while fully resolved tasks jumped from 43 to 72. Holding the underlying model constant, orchestration layer optimizations alone can drive substantial performance gains.

"Holding the model constant and modifying the harness can dramatically alter coding agent benchmarks when context budgets are constrained."

• Context Windows
  When the context window expanded to 262,144 tokens, performance between the control and experimental groups converged almost entirely, indicating that orchestration gains are concentrated in genuinely context-constrained scenarios. Engineering teams must evaluate models, tooling, context strategies, and runtime controls concurrently rather than attributing all performance variances to the foundation model.

• Factual Recall
  A Google Research study testing 2,150 facts across 13 models found that while frontier models encode 95% to 98% of facts within their weights, direct closed-book prompting still fails to recall 26% to 34%, and enabling reasoning tokens still leaves an 11% to 12% failure rate. Expanding parameter size solves encoding deficits ("not stored"), whereas retrieval and prompting are better suited for recall failures ("stored but unretrieved").

"Frontier models encode 95% to 98% of facts in their weights, yet closed-book direct queries still fail on 26% to 34%, and enabling thinking mode still leaves an 11% to 12% failure rate."

• Open-Ended Research Agents
  In an experiment granting an AI agent 6 days and a $3,000 budget, the agent produced 2 research papers that were both rejected, while leaving over half the budget unspent. Although the agent executed hundreds of experiments, recovered from GPU failures, and compiled papers, it failed on critical judgment: when confronted with negative feedback, it merely narrowed its claims rather than redesigning the experimental setup.

• Long-Horizon Agent Security
  When long-horizon agents persist memory across iterations, security monitoring must retain equivalent context. Attackers can distribute malicious payloads across multiple seemingly benign steps, causing single-trajectory monitors to miss combinatorial risks. Enterprise deployments require persistent stateful audits rather than isolated short-horizon safety checks.

"Attackers can distribute malicious evidence across multiple individually benign-looking steps, preventing single-trajectory monitors from capturing sufficient context to distinguish an attack from legitimate workflow."

• Reinforcement Learning Monitoring
  Verifiable reward signals accelerate model optimization but risk inducing unintended behaviors. Case studies from OpenAI demonstrate the necessity of monitoring chain-of-thought reasoning and behavioral drift alongside reward metrics; otherwise, models may nominally satisfy the target through unintended exploits.

• Anthropic Subscription Transparency
  Claude Max’s "5x or 20x" multiplier applies to a rolling 5-hour session window rather than an aggregate weekly cap. Lawsuits allege that a $200 tier yields only roughly 2x the weekly volume of a $100 tier. SaaS pricing models must transparently detail session windows, weekly limits, and model-specific quota allocations.

"Within each 5-hour window, the '20x' claim holds true. However, weekly caps create an entirely different reality."

• Codex Quotas
  ChatGPT Work and Codex reset paid subscription usage caps after reaching 25 million active users, clarifying that Pro's "20x" capacity represents 20 times the weekly limit of Plus, with neither Pro tier subject to a 5-hour rolling cap. One-off resets operate under different rules than standard weekly quotas, and users must refer to active account dashboards.

• OpenAI Semiconductor Design Tooling
  OpenAI stated that its new hardware programming environment paired with AI implemented over half of a chip core and optimized power, performance, and area (PPA) in the final phase, successfully fitting the layout into floorplan modules to meet tape-out deadlines. While AI has expanded from software coding into semiconductor implementation toolchains, outputs still require strict timing closure, power validation, and volume fabrication verification.

"In practice, by utilizing AI for PPA optimization, we successfully fit the design in, met the target schedule, and achieved our originally intended datasheet specifications."

• Checkpoint Storage
  Training clusters periodically persist checkpoints; if the storage tier cannot write or read efficiently, the entire compute cluster sits idle. MLCommons launched MLPerf Storage to elevate storage throughput from an auxiliary hardware specification into a benchmarkable metric. Scheduling software must simultaneously manage checkpoint frequency, recovery time, and training loss exposure.

• Data Tiering
  Model weights, context windows, and external documents must dynamically traverse bulk storage, host system memory, and near-accelerator memory. Placing high-frequency data closer to compute cores minimizes latency and energy draw, while cold data resides in lower-cost tiers. The competitive value of inference software is increasingly defined by intelligent data placement.

• Inference Pipeline Latency
  AI search and conversational agents concurrently query dialogue history, web search results, and external corpora. User-perceived latency is therefore driven not just by model FLOPs, but by data ingestion, reranking, and memory transport. Optimizing inference requires end-to-end pipeline visibility rather than isolated peak GPU throughput metrics.

• Software and Interconnect Co-Design
  Parallelism and memory offloading alleviate HBM capacity bottlenecks but transfer pressure to interconnect latency. Software orchestration must strike an optimal balance among parallel scale, communication overhead, and memory access; excessive partitioning risks leaving larger accelerator clusters idling at synchronization barriers.

• Model Evaluation Methodology
  Unencoded facts and unretrieved facts represent fundamentally distinct failure modes. Binary pass/fail benchmarks mistakenly attribute all errors to inadequate parameter capacity or dataset coverage. A more rigorous evaluation protocol systematically tests rephrased prompts, multiple-choice formats, reasoning chains, and retrieval before justifying capital allocation toward larger model scale.

• Enterprise AI Acceptance Criteria
  Benchmark scores serve only as a baseline in enterprise AI procurement. Pragmatic acceptance metrics include real-world task resolution rates, human rework overhead, tool-calling failures, context exhaustion, cost per task, and system resilience. Research agents that execute flawlessly yet fail on scientific judgment illustrate that high automation throughput does not guarantee commercially viable output.

Consumer Electronics & Intelligent Vehicles

• Samsung Electronics
  Samsung has reportedly solidified its #1 position in the smartphone market with a 32% share, spanning the budget-to-flagship spectrum with the Galaxy A07, A17, and Galaxy S26. While rising memory prices disproportionately squeeze margins on low-end handsets, Samsung's captive memory supply and comprehensive product tiering better position it to absorb component cost volatility in the near term.

"Amid tightening memory supply, budget-focused vendors like Xiaomi and Transsion faltered, while Samsung Electronics—leveraging its extensive product portfolio—successfully reinforced its #1 position with a 32% market share."

• Apple
  Due to virtualization licensing restrictions, AI labs are procuring physical Mac mini units in volume. While this provides a near-term boost to hardware demand, it inflates the operational cost of cloud-based Mac Agents. If Linux cloud Agents establish a decisive cost advantage, Apple will need to revise its licensing and virtualization frameworks to convert one-off hardware sales into a sustainable developer ecosystem.

"You cannot spin up vast fleets of Mac virtual machines to train AI Agents, forcing the purchase of physical hardware."

• NVIDIA AI PCs
  Initial batches of RTX Spark PCs have reportedly sold out. Powered by TSMC 3nm Grace-plus-Blackwell chips co-designed by NVIDIA and MediaTek, the N1X and N1 are scheduled to ship this fall as Windows on Arm devices. With premium configurations starting around $3,400, the sellout reflects constrained initial supply rather than a broad-based inflection in PC demand.

"Grace-plus-Blackwell superchips co-designed with MediaTek on TSMC's 3nm process will ship this fall as Windows on Arm devices."

• MediaTek
  Co-designing the N1X and N1 marks MediaTek's expansion from smartphone SoCs into premium Windows on Arm computing platforms. Sustained dollar content hinges on incremental silicon allocations from NVIDIA, OEM SKU expansion, and software compatibility; early sellouts confirm initial niche demand rather than scaled market share.

• ASUS
  ASUS reported that channel pre-orders have cleared initial N1X inventory, prompting the OEM to lobby NVIDIA for additional supply. The sellout of premium AI laptops in a sluggish PC market highlights core enthusiasts' willingness to pay for a full NVIDIA technology stack; the key next metrics to track are restocking velocity, return rates, and active device telemetry.

"An Arm AI laptop priced above $3,000 selling out is an outlier, not an indicator of broader market conditions."

• MSI
  MSI indicated that N1X inventory allocated to the US and Chinese markets is essentially sold out, and it is likewise seeking incremental allocations. While regional stockouts confirm that early demand outpaced initial supply, the product remains in a high-end niche; if memory and other component costs continue to climb, pushing price points into mainstream tiers will face higher friction.

• Dell Technologies
  Dell was included in the initial roster of N1X brands, even as its server business contends with tight DRAM and NAND supply. Because client PCs and AI servers compete for memory components, supply chain capacity, and customer IT budgets, component allocation will directly dictate gross margins across both units. On the client side, the critical indicators to watch are commercial SKU volume and channel replenishment rather than initial brand marketing.

• HP
  HP is also part of the first-wave N1X OEM lineup. While Windows on Arm was historically hampered by application compatibility, broadened support from NVIDIA and gaming studios offers HP a differentiated high-end platform. However, absent confirmed shipment volumes and pricing structures, this remains an isolated product opportunity rather than the catalyst for a broader PC replacement cycle.

• Lenovo Group
  Lenovo is among the first-wave N1X launch partners, leveraging its global enterprise channels and commercial client base. Provided NVIDIA maintains stable chip supply, Lenovo is well positioned to drive enterprise trials for Arm AI PCs. Key validation metrics include commercial software compatibility, channel restocking rates, and enterprise deployment scale, rather than early sellouts of limited SKUs.

• Microsoft
  Microsoft serves as both an ecosystem launch partner and the architectural backbone for Windows on Arm software compatibility. Expanding support across gaming studios has narrowed legacy application gaps, but platform scalability still depends on driver maturity, virtualization, enterprise manageability, and whether legacy x86 translation delivers true functional parity.

• PC Memory Costs
  Q3 PC DRAM contract prices are projected to rise 18% to 23% QoQ despite persistent weakness in notebook demand. While OEMs are building over 10 weeks of inventory to buffer against potential shortages in 2027, they are simultaneously reassessing high-BOM configurations. AI PCs must justify rising memory costs through demonstrable performance gains or compelling new feature sets to prevent retail price points from dampening unit volume.

• Windows on Arm Ecosystem
  The N1X leverages NVIDIA's full hardware and software stack to penetrate the premium tier, addressing the historical lack of a true flagship platform for Arm-based PCs. Whether initial traction spreads from enthusiasts to mainstream users will depend on application compatibility, gaming support, battery efficiency, and local AI execution, rather than peak synthetic compute benchmarks.

• Waymo
  Autonomous driving miles still account for a negligible fraction of total vehicle miles traveled, meaning Waymo should not be prematurely written off in light of Tesla's progress. More meaningful indicators to track include driverless operating miles, municipal expansions, unit cost per mile, and disengagement frequencies, rather than extrapolating an industry endgame from a single competitor's product announcements.

"Tesla's success does not automatically spell defeat for Waymo or Uber, particularly when autonomous miles still represent only a tiny fraction of total vehicle miles traveled."

• Tesla
  Tesla's autonomous driving initiatives expand consumer awareness and the broader addressable market, but its model still requires validation through commercial operating data on safety and unit economics. While its execution may accelerate overall industry penetration, it does not inherently crowd out Waymo or alternative platforms; market competition will primarily hinge on geographic footprint, fleet utilization rates, and regulatory approvals.

• Uber
  Uber operates primarily as a demand aggregator and fleet operating network rather than a single autonomous technology developer. With the autonomous mobility sector in its early stages, multiple technology providers are likely to integrate into its network simultaneously. Going forward, investors should monitor partner fleet scale, autonomous trip share, and platform take rates, rather than treating Tesla's pipeline as an immediate direct replacement.

Related Reading

404K SEMI-AI Evening Brief, August 27, 2026 — Bottlenecks Spill Over: Nvidia Raises Demand Outlook as Memory, Packaging, and Power All Tighten

404K Tech Morning Brief — August 18, 2026 — Repricing the Bottlenecks: Memory Strengthens as Compute Expansion Shifts to Power, Packaging, and Optical Interconnects

404K SEMI-AI Morning Brief | August 31, 2026 — Cascading Bottlenecks: AI CapEx Expansion Hits Security, Memory, and Advanced Packaging Constraints

404K SEMI-AI Evening Brief, 2026-08-19 — Bottlenecks Spill Over: Inference Demand Shifts Pressure to Memory, Packaging, and Power

404K SEMI-AI Evening Brief 2026-08-04 — Bottlenecks Spread: AI Demand Extends to Memory, Optical Interconnects, and Advanced Packaging

• I. The Bottom Line: AI Memory Is an Ever-Expanding Pyramid

• II. Training Demand: HBM Sets the Ceiling; DRAM and SSDs Keep the System Running

• III. Inference Has Two Distinct Stages: Prefill Is Compute-Bound; Decode Is Memory-Bound

• IV. Multi-Agent Systems and Long Context: KV Cache Becomes a Commercial Bottleneck

• V. HBM Still Captures the Most Value, but Profit Allocation Is Becoming More Complex

• VI. Conventional DRAM Is the Underappreciated Second Pillar

• VII. NAND’s Re-Rating Comes from Tiering, Not Wholesale DRAM Replacement

• 8. HBF Offers Enormous Potential—and Equally Formidable Technical Barriers

• 9. Two Hierarchies Matter: Architectural Position and Physical Media Are Not One-to-One

• 10. Industry Profit Pools: DRAM Dominates Revenue, While High-Performance SSDs Pursue Incremental Upside

• 11. Emerging Technology Pathways: Every “Breakthrough” Introduces a New Cost

• 12. The Ten Most Important Indicators to Track

• 13. Risk Framework: Mass Production, Cost, and Cyclicality Still Separate Strong Demand from High Valuations

• 14. Final View: Upgrade the Thesis from “Buy HBM” to “Buy Data-Tiering Efficiency”

本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读

AI servers are developing a storage architecture far more complex than “GPU plus HBM.” The real opportunity lies in precisely matching each workload to the right storage tier.

I. The Bottom Line: AI Memory Is an Ever-Expanding Pyramid

The greatest value of Bernstein’s Global Memory: An AI Memory Primer is that it breaks the familiar “AI drives storage” narrative into a verifiable technical chain. Training, inference, retrieval-augmented generation (RAG), and multi-agent workflows have different requirements for capacity, bandwidth, latency, power, endurance, and cost. All support HBM growth, but HBM will not capture the entire incremental opportunity.

An AI system can be viewed as a pyramid in which data becomes progressively colder and capacity progressively larger toward the base. At the top is on-chip XPU SRAM—the smallest and fastest tier. Next to the XPU is HBM, which holds the hottest model weights, activations, and KV cache. Below that are CPU-side system DRAM and CXL-expanded memory, followed by Storage Next, local SSDs, and CMX data nodes. Shared SSDs, HDDs, and even tape form the bottom tier. The hotter the data, the closer it must reside to the compute chip; the colder the data, the more it should migrate to high-capacity media with a lower cost per unit of storage.

Figure 1: Bernstein’s AI architecture storage hierarchy. The closer data resides to the XPU, the faster the access and the higher the cost per unit of capacity. Sources: NVIDIA, Solidigm, Bernstein analysis; Bernstein, August 28, 2026, p. 7.

This hierarchy means the central challenge in AI storage has shifted from “is there enough capacity?” to “which processor must access which data, and when?” Putting everything in HBM would make system costs and packaging area unmanageable; relying too heavily on inexpensive NAND would allow latency and IOPS to bottleneck costly GPUs. The industry is therefore moving toward retaining hot data in high-bandwidth media while offloading colder data to the next tier more quickly and transparently.

This creates three investment themes. First, HBM remains the highest-value incremental opportunity because it directly determines available bandwidth for large-model training and the decode stage. Second, conventional DRAM is assuming a larger role in training-data pipelines, RAG indexes, CPU tool calls, and KV cache offloading. Third, NAND will benefit from more than capacity growth: high IOPS, low latency, direct GPU connectivity, and networked data nodes are segmenting enterprise SSDs into multiple profit pools.

This also highlights a common misconception: a strong HBM cycle does not automatically translate into simultaneous earnings growth for every memory company. Wafer allocation, advanced packaging, control of base dies, media type, controller capabilities, and software integration determine how much value each company can capture. AI is creating a longer and more complex value chain, but also greater execution risk.

II. Training Demand: HBM Sets the Ceiling; DRAM and SSDs Keep the System Running

Training most readily creates the impression that “HBM is everything,” because model parameters, activations, and frequently accessed compute data must move rapidly near the XPU. Bernstein argues that memory bandwidth may set the absolute ceiling for model scale and compute speed. Regardless of GPU performance, if weights and intermediate results cannot reach the compute units in time, expensive processing capacity will sit idle waiting for data.

But a training system involves far more than HBM. Raw training data is typically stored in lower-cost shared storage. Before training, it must be cleaned, sharded, shuffled, and prefetched, then pipelined through local SSDs and system DRAM before reaching HBM. The longer training runs, the more important fault recovery becomes. Systems must periodically write model state, optimizer state, and training progress to checkpoints. These checkpoints pass through system DRAM and local SSDs before entering networked shared storage, preventing a single hardware failure from wiping out days or even weeks of computation.

Training therefore drives storage demand in sequence: larger models raise HBM capacity and bandwidth requirements; larger datasets expand local SSD and shared-storage capacity; greater training parallelism increases system DRAM caching and network throughput; and more expensive training workloads demand more frequent and reliable checkpointing. Focusing only on HBM bit demand misses three other equally indispensable sources of growth.

Training workloads also explain why memory supply cannot be assessed solely by total bit output. HBM and conventional DRAM share front-end wafer capacity, while HBM requires more dies, more complex TSVs, stacking, and advanced packaging. As manufacturers allocate more wafers and capital to HBM, conventional server DRAM supply may become less elastic. Even when AI demand is concentrated in premium products, capacity displacement can affect pricing across the entire DRAM market.

The NAND implications are more differentiated. Training datasets and checkpoints both require high-capacity SSDs, but their latency, write-endurance, and throughput needs differ. Dataset reads emphasize sequential throughput and capacity; checkpoint writes require burst bandwidth and reliability; and local caches must withstand frequent reads and writes. Enterprise SSD controllers, firmware, parallel channels, and endurance configurations therefore matter more than NAND layer count alone.

III. Inference Has Two Distinct Stages: Prefill Is Compute-Bound; Decode Is Memory-Bound

After a user request enters a large model, it passes first through prefill and then through decode. Both stages use the same model but face entirely different bottlenecks.

Prefill processes the full input prompt and performs large volumes of parallelizable matrix operations. The report describes it as a generally compute-bound stage. Time to first token, or TTFT, includes scheduling and network time as well as prefill computation. Research cited by Bernstein suggests that chatbot TTFT below 0.5 seconds is ideal, 1 second remains acceptable, and more than 2 seconds may cause frustration; complex research or large tasks can tolerate longer waits. The key variables are XPU compute capacity, input length, batching efficiency, and HBM bandwidth—not all first-token latency should be attributed to memory capacity.

Decode is the autoregressive process of generating tokens one at a time. For every new token, the system must read the model weights and access the Keys and Values associated with prior tokens. Model weights are static and can be shared across users; the KV cache updates as the sequence grows and is unique to each user. Decode is therefore more likely to be constrained by memory capacity and bandwidth.

In simplified terms, KV cache size grows with the number of model layers, hidden dimension, numerical precision, context-token count, and concurrent sequences. Bernstein highlights two direct relationships: the KV cache for an individual sequence expands linearly with token count, and KV caches cannot be directly shared across users. Long contexts and high concurrency therefore create a multiplicative effect. A model’s ability to accept longer documents does not mean a provider can serve more users simultaneously on the same hardware.

The user-experience metric also shifts from TTFT to TPOT—the time between consecutive output tokens. Standard text chat may tolerate around 50 milliseconds, equivalent to 20 tokens per second, while real-time voice, coding, and multi-agent applications may require less than 10 milliseconds. The tighter the TPOT requirement, the smaller the time budget for retrieving cold KV cache after offloading it from HBM, increasing the need for high-speed DRAM, low-latency SSDs, and efficient interconnects.

This chain reveals the core economics of inference: revenue depends on the number of users served, tokens per user, and service quality, while costs are shaped by XPU utilization, HBM occupancy, and cache hierarchy. If large volumes of cold KV cache remain in expensive HBM, providers sacrifice concurrency; if data is offloaded too deeply, latency degrades the user experience. Much of the future competition in inference infrastructure will center on cache scheduling, data tiering, and interconnect efficiency.

IV. Multi-Agent Systems and Long Context: KV Cache Becomes a Commercial Bottleneck

Multi-agent systems further amplify these challenges. One agent’s output often becomes the next agent’s input, while downstream nodes may invoke search, code, databases, or enterprise systems. Each handoff can lengthen the prompt, trigger another prefill, and create a new KV cache for the ensuing generation. As the chain grows, so does the coordination overhead across CPU tool calls, system DRAM staging, network access, and GPU inference.

What distinguishes these workloads is that a single user-visible task may involve multiple model calls behind the scenes. Average tokens per request, peak context length, the number of parallel branches, and retry counts all increase. Providers that continue to plan capacity around single-turn chat can easily underestimate peak HBM usage and cache-migration traffic.

Multi-agent workflows also redefine what counts as “hot” or “cold.” Newly generated intermediate results that are consumed immediately by the next step are short-lived hot data. When a task pauses for an external tool, the associated KV cache may temporarily cool; once the task resumes, it must be retrieved quickly. An ideal system would determine in real time which caches remain in HBM, which spill into system DRAM, and which can move to SSD, then prefetch them before execution resumes. This is precisely the gap that CXL, Storage Next, and CMX aim to fill.

Commercially, long context is not merely a model-feature upgrade. It increases memory consumption per active session and may reduce concurrent users per GPU. If service pricing does not cover the additional memory and interconnect costs, stronger model capabilities can actually worsen unit economics. Assessing AI inference providers’ margins will therefore require tracking not only chip acquisition costs, but also context length, average output tokens, KV-cache hit rates, offload ratios, and service-level requirements.

V. HBM Still Captures the Most Value, but Profit Allocation Is Becoming More Complex

HBM sits alongside the XPU, using wide buses and advanced packaging to deliver extremely high bandwidth. Using a Rubin GPU package as an example, the report illustrates eight HBM stacks arranged around a single package. Adding stacks increases capacity and bandwidth, but consumes more package area and raises the complexity of TSVs, base dies, interconnects, and thermal design.

HBM derives its value from three layers of scarcity. The first is DRAM dies and yield; the second is stacking, bonding, and testing; and the third is the capacity and design expertise required to co-package HBM with the XPU. The market often treats all three layers as part of memory vendors’ competitive moat, but emerging architectures could alter how that value is distributed.

Bernstein focuses on zHBM and NVHBM. zHBM seeks to move the HBM stacks from beside the XPU to above it, shortening interconnects and increasing bandwidth. However, the upper DRAM must withstand heat from the XPU beneath it, while hybrid-bonding yield and cost remain unresolved. NVHBM, meanwhile, would move controller and related functions into a custom base die designed by NVIDIA and potentially manufactured by TSMC. If adopted, memory vendors would continue supplying DRAM stacks but could cede some of the value associated with base-die IP, foundry production, and system differentiation.

This does not mean incumbent HBM suppliers will immediately lose pricing power. High-yield DRAM stacking, TSV processes, thermal management, and customer qualification still present formidable barriers, while HBM also competes with conventional DRAM for wafer capacity. A more balanced conclusion is that the HBM profit pool will continue to expand, but incremental value will be redistributed among memory vendors, XPU designers, foundries, advanced-packaging providers, and the interconnect ecosystem.

HBC offers another approach. It uses LPDDR and conventional packaging, avoiding costly CoWoS and reducing power consumption, at the expense of fewer interconnects and lower performance than HBM. The corporate roadmap cited in the report indicates that Qualcomm plans to ship first-generation HBC in fiscal 2027 and launch a second generation in fiscal 2028. This timeline should be treated only as a product-roadmap milestone, not as evidence of validated mass production. HBC is more likely to serve cost- and energy-sensitive applications with less demanding peak-bandwidth requirements than to replace HBM without performance trade-offs.

VI. Conventional DRAM Is the Underappreciated Second Pillar

System DRAM in AI servers performs four functions. First, it supports preprocessing and pipeline buffering before training data enters HBM. Second, it enables RAG vector-database creation and online index retrieval. Third, it supports CPU-based tool execution, scheduling, and control. Fourth, it acts as a front-end cache for NAND, masking SSDs’ relatively high access latency.

RAG particularly illustrates DRAM’s independent value. During database construction, large volumes of unstructured data are stored on SSDs, while the system uses DRAM to generate vectors and indexes. This is generally a pre-deployment, offline cost that does not recur with every online query. During online retrieval, vector indexes reside primarily in system DRAM. Only after matching results are added to the prompt do they re-enter prefill and decode, driving demand for XPUs, HBM, and KV cache. Attributing all RAG-related demand to HBM would simultaneously overstate HBM demand and understate the importance of system DRAM.

Module form factors are also evolving. SOCAMM2 uses LPDDR, lower voltage, and a lower-profile design, while expanding the data bus from 64 bits on a conventional DIMM to 128 bits. This increases bandwidth per module, reduces power consumption, and improves airflow. The report’s Vera CPU example assigns 8 SOCAMM2 modules to each CPU, or 16 modules across two CPUs. The investment implication extends beyond simply installing more memory per server: it also includes higher value per module and upgrades to power delivery, cooling, and board-level layout.

CXL enables memory expansion and pooling over PCIe, reducing the waste created when each server reserves large amounts of idle DRAM for peak demand. Its value depends on whether software can distinguish latency across memory tiers and place data appropriately. For a CXL-connected NAND SSD to perform a “DRAM-like” role, it will generally also require DRAM caching, controller optimization, and greater parallelism.

The risks to conventional DRAM should not be overlooked. HBM’s consumption of wafer capacity can tighten supply, but supply elasticity may also recover quickly when manufacturers expand capacity, migrate process nodes, or encounter weaker-than-expected AI demand. Server DRAM remains cyclical; structural capacity growth will not automatically eliminate inventory and pricing volatility.

VII. NAND’s Re-Rating Comes from Tiering, Not Wholesale DRAM Replacement

NAND and DRAM differ by several orders of magnitude in access latency. The typical parameters summarized by Bernstein on page 23 show SRAM below 10 nanoseconds, DRAM at approximately 35 nanoseconds, MRAM at approximately 10–50 nanoseconds, PCM above 100 nanoseconds, and RRAM at approximately 100 nanoseconds. NAND is approximately 25–100 microseconds, while HDD is approximately 3 milliseconds. This gap means that inexpensive capacity cannot directly become high-performance memory.

Figure 2: Comparison of the physical characteristics of major storage technologies. NAND’s cost and density advantages come with microsecond-scale access latency and lower write endurance. Sources: Applied Materials, Toshiba, Everspin, Yole, Bernstein analysis; Bernstein, August 28, 2026, page 23.

Endurance and cell area also define the boundaries. The report lists endurance of approximately 10^15 cycles for SRAM and DRAM, 10^9–10^12 for MRAM, 10^6–10^7 for PCM, 10^5–10^6 for RRAM, and 10^5 for NAND. SRAM cell area is approximately 160–280F², versus approximately 6F² for DRAM, while 3D NAND can amortize unit area through hundreds of stacked layers. NAND leads in capacity density and cost; DRAM leads in latency, write endurance, and fine-grained access.

AI is therefore driving further segmentation within NAND. XL-FLASH and Z-NAND use SLC- or MLC-based approaches, sacrificing capacity density for higher IOPS and lower latency, making them suitable for tiers closer to compute, such as Storage Next. TLC offers a more balanced trade-off among performance, endurance, and cost and can be used in CMX. QLC targets lower cost per bit and greater capacity, making it better suited to shared storage and competition with HDDs. No medium is universally superior; workload fit is what matters.

Kioxia Holdings claims that its GP-series SSDs can reach 10 million IOPS in the report year and has set a target of more than 100 million IOPS for 2027. These figures reflect the company’s ambition to move NAND closer to the DRAM tier, but they remain corporate targets and should not be interpreted as capabilities already available across all products. Actual performance will also depend on the read/write mix, queue depth, access granularity, controller, thermal management, and software stack.

Storage Next sits between DDR DRAM and local Compute SSDs, aiming to deliver lower latency and greater parallelism than conventional SSDs. CMX instead deploys capacity in dedicated data nodes connected to compute nodes through the BlueField-4 DPU, Spectrum-X, and Ethernet, allowing GPUs to bypass the CPU and access that capacity directly under suitable software and network conditions. Both address the same problem: enabling GPUs to retrieve data that does not merit permanent residency in HBM with sufficient speed.

8. HBF Offers Enormous Potential—and Equally Formidable Technical Barriers

HBF seeks to stack NAND in much the same way that HBM stacks DRAM, placing HBF alongside HBM around the XPU package. Citing a SanDisk concept, the report compares a pure-HBM configuration with 192GB of total memory against hybrid HBF configurations with 3,120GB and 4,096GB. The capacity advantage is compelling, but these figures illustrate a vendor proposal and should not be treated as generally available production specifications.

Figure 3: SanDisk’s HBF concept. The 3,120GB and 4,096GB figures are vendor target configurations, not generally available production specifications. Sources: SanDisk; Bernstein, August 28, 2026, p. 9.

The commercial rationale for HBF is straightforward: use NAND’s cost and capacity advantages for colder data while keeping that data as close to the XPU as possible, reducing retrieval across servers and networks. However, HBF must bridge multiple performance gaps—from NAND to DRAM and then to HBM. Bernstein explicitly assesses the technical barriers as high.

There are at least five specific challenges. First, NAND read latency is far higher than DRAM latency. Second, write endurance is weaker. Third, the access granularity may not match GPU workloads. Fourth, thermal management becomes more complex after stacking and packaging. Fifth, controllers and software must make data migration as transparent as possible. Failure on any front could leave HBF as an expensive component with enormous capacity that GPUs cannot afford to wait for.

HBF is therefore better understood as a complementary tier to HBM. If successful, it would expand XPU-adjacent capacity, reduce some data movement between networks and remote SSDs, and increase concurrency for ultra-long-context workloads or large-model inference. In the near term, however, it is unlikely to replace the HBM bandwidth required by the hottest weights and active KV cache.

For NAND vendors, HBF opens a high-value market spanning advanced packaging and XPU integration. For HBM vendors, it could ease the cost and area pressures associated with stacking ever more DRAM. For system vendors, it adds another layer of scheduling complexity. Value capture will depend on who controls the media, controllers, packaging, interconnects, and caching software.

9. Two Hierarchies Matter: Architectural Position and Physical Media Are Not One-to-One

Bernstein presents both an “architectural hierarchy” and a “hardware hierarchy,” a critical distinction for understanding emerging product categories. The architectural hierarchy addresses how far data resides from the XPU, who accesses it, and through which interconnect. The hardware hierarchy addresses the underlying storage medium, its latency and cost, whether it is volatile, and its endurance.

Figure 4: Bernstein’s storage hardware hierarchy. Moving down the hierarchy lowers capacity cost but increases access latency; volatile and non-volatile media serve different roles. Sources: Bernstein analysis; Bernstein, August 28, 2026, p. 18.

The same physical medium can appear at multiple architectural tiers. DRAM can serve as HBM, system memory, CXL-expanded memory, or SSD cache. NAND can support local SSDs, CMX, and shared storage, while XL-FLASH, Z-NAND, and HBF seek to move it into higher-performance tiers. Conversely, a single architectural tier may combine multiple media, using caches and controllers to mask differences in the underlying technologies.

This will make product naming increasingly complex. CXL, Storage Next, CMX, HBF, and HBC are not simple substitutes; they address different bottlenecks and occupy different positions in the architecture. Investors should look beyond the labels and answer at least four questions: What is the underlying medium? How far is it from the XPU? How do the CPU or GPU access it? Relative to the tier above, how much cost is saved and how much latency is added?

If a product does not clearly answer these four questions, the purported “new AI storage product” may simply be an old medium under a new label. The genuinely informative metrics are sustained IOPS, tail latency, bandwidth, power consumption, endurance, the share of GPU-direct access, software compatibility, and customer qualification—not a single peak specification.

10. Industry Profit Pools: DRAM Dominates Revenue, While High-Performance SSDs Pursue Incremental Upside

The report’s long-term revenue mix shows that DRAM accounts for most global memory and storage revenue. This is unsurprising: DRAM commands a higher price per bit than NAND, while HBM adds further value through stacking and packaging. AI growth therefore expands the DRAM revenue pool first, with HBM increasing value per unit.

Figure 5: Global memory and storage revenue mix, 2020–2025. The original chart does not label decimal values for each segment, so precise shares should not be inferred from bar heights. Sources: Company reports, Gartner, Bernstein analysis; Bernstein, August 28, 2026, p. 19.

The first group of beneficiaries comprises memory vendors with HBM mass-production capabilities, advanced DRAM processes, and qualification from major customers. Key variables include HBM generation transitions, yields, wafer allocation, advanced-packaging supply, customer share, and conventional DRAM pricing. In strong markets, HBM and conventional DRAM may benefit from simultaneous volume and pricing gains. In periods of stress, failed generation qualifications or lagging yields can rapidly widen the gap between vendors.

The second group consists of the advanced-packaging, base-die, and interconnect ecosystems. zHBM, NVHBM, and taller stacks all require stronger bonding, testing, thermal-management, and foundry capabilities. If XPU vendors control customized base dies, foundries and packaging companies could capture more value, leaving traditional memory vendors increasingly positioned as suppliers of high-performance DRAM.

The third group comprises enterprise NAND and SSD vendors. Commodity capacity growth offers limited earnings leverage. The segments that merit genuine re-rating are low-latency SLC/MLC products, enterprise TLC, controllers, DRAM cache, GPU-direct connectivity, and data-node solutions. A vendor’s ability to enter the Storage Next or CMX ecosystem and consistently deliver low tail latency and high endurance may matter more than peak sequential read/write speeds.

The fourth group comprises HDDs and cold storage. AI-generated and training data will continue to accumulate, while KV cache and checkpoints may also migrate to lower tiers. HDDs and tape remain suitable for high-capacity cold data outside the critical path. They will not achieve HBM-like value per unit, but could benefit from growth in total data volumes and nearline-storage demand.

The supply-chain landscape should be assessed by capability rather than label. Samsung Electronics, SK hynix, and Micron participate in DRAM/HBM; Samsung, SK hynix/Solidigm, Micron, Kioxia Holdings, and SanDisk participate in NAND and enterprise SSDs; Seagate Technology and Western Digital are positioned in HDDs and high-capacity storage; and NVIDIA, Dell, and the DPU and networking ecosystems are advancing CMX and GPU-direct connectivity. This analysis uses only the report’s technical and industry evidence and excludes the point-in-time ratings and price targets on page 2.

11. Emerging Technology Pathways: Every “Breakthrough” Introduces a New Cost

4F² DRAM reduces cell area by changing the cell layout, but Bernstein views this as a one-time structural gain, after which further progress will again depend on lateral scaling. 3D DRAM may offer a genuinely sustainable path, but commercialization remains years away. Process upgrades can alleviate capacity constraints but are unlikely to eliminate the supply cycle entirely.

PIM moves some computation into memory, theoretically reducing data movement between logic and memory and thereby easing bandwidth and power constraints. The challenge is that it departs from the existing von Neumann architecture, requiring processors, software, networks, and manufacturing supply chains to adapt. The report notes that real-world adoption remains extremely limited. PIM is a long-term technology option, but current adoption evidence is insufficient to support assumptions of large-scale revenue.

XBM seeks to build DRAM cells in the back-end interconnect layers of logic chips, while ZAM rotates DRAM dies by 90 degrees to improve cooling. These approaches target bandwidth, stacking, and thermal design, but remain at the patent or proof-of-concept stage. Laboratory peaks demonstrate only technical potential; commercial outcomes will depend on production yields, cost, thermal-cycle reliability, customer software adaptation, and supply-chain replicability.

Edge AI is also producing approaches such as zNAND-O, which stacks NAND using TSVs and places it close to the processor package to provide greater capacity for local models. Edge applications are more sensitive to power, form factor, and cost, while requiring less extreme bandwidth than data centers, and may therefore follow a distinct development path. The report focuses on Transformers and AI systems, and its conclusions should not be extrapolated unconditionally to all edge devices.

12. The Ten Most Important Indicators to Track

To determine whether the AI storage boom is translating from a thematic narrative into earnings, investors should monitor ten indicators.

First, HBM generation qualification and mass-production yields. Customer qualification determines market share; yields determine actual profitability. Second, wafer allocation between HBM and conventional DRAM. If HBM expansion materially crowds out conventional DRAM, pricing elasticity will strengthen; simultaneous supply growth would weaken this effect. Third, CoWoS and hybrid-bonding capacity. Advanced-packaging shortages will defer demand rather than eliminate it.

Fourth, the average context length and number of concurrent sequences in AI inference, which jointly determine KV cache capacity. Fifth, KV cache offload ratios and hit rates across HBM, system DRAM, and SSDs—the core operating metrics for inference unit economics. Sixth, TTFT and TPOT service-level requirements. More demanding real-time applications require higher-end memory and lower-latency interconnects.

Seventh, server penetration of SOCAMM2, MRDIMM, and CXL memory, which will determine whether conventional DRAM can secure structural growth independent of HBM. Eighth, actual customer deployments, sustained IOPS, and tail latency for Storage Next and CMX; peak-performance demonstrations are no substitute for production workloads. Ninth, the media mix in enterprise SSDs. Shifts among SLC/MLC, TLC, and QLC will affect value per unit and gross margins.

Tenth, control of the supply chain. Who designs and manufactures custom base dies, and who supplies the software and networking for direct GPU connectivity, will determine whether incremental profits accrue to memory vendors, XPU vendors, foundry and packaging providers, or systems-equipment companies.

13. Risk Framework: Mass Production, Cost, and Cyclicality Still Separate Strong Demand from High Valuations

The first risk is demand extrapolation. Long-context models, multi-agent systems, and RAG clearly increase memory requirements, but model compression, quantization, sparsity, paged attention, cache reuse, and scheduling optimization can also reduce memory consumption per token. If efficiency improves faster than usage grows, hardware demand will fall short of linear extrapolations.

The second risk is technological maturity. HBF must bridge the latency gap between NAND and DRAM; zHBM faces thermal constraints and hybrid-bonding yield challenges; HBC trades performance for cost; PIM requires an ecosystem overhaul; and 3D DRAM remains far from large-scale commercialization. Supplier roadmaps indicate direction but cannot substitute for evidence of mass production.

The third risk is the pricing and inventory cycle. Structural HBM growth will absorb capital and wafer capacity, but the memory industry remains exposed to delayed capacity additions, inventory corrections, and price elasticity. If several manufacturers expand simultaneously or AI customers slow capital spending, even high-end products could face order adjustments.

The fourth risk is value migration. Custom architectures such as NVHBM could shift the value of base-die and system design to Nvidia and semiconductor foundries; CMX could direct more spending toward DPUs, networking, and data nodes. Growth in the overall market does not guarantee that every traditional memory vendor will achieve corresponding gains in market share and gross margins.

The fifth risk is power consumption and thermal management. Greater capacity and bandwidth increase data-movement requirements and package heat density. SOCAMM2, HBC, and PIM all address energy efficiency, but the report does not provide a complete system-level energy-per-token model and therefore cannot quantify total data-center power consumption or cooling costs.

The sixth risk concerns the scope of the research. This is a technical primer; it does not provide AI server shipments, HBM/DRAM/NAND bit demand, TAM, gross-margin sensitivity, or free-cash-flow sensitivity. The technology roadmap helps identify where value may accrue across the supply chain, but it cannot independently support price targets or earnings forecasts.

14. Final View: Upgrade the Thesis from “Buy HBM” to “Buy Data-Tiering Efficiency”

The investment thesis for memory in the AI era is expanding along two dimensions. The first extends from GPUs to HBM, because compute requires high-bandwidth data delivery. The second extends from HBM to the entire data hierarchy, because long contexts, high concurrency, RAG, and multi-agent systems make capacity, latency, and cost simultaneous constraints.

HBM remains the most direct beneficiary and offers the highest value density. It determines whether the hottest data can keep pace with XPUs and also affects conventional DRAM supply through shared wafer capacity. The value of system DRAM is underappreciated: it supports training pipelines, RAG indexes, CPU workloads, and KV cache offloading. NAND’s opportunity lies in product segmentation. High-performance media can enter Storage Next and CMX; high-capacity QLC and HDDs can absorb colder data; and HBF seeks to move part of the NAND layer closer to the XPU.

Sustainable competitive advantage will depend on combining four capabilities: media performance, advanced packaging, interconnect architecture, and cache software. Any one capability alone is unlikely to convert capacity into low-latency data that GPUs can use. Coordinating all four could increase XPU utilization, support greater concurrency, extend context length, and improve cost per token.

Accordingly, the central question when evaluating AI storage companies is no longer simply, “How many GB can they sell?” A more useful framework asks: Which workload does the company serve? At which tier? How much expensive HBM capacity or GPU waiting time does it save customers? And who captures the resulting value? Viewed through this lens, opportunities in HBM, conventional DRAM, enterprise SSDs, data nodes, interconnects, and cold storage can all coexist—but their realization timelines, margins, and risks differ materially.

Bernstein’s primer does not provide a plug-and-play earnings model, but it offers a more durable analytical framework. The end state of AI memory will not be determined by a single winning medium. It will be shaped by a tiered system capable of continuously placing data in the right location and delivering it to the right processor at the right time.

Related Reading

404K SEMI-AI Evening Brief, August 27, 2026 — Bottlenecks Spill Over: Nvidia Raises Demand Outlook as Memory, Packaging, and Power All Tighten

404K SEMI-AI Evening Brief, 2026-08-19 — Bottlenecks Spill Over: Inference Demand Shifts Pressure to Memory, Packaging, and Power

404K Tech Morning Brief — August 18, 2026 — Repricing the Bottlenecks: Memory Strengthens as Compute Expansion Shifts to Power, Packaging, and Optical Interconnects

404K SEMI-AI Evening Brief, 2026-08-10 — Efficiency Takes the Baton: AI Expansion Shifts Toward Coordination Across Memory, Power, and Interconnects

HBF Standard Takes Shape: How 512GB and 3.0TB/s Turn the AI Inference Memory Wall into a NAND Opportunity

• 1. Robust Revenue Growth, Yet Operating Profit Still Misses Expectations

• 2. AIDC Power Volume Ramps Up, but Profit Realization Remains Anchored in 2027

• 3. Vera Rubin Advances Thesis to Delivery; Liquid Cooling and ASICs Will Dictate the Next Trajectory

• IV. Legacy Businesses Are Still Growing, but Gross Margins Reveal Pricing Pressure

• V. Cash Flow: The Critical Verification Beyond the Profit Narrative

• VI. Near-Term Cuts vs. Long-Term Hikes: The Timing Gap in the Earnings Inflection

• VII. Target Price Raised to RMB 120: Why Goldman Sachs Stays Neutral

• 8. Six Key Metrics to Track Going Forward

本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读

Megmeet has begun delivering Vera Rubin power supplies, moving its AIDC growth thesis into the stage of physical validation. However, 2Q operating profit significantly missed expectations, as legacy business gross margins, cash flow pressures, and an elevated valuation continue to constrain near-term upside.

1. Robust Revenue Growth, Yet Operating Profit Still Misses Expectations

For Megmeet's 2Q results, the metric requiring the closest examination is operating profit rather than top-line growth. Revenue reached RMB 3.273 billion (+39% YoY), missing Goldman Sachs' estimate by just 1%. Gross profit came in at RMB 685 million (+37% YoY), 10% below Goldman Sachs' forecast. Operating profit was only RMB 92 million; despite an 89% YoY increase, it missed Goldman Sachs' estimate by 47% and Visible Alpha consensus by 53%.

Net profit reached RMB 137 million (+107% YoY), missing Goldman Sachs' estimate by a mere 2%. This divergence between operating profit and net profit demonstrates that focusing solely on headline net income risks overstating core operational quality. Goldman Sachs flagged operating profit as the primary shortfall because it directly captures the combined impact of core gross margins and operating expenses.

Second-quarter gross margin was 21%, trailing Goldman Sachs' forecast by 2 percentage points. Operating margin was 3%, also 2 percentage points below estimate, while net margin was 4%, in line with Goldman Sachs' expectations. Rapid revenue growth failed to convert proportionally into operating profit, as incremental product mix, margin erosion in legacy segments, and cost-passthrough friction weighed on near-term earnings elasticity.

Operating expense discipline demonstrated emerging economies of scale. The SG&A; expense ratio dropped to 16%, down 2 percentage points YoY and 1 percentage point QoQ. In other words, while operating leverage exists, the benefits of lower overhead were largely offset by gross margin pressure. Going forward, only when gross margins stabilize can expense reductions translate visibly into profit growth.

These figures establish the report's first key takeaway: while incremental revenue from AIDC has materialized, profit conversion remains in its early stages. Investors must evaluate top-line growth and earnings quality separately rather than assuming they will materialize in tandem.

Second-quarter revenue missed consensus by 18%, whereas operating profit missed by 53%—a pronounced negative leverage effect. Top-line variance explains only part of the earnings gap, with gross margins and business mix exerting far greater drag. To rebuild market expectations, sustaining high top-line growth alone will not suffice; gross profit and operating profit growth must realign with revenue momentum.

Net profit missed consensus by 41%, also reflecting a wide gap with operating profit. Because the report does not break down all non-operating items, it is unclear which specific gains offset the shortfall. A more reliable analytical approach is to treat operating margin as the true gauge of core operating health and net profit as the bottom-line output; a genuine inflection in earnings quality requires concurrent improvement across both metrics.

In the upcoming quarter, if revenue continues to grow while gross margins remain stuck around 21%, lower expense ratios will provide only a limited buffer. Only if gross margins rebound alongside an expanding operating margin will it signal that an improved AIDC product mix or legacy price adjustments are taking effect. This serves as the primary benchmark to determine whether 2Q's profit gap was a transient fluctuation or a structural headwind.

2. AIDC Power Volume Ramps Up, but Profit Realization Remains Anchored in 2027

The power supply segment is currently Megmeet's strongest growth engine. In 1H26, power supply revenue reached RMB 1.841 billion (+59% YoY), with 2Q revenue at RMB 1.024 billion (+54% YoY). While Goldman Sachs clearly identifies AIDC power supplies as the core driver, the report notes that the company has not disclosed absolute AIDC power revenue.

This informational boundary is critical. Power supply segment revenue cannot be equated directly with AIDC revenue, nor does rapid segment growth prove that Vera Rubin has entered high-volume mass production. Although the market has visibility into the new strategic direction, it still lacks verifiable quarterly data on AIDC revenue, shipment volumes, customer mix, and product breakdown.

Goldman Sachs' projections illustrate why the market is focused on this segment: the firm expects Megmeet's AIDC power supply revenue to reach RMB 1.7 billion in 2026 and RMB 5.2 billion in 2027. This represents a near-tripling of revenue within one year, lifting its share of total revenue from 13% to 29%.

AIDC's impact on profit structure could be even more decisive. Goldman Sachs projects the business to contribute 20% and 41% of total gross profit in 2026 and 2027, respectively. The fact that the 2027 gross profit share significantly exceeds its revenue share implies that, if realized as projected, the AIDC mix will become a key driver of corporate margins and operating leverage.

A 29% revenue share versus a 41% gross profit contribution carries a clear implication: by 2027, Goldman Sachs' model relies on AIDC to drive the company's profit mix upgrade. Even if legacy businesses continue to expand, without a corresponding recovery in their gross margins, overall profit improvement will become increasingly dependent on AIDC.

This structural shift requires four conditions to hold simultaneously: customer orders must convert into shipments on schedule, manufacturing yields must support mass production, dollar-per-watt pricing cannot deteriorate rapidly, and component supply must remain stable. Deviation in any of these factors could delay the RMB 5.2 billion revenue or the 41% gross profit contribution.

Goldman Sachs did not disclose customer concentration or platform share, noting only that Megmeet has entered the NVIDIA ecosystem and is advancing custom ASIC projects with multiple US cloud service providers (CSPs). This confirms expanding pipeline opportunities, but does not prove the company has locked in a sticky, long-term market share. Re-qualification cycles across platform upgrades will continue to dictate competitive positioning in next-generation products.

This explains why Goldman Sachs can maintain near-term caution while upgrading long-term forecasts. The year 2026 remains a period of order conversion, product validation, and capacity ramp-ups, where incremental revenue may not immediately offset gross margin pressure in legacy segments. Only by 2027, if AIDC revenue and gross profit shares approach forecast levels, will the overall profit structure undergo a meaningful transformation.

The assumption of high AIDC gross margins is also conditional. The report notes that price-per-watt pricing is sensitive to near-term supply-demand dynamics and component availability. As product volumes scale, increased competition, client pricing power, and supply easing could alter pricing. Current forecasts reflect Goldman Sachs' assessment of order visibility, product mix, and operating leverage rather than locked-in profitability.

3. Vera Rubin Advances Thesis to Delivery; Liquid Cooling and ASICs Will Dictate the Next Trajectory

Progress on Vera Rubin has shifted Megmeet from merely "entering the AI server power supply chain" to "commencing initial deliveries." The company has delivered air-cooled Vera Rubin power supplies to customers, with liquid-cooled versions scheduled for 4Q26 to 1Q27. However, the report does not disclose air-cooled shipment volumes; thus, current visibility confirms supplier status rather than the magnitude of revenue contribution.

Management expects new AIDC orders to grow significantly sequentially in 2H26 and accelerate rapidly in 2027. Within the product mix, demand share for Vera Rubin is expected to rise, while contributions from GB200 and GB300 decline. This shift underscores that Megmeet must continually pass qualification cycles on new-generation platforms, as legacy platform orders do not automatically convert into next-generation share.

Megmeet is also advancing custom ASIC projects with multiple US CSPs, involving ~5.5 kW power supply solutions. While ASIC initiatives could reduce reliance on single GPU platform release cycles, customer validation, mass-production yields, and datacenter buildout schedules will continue to influence quarterly shipments. A time lag remains between order placement and recognized revenue.

The 800V DC architecture provides a longer-term product roadmap. Megmeet is developing power racks, 800V-to-50V power shelves, and DC-DC power bricks, with plans to introduce solid-state transformers in 1H27. However, management notes that because safety standards remain incomplete, 800V DC is unlikely to become mainstream across the industry by 2027, and widespread adoption of solid-state transformers may not occur until after 2029.

This timeline reminds investors to distinguish among three phases: sampling and qualification, initial deliveries, and high-volume mass production. While Megmeet has crossed the first two stages for certain products, commercial-scale mass production still requires verification via quarterly revenue and shipment data. Any estimate that equates "initial deliveries" directly with "meaningful profit contribution" overlooks customer deployment schedules and volume ramp execution risks.

The path from orders to profit is also mediated by customer datacenter buildout timelines. The report explicitly highlights that quarterly shipments and sales will depend on facility construction progress. Even if underlying annual demand remains intact, fluctuations in facility handovers, server deployments, and customer acceptance can cause noticeable quarterly revenue volatility.

Product efficiency and portfolio breadth will determine whether the company can defend its market share. Goldman Sachs notes that Megmeet's high-efficiency product lineup still lags behind Tier-1 incumbents like Delta Electronics. As new entrants proliferate, competition will evolve from "supply availability" to whether a vendor can simultaneously meet benchmarks across efficiency, reliability, delivery, and cost.

The 800V DC architecture and solid-state transformers are best viewed as long-term call options rather than foundational revenue drivers for 2027. Management itself has not brought forward its timeline for industry-wide 800V DC adoption. Factoring this revenue into near-term valuations prematurely conflates product introduction with commercial adoption at scale.

IV. Legacy Businesses Are Still Growing, but Gross Margins Reveal Pricing Pressure

Megmeet's legacy businesses are not in broad contraction; rather, the core challenge lies in the quality of growth. In 1H26, home appliance revenue grew 19% YoY, accelerating to 40% YoY in 2Q26. While recovering overseas demand provided support, domestic momentum was constrained by a soft real estate market.

The new energy vehicle (NEV) and rail transit components segment grew revenue by 11% and 9% YoY in 1H26 and 2Q26, respectively, while industrial automation grew 15% and 13% YoY over the same periods. Although all three legacy segments maintained positive top-line growth, their 1H26 gross margins came in at 18%, 13%, and 28%—contracting by 4, 1, and 1 percentage points YoY, respectively.

Rising raw material costs, sluggish domestic demand, and RMB appreciation have combined to pressure margins. Although Megmeet issued a company-wide price increase notice in June, customer pushback has been relatively mild overseas but significantly stiffer domestically, particularly within the automotive sector. The price hike notice itself does not guarantee an immediate margin rebound, as actual realization depends on customer acceptance and the timing of contract renewals.

These legacy operations will dictate whether AIDC growth successfully translates into corporate-level profit expansion. If legacy gross margins continue to slide, incremental gross profit from AIDC will merely plug legacy earnings deficits. Only if price increases take hold and raw material cost pressures subside will incremental AIDC contributions lift company-wide margins.

The proportion of overseas revenue represents another critical variable. Management expects roughly 20% of AIDC orders from 2H26 through 2027 to come from China and 80% from overseas, targeting an increase in overall overseas sales to 50%–60% by 2027. While overseas clients present less friction on pricing pass-through, expanding international operations introduces heightened demands on customer qualification, foreign exchange management, and global fulfillment.

There are three potential pathways for legacy business recovery: first, a rebound in domestic demand that improves capacity utilization and product mix; second, gradual price increases that offset raw material inflation; and third, an expanding overseas revenue share that leverages smoother cost pass-through to alleviate margin pressure.

These three pathways operate on differing timelines. Raw material prices and currency exchange rates can fluctuate rapidly, whereas customer price adjustments require contractual renewal cycles, and domestic demand recovery remains dependent on broader macroeconomic conditions. Upgrading full-year gross margin expectations based solely on a single price-hike announcement overlooks these timing lags.

The 40% YoY revenue surge in home appliances in 2Q26 alongside a 4-percentage-point YoY drop in 1H26 gross margin underscores that rapid top-line growth does not inherently deliver stronger profitability. In upcoming segment disclosures, revenue growth and gross margins must be assessed in tandem; viewing either metric in isolation risks reaching premature conclusions regarding fundamental recovery.

V. Cash Flow: The Critical Verification Beyond the Profit Narrative

In 1H26, Megmeet generated just RMB 30mn in operating cash flow (OCF), down sharply from RMB 192mn in the prior-year period. Over the same timeframe, capital expenditures reached RMB 190mn, up 9% YoY, as the company continued expanding manufacturing capacity across China and Thailand while building R&D; and testing facilities for AIDC products.

The report did not provide a full breakdown of the OCF decline, making it impossible to assign the drop entirely to accounts receivable, inventory, or any single line item. What is clear is that the company is going through a phase where accounting profit and capital deployment are disconnected. Rapidly expanding AIDC orders are likely front-loading cash requirements for raw material procurement, qualification, capacity expansion, and customer fulfillment.

Operating cash flow does not need to match net profit in every individual quarter, but sustained divergence raises the cost of expansion. Because Megmeet's forward valuation prices in aggressive volume scaling, cash conversion cannot remain absent indefinitely. Going forward, key metrics to track include the OCF-to-net-profit ratio, working capital shifts across inventory and receivables, and whether capital investments generate visible output.

Cash flow also dictates valuation margin of safety. If earnings growth relies on persistent capital outlays, enterprise value will appreciate at a slower pace than the income statement suggests. Only if OCF improves substantially as AIDC revenue ramps up can the market confirm that this is a scalable, cash-generative business rather than growth existing solely within order books and earnings models.

In 1H26, OCF covered only ~16% of capex. While this does not directly indicate poor return on investment—since new capacity and R&D; facilities require lead time before generating revenue—it highlights that free cash flow is under significant strain, placing elevated demands on financing capabilities and working capital management during this expansionary phase.

Leading indicators for cash flow can be divided into two clusters: working capital metrics (inventory levels, accounts receivable, and cash collections) and capex conversion (when capacity under construction transfers into fixed assets and begins shipping). The report lacked these granular details, leaving upcoming earnings reports to provide the requisite evidence.

If AIDC revenue ramps alongside a corresponding rebound in OCF, the capacity expansion thesis will gain substantial credibility. Conversely, if revenue expands while cash flow continues to deteriorate, investors will need to determine whether customer payment terms, inventory pre-stocking, and capacity ramp-ups are tying up capital over the long run—a dynamic that erodes the enterprise value created by each dollar of reported net income.

VI. Near-Term Cuts vs. Long-Term Hikes: The Timing Gap in the Earnings Inflection

Goldman Sachs trimmed its 2026 and 2027 net profit forecasts for Megmeet by 10% and 6%, respectively. The downgrades reflect weaker-than-expected legacy operations, sustained gross margin headwinds, and the reality that initial AIDC volume ramps remain insufficient to fully offset these legacy drags.

Concurrently, Goldman Sachs raised its 2028–2030 net profit forecasts by an average of 11%, driven by expanding AIDC scale and operating leverage. These revisions delineate a clear timing divergence: visibility over the next two years has weakened, even as the long-term earnings ceiling has been lifted.

Goldman Sachs forecasts Megmeet's revenue to reach RMB 12.661bn in 2026, RMB 17.643bn in 2027, and RMB 25.898bn in 2028, with corresponding EPS of RMB 0.98, RMB 2.35, and RMB 4.16. Both top-line revenue and EPS require two consecutive years of hyper-growth to gradually compress the current valuation.

The stock trades at projected P/E multiples of ~128.2x for 2026 and 53.7x for 2027, before easing to 30.3x in 2028. Valuation multiple compression depends almost entirely on denominator expansion—specifically, rapid EPS acceleration. Any delay across order conversion, product mix, gross margins, or mass-production timelines will lengthen the multiple digestion period.

Based on Goldman Sachs' projections, revenue is set to grow ~39% YoY in 2027 and another ~47% YoY in 2028. EPS growth is modeled even faster, surging ~140% YoY in 2027 and ~77% YoY in 2028. The model implicitly assumes that profit growth will significantly outpace revenue expansion, predicated on product mix optimization and operating leverage.

This earnings trajectory is highly sensitive to gross margins. If revenue meets expectations but gross margins undershoot model assumptions, operating expense dilution may not suffice to offset the gap. Conversely, if AIDC gross profit contribution rises while legacy margins stabilize, earnings growth could significantly outrun revenue growth.

The downward revisions to near-term estimates also steepen the climb toward the 2028 target. With the 2026 EPS baseline lowered to RMB 0.98, reaching RMB 4.16 over the subsequent two years requires a more than three-fold cumulative increase. While achievable, any quarterly execution slippage will be amplified across long-term valuation multiples.

These projections also establish explicit criteria for thesis falsification. If 2027 AIDC revenue falls materially below RMB 5.2bn, or if its gross profit contribution fails to approach 41%, the upward revisions to long-term earnings will require re-evaluation. Furthermore, if legacy gross margins fail to bottom out, incremental AIDC profit gains will continue to be diluted.

VII. Target Price Raised to RMB 120: Why Goldman Sachs Stays Neutral

Goldman Sachs raised its 12-month target price from RMB 112 to RMB 120, applying a 32x 2028E P/E multiple discounted back to 2027 at an 11% cost of equity. The higher target price primarily reflects upgraded long-term earnings forecasts rather than increased optimism regarding near-term operating performance.

Prior to the report's release, Megmeet closed at RMB 125.95, roughly 4.7% above Goldman Sachs' target price. Consequently, the target price increase offers no upside headroom. Goldman's Neutral rating signals that growth upside is roughly balanced against rich valuation multiples and execution risks.

A 32x 2028E P/E multiple embeds aggressive growth assumptions. The median 2026E and 2027E P/E across Goldman Sachs' China Industrial Tech coverage universe sits at ~26x and ~23x, compared to ~128x and ~54x for Megmeet over the same periods. Megmeet must deliver significantly faster earnings growth simply to digest this steep valuation premium.

Megmeet's 2026E P/E is nearly five times the coverage median, and remains more than double in 2027E. This premium prices in global AIDC supply chain entry, Vera Rubin deliveries, and long-term margin expansion. If only a fraction of these catalysts materialize, the valuation premium will lack sufficient fundamental earnings support.

Valuing the company on 2028 earnings and discounting back to 2027 creates substantial reliance on long-dated assumptions. As the forecast horizon extends, uncertainties around order volumes, competition, pricing, and technology roadmaps increase significantly. While an 11% cost of equity accounts for the time horizon and execution risks, it cannot eliminate the forecasting error embedded in the earnings estimates themselves.

The risk-reward profile can be categorized into three scenarios: a bull case where Vera Rubin and ASIC projects ramp faster than anticipated and drive an earlier mix shift toward AIDC gross profits; a base case where AIDC expands as scheduled but persistent legacy headwinds delay overall profit improvement; and a bear case where qualification, mass production, or customer rollout timelines slip alongside valuation multiple contraction.

Goldman Sachs did not provide standalone target prices for each scenario, outlining only the primary valuation drivers. Consequently, RMB 120 is best interpreted as the output of a specific set of baseline assumptions rather than a fixed anchor immune to fundamental shifts. Maintaining a Neutral rating reflects this balance between conflicting forces.

The valuation framework hinges on three key variables: first, whether 2028E EPS of RMB 4.16 is delivered; second, whether the market remains willing to assign a 32x forward P/E; and third, whether an 11% cost of equity sufficiently reflects customer concentration, mass-production, and technology roadmap risks. Any compression in these assumptions would directly impair the target price.

A Neutral rating does not represent a rejection of the AIDC business. Rather, it indicates that positive drivers must translate into quantifiable revenue, earnings, and cash flow to support the growth expectations already embedded in the stock price. With the market already paying a premium for long-term growth, every subsequent quarter will serve as a critical checkpoint for execution.

8. Six Key Metrics to Track Going Forward

Evaluating Megmeet's forward trajectory hinges on six sets of publicly reported figures rather than unverified order flows and technical narratives.

First, absolute AIDC power supply revenue and its share of total power supply revenue. Current reports only identify AIDC as the primary growth driver without disclosing absolute figures. This metric will determine whether the RMB 1.7 billion and RMB 5.2 billion revenue forecasts materialize on schedule.

Second, quarterly shipments of air-cooled Vera Rubin products, along with the validation and mass-production status of liquid-cooled versions in 4Q26 to 1Q27. Product launches, customer qualification, and revenue recognition must be evaluated as separate milestones.

Third, power supply segment gross margin and corporate blended gross margin. If AIDC's gross profit contribution increases, blended gross margin should gradually find support. Conversely, if power supply revenue grows while blended gross margin continues to compress, legacy business margin pressure and AIDC pricing dynamics must be closely scrutinized.

Fourth, the effective implementation of price hikes across home appliances, NEVs/rail transit, and industrial automation. Post-hike gross margin trajectory is far more informative than price-increase notices alone, with domestic customer acceptance being particularly critical.

Fifth, the relationship among operating cash flow, capex, and net income. If top-line expansion is accompanied by persistent cash flow weakness, the quality of business expansion warrants caution. Operating cash flow must improve alongside the AIDC volume ramp to prove that order growth is converting into realized economic returns.

Sixth, execution against 2027 revenue and EPS forecasts. Goldman Sachs' projected RMB 17.643 billion in revenue and RMB 2.35 in EPS represent the critical stepping stones for valuation to compress from 128x to 54x P/E. Any significant deviation will disrupt the market's pace of digesting current multiples.

This checklist places technology, orders, and financials along a single validation chain. Improvement in any single metric is insufficient to confirm the overarching thesis: expanding air-cooled shipments must translate into revenue, revenue growth must flow through to gross profit, and gross margin expansion must ultimately convert into cash flow.

The primary value of Goldman Sachs' report lies in mapping near-term earnings gaps and long-term product opportunities onto the same timeline. Second-quarter results show that the company has not yet escaped legacy business headwinds, while Vera Rubin deliveries demonstrate tangible progress in its new growth vector. Both realities coexist, and any conclusion focusing solely on one remains incomplete.

Investors should dynamically update their assessment over the coming quarters rather than placing an all-or-nothing bet on 2028 outcomes. If AIDC revenue, liquid-cooled product milestones, segment gross margins, and operating cash flows improve in sequence, the case for forward earnings upgrades will strengthen. Conversely, if two or more of these metrics consistently disappoint, market expectations for both growth velocity and valuation multiples must be adjusted downward.

Megmeet has secured a critical entry ticket into next-generation AI server power supply competition, and the commencement of air-cooled Vera Rubin deliveries marks substantial progress. However, the company's ultimate intrinsic value depends on whether this entry ticket can consistently translate into sustained orders, gross profit, and cash flow. Legacy gross margin drag, volume execution risks, and elevated valuation leave little margin for error—explaining why Goldman Sachs maintained a Neutral rating even after raising its long-term price target.

Related Reading

404K SEMI-AI Morning Brief — August 14, 2026 — Demand Broadens: AI Orders Accelerate, but Profitability and Delivery Remain Unproven

404K SEMI-AI Evening Brief 2026-08-05 — Broadening Constraints: AI Demand Remains Strong, While Gross Margins and Cash Flow Begin to Differentiate Company Performance

404K SEMI-AI Morning Brief, 2026-08-12 — Financing Broadens: $500 Billion Platforms Drive Compute Buildout, While Power and Utilization Determine Returns

404K SEMI-AI Morning Brief 2026-07-31 — Returns Begin to Diverge: Improving Cloud Revenue, With Power and Storage Constraints Determining the Pace of Monetization

404K SEMI-AI Evening Brief, August 27, 2026 — Bottlenecks Spill Over: Nvidia Raises Demand Outlook as Memory, Packaging, and Power All Tighten

• I. The Bottom Line: The Same Model Business Can Produce Two Very Different Income Statements

• II. The First Hurdle: Put Gross versus Net Reporting and ARR Definitions on a Common Basis

• III. Three Product Models: Subscriptions, Direct APIs, and Indirect APIs Monetize Differently

• IV. The 2026 Margin Inflection: Lower Costs Matter More Than Higher Model Prices

• 5. Why Cloud Providers Remain Major Winners: They Monetize Compute, Distribution, and Delivery Certainty

• 6. Training Costs Change the Conclusion: High Inference Margins Do Not Guarantee High Free Cash Flow

• 7. From $7 Billion to $690 Billion: The Forecast’s Structure Matters Most

• 8. The 2028 Inflection Point: Labs Will Move More Compute Beyond General-Purpose Public Clouds

• IX. Who Has the Edge: Focus on Contracts, Distribution, and Utilization—not a Single Model Leaderboard

• X. The Strongest Counterevidence and Key Metrics to Track: Five Sets of Numbers Will Determine Whether the Framework Holds

本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读

The defining shift in 2026 is the sharp decline in unit inference costs: AI labs are rebuilding gross margins, while cloud providers continue to capture high-margin compute revenue. Winners will be determined by their business mix, accounting treatment, and the infrastructure transition around 2028.

I. The Bottom Line: The Same Model Business Can Produce Two Very Different Income Statements

AI lab unit economics have entered a new phase. Barclays estimates that inference compute accounted for $61—79 of every $100 of revenue in 2025; by 2026, that figure had fallen to approximately $30. Paid-inference margins rose from 14%—18% to 48%—65%. After deducting the training costs of production models, the two hypothetical labs achieved blended gross margins of 38% and 55%, respectively.

These figures show that improvements in model capability are beginning to translate into commercial efficiency. Quantization, speculative decoding, batching, caching, routing optimization, and higher utilization among enterprise customers have collectively reduced the compute cost of each effective task. Model prices have not declined nearly as sharply, allowing the cost savings to accrue first to lab gross margins. Growth in enterprise agent workflows has also improved the customer mix: high-utilization, billable workloads with defined usage caps now represent a larger share, reducing the drag from free users and inefficient interactions.

The same underlying improvement does not automatically produce comparable financial statements. Revenue recognition, channel costs, and partner revenue shares are treated differently across subscriptions, direct API sales, and indirect API sales through cloud marketplaces. One lab may recognize cloud-marketplace transactions on a gross basis, while another records only the amount net of channel fees. One may classify strategic-partner revenue sharing as cost of revenue, while another may not recognize revenue hosted by its cloud partner at all. Reported revenue growth, gross margin, and R&D; expense ratios therefore cease to be directly comparable.

Investors should compare three layers of economics: what customers pay for model services; what the lab retains after inference and channel costs; and how much sustainable profit and cash flow remains after production-model and R&D; training expenses. GAAP financial statements remain important, but the underlying contracts must first be mapped onto the same value chain.

II. The First Hurdle: Put Gross versus Net Reporting and ARR Definitions on a Common Basis

Accounting differences can make two operationally similar companies appear to be at entirely different stages of development. Barclays illustrates the issue with Lab A and Lab B. The two labs serve similar customer demand and incur comparable underlying inference costs, yet differences in revenue recognition and strategic-partnership terms produce wide gaps in reported revenue, costs, and gross margins.

Indirect API sales are the most prone to misinterpretation. When customers purchase model services through AWS, Azure, or Google Cloud, the cloud provider handles billing, sales, and customer-relationship management. If the lab is the principal in the transaction, it may recognize $100 of gross revenue and then pay $20 in channel or cloud fees. If the cloud provider is the principal, the lab may recognize only $80 of net revenue. Customer spending is identical, but the lab’s reported revenue differs by $20.

Strategic partnerships introduce another layer of divergence. Barclays assumes that Lab B pays a strategic partner a 20% revenue share, while Lab A has no such provision. Lab B’s underlying product economics may be no worse, but the revenue share reported on its income statement reduces both paid-inference margin and blended gross margin. If partner-hosted revenue is excluded entirely from Lab B’s financial statements, the market may also underestimate its true end demand and model usage.

ARR also requires disaggregation. Annualized subscription contract value, current-period API consumption, committed cloud-marketplace spending, and strategic-partner-hosted revenue may all be annualized differently. Page 2 of the report even uses customer spending of $101 and $102 to show that the two labs define $100 of reported revenue differently. Comparing disclosed ARR directly can therefore mistake revenue-recognition differences for shifts in market share.

A more reliable framework tracks four metrics: end-customer spending, paid Tokens or effective task volume, contribution profit after inference costs, and cash allocation between the lab and its cloud partners. Once these four measures are aligned, the apparent gap created by gross versus net revenue reporting narrows substantially. As AI labs begin providing more comprehensive GAAP disclosures, the market may initially enter a phase in which more data paradoxically reduces comparability.

III. Three Product Models: Subscriptions, Direct APIs, and Indirect APIs Monetize Differently

Subscriptions sell reliable access and a consistent user experience. Users pay a fixed monthly or annual fee, while the lab absorbs actual Token consumption. Tools such as Claude Code and Codex control costs through usage allowances, concurrency limits, model routing, and enterprise contract caps. Barclays estimates that mature subscription products generate inference margins of approximately 70%. This is already a high level, but it remains sensitive to power users, free trials, quota resets, and customer-acquisition subsidies.

The subscription model offers predictable revenue, strong product stickiness, and a natural entry point into customer workflows. Its risk is asymmetric usage: a small group of intensive users can consume substantial compute, while fixed pricing cannot immediately reflect the associated cost. Labs must continually adjust pricing tiers, rate limits, and model segmentation. Enterprise contracts with explicit usage caps and overage charges should deliver more stable margins than unlimited consumer plans.

Direct APIs monetize measurable model calls. Developers such as Cursor and Figma purchase Tokens directly from the lab, with pricing tied to usage. Barclays says paid-inference margins for some direct APIs exceeded 80% in Q2 2026, although the report uses a more conservative 70% assumption. Direct APIs translate inference-efficiency gains into profit more quickly because unit pricing is transparent and cost savings do not need to be passed on to customers immediately.

These high margins still have limits. API pricing may fall as competition among frontier models intensifies; routing to open-source and smaller models will reduce the share of calls handled by premium models; and larger customers will demand volume discounts. Current margins include premiums derived from scarce compute, superior capabilities, and switching costs. Long-run normalized margins will probably be below peak levels.

Indirect APIs monetize channel reach. Cloud providers integrate models into their existing procurement, compliance, billing, and sales systems, helping labs reach large enterprises. Labs exchange channel fees or revenue sharing for lower customer-acquisition costs and faster scaling. The economics depend on bargaining power: when models are scarce, labs can retain a larger share; as models become more substitutable, cloud providers can more easily compress that share and control the customer relationship.

Business mix therefore determines the shape of the income statement. Barclays assumes that Lab A’s revenue mix is 30% subscriptions, 45% direct API (API45), and 25% indirect API (API25); for Lab B, the mix is 80% subscriptions, 10% direct API (API10), and 10% indirect API (API10). Both incur approximately $30 of inference costs, but Lab A pays higher channel fees without a strategic-partner revenue share, generating $65 of paid-inference profit. Lab B pays a $20 partner revenue share and generates only $48. Despite similar product efficiency, contract structures create a 17-percentage-point margin gap.

IV. The 2026 Margin Inflection: Lower Costs Matter More Than Higher Model Prices

The shift from 2025 to 2026 was driven almost entirely by lower inference costs. Lab A’s inference compute cost per $100 of revenue fell from $79.1 to $30, while Lab B’s declined from $61.1 to $30. Their paid-inference margins rose from 14% and 18% to 65% and 48%, respectively. Even after deducting $10 of production-model training costs, blended gross margins increased from 4% and 8% to 55% and 38%.

Cost improvement has followed three paths. First, greater model-level efficiency means that the same task requires fewer Tokens or can achieve similar quality at lower precision. Second, serving-layer gains—including batching, caching, parallel scheduling, speculative decoding, and higher chip utilization—reduce compute consumption per Token. Third, the demand mix has improved: enterprise APIs and agent tasks carry higher paid-usage rates, reducing the drag from free interactions on the revenue denominator.

Agent workflows are especially important. Ordinary chats often involve only a few rounds of interaction, whereas agents repeatedly call models, tools, and external systems. Each task consumes more Tokens, and customers are more willing to pay for measurable productivity gains. Higher call volumes increase compute demand, but contribution profit can continue expanding as long as efficiency improves faster than unit prices decline.

The margin inflection should not be extrapolated mechanically. The report also cautions that current levels may exceed the long-run steady state. Competition at the model frontier will push prices lower, expanding compute supply will erode scarcity premiums, and larger enterprise customers will negotiate discounts. Efficiency gains will ultimately be redistributed among labs, cloud providers, and customers. In the near term, they accrue first to lab income statements; over time, some will likely be passed on to customers.

Testing this thesis requires tracking whether unit costs are falling faster than unit prices. Public disclosures may provide five relevant signals: price per million Tokens, Token consumption for the same benchmark task, cache-hit rates or batching ratios, inference compute cost as a percentage of revenue, and the margin differential between APIs and subscriptions. If call volumes continue to rise but the inference-cost ratio stops declining, efficiency gains may be approaching a plateau—or price competition may be consuming the cost savings.

5. Why Cloud Providers Remain Major Winners: They Monetize Compute, Distribution, and Delivery Certainty

Improving gross margins at AI labs have not diminished the value of cloud providers, at least at this stage. Barclays estimates that roughly $35—$40 of every $100 in lab revenue will flow to hyperscalers in 2026, generating $10—$20 in operating profit for the cloud providers and implying an operating margin of approximately 35%—45%. Partnership terms may alter the accounting allocation, but the underlying Token economics could be similar.

Cloud providers capture returns in three ways. First, IaaS revenue, as labs directly purchase GPUs, networking, and storage. Second, distribution revenue, as cloud marketplaces resell model APIs and charge fees. Third, strategic-partnership revenue sharing, through which cloud providers exchange capital, compute commitments, and go-to-market access for a share of lab revenue. All three may coexist, making it difficult to isolate AI’s contribution from headline cloud growth alone.

What cloud providers really sell is delivery certainty. Training and serving frontier models require large-scale clusters, reliable networks, sufficient power and cooling, and security and compliance frameworks already accepted by global enterprise customers. Labs can optimize their models, but they cannot replicate this infrastructure quickly. Cloud providers can also share data centers, power procurement, and sales teams across AI demand and traditional cloud workloads, lowering marginal customer-acquisition and operating costs.

This advantage still depends on utilization. GPU procurement and data-center construction incur capital expenditure upfront, while revenue is realized gradually. If lab demand falls short of commitments, cloud providers may have contractual protection; if end-customer demand falls short of lab expectations, the entire value chain faces low utilization and depreciation pressure. A high operating margin does not automatically translate into a high return on capital, particularly when investment in new data centers, power, and networking grows faster.

Competitively, cloud providers with frontier-model partners, ample power, and mature enterprise distribution are better positioned to capture AI revenue. The report assumes that by 2028, AWS and Google Cloud will each account for approximately 30% of hyperscaler AI revenue, with other clouds representing the remaining 40%. Azure is not shown separately and is included in other clouds. Readers should therefore not interpret “other clouds” as a collection of small providers; it may include platforms as large as Microsoft.

6. Training Costs Change the Conclusion: High Inference Margins Do Not Guarantee High Free Cash Flow

Inference gross margin answers only how much profit remains from serving one additional paying customer. It does not capture the full cost of maintaining model competitiveness. Training next-generation models, acquiring data, hiring researchers, absorbing failed experiments, and optimizing models before deployment all consume substantial cash. Barclays assumes that approximately 10% of training costs relate to the final production model and are recognized in cost of sales, while the remaining 90% is recorded as R&D; expense.

This allocation directly affects gross margin. The more production-training costs recognized in cost of sales, the lower the gross margin; allocating more training costs to R&D; makes gross margin appear higher, although operating profit is still affected. Labs may classify model training, post-training, fine-tuning, and production deployment differently. Gross-margin comparisons must therefore be assessed alongside R&D; expense, capitalization policies, and cash expenditure.

Barclays’ industry model projects training costs rising from $7 billion in 2024 to $210 billion in 2028. As a share of AI lab revenue, training costs decline from 96% to 30%, indicating that their fixed-cost characteristics become more evident at scale. Absolute spending, however, continues to grow rapidly. Even if inference generates high contribution margins, the continuing race to develop frontier models could absorb most of the resulting cash.

Inference and training costs also follow different cycles. Inference costs rise with usage, while unit costs can decline through higher utilization and Serving optimization. Training costs are more step-like: each new model generation requires new clusters, data, and R&D; investment. If gains in model capability begin to slow, the marginal return on training investment will decline. Genuine operating leverage emerges only if application revenue grows fast enough for training costs to fall as a share of revenue.

High-quality earnings at AI labs therefore require three conditions: stable margins on paid inference; training costs growing more slowly than revenue; and R&D; investment that produces durable model differentiation or product stickiness. Looking only at API gross margins risks overstating cash-generation capacity, while focusing only on R&D; losses may understate the contribution-profit foundation already established by inference.

7. From $7 Billion to $690 Billion: The Forecast’s Structure Matters Most

Barclays expects total AI lab revenue to rise from $7 billion in 2024 to $26 billion in 2025, $137 billion in 2026, $376 billion in 2027, and $690 billion in 2028. The corresponding growth rates are 276%, 418%, 175%, and 84%. Year-end AI ARR is projected to increase from $44 billion in 2025 to $782 billion in 2028.

The absolute forecasts are highly aggressive, but the structural changes are more analytically useful. In 2026, accelerating revenue growth coincides with improving unit economics. Revenue grows 418%, while the inference cost ratio falls from 66% in 2025 to 42% and the training cost ratio declines from 70% to 48%. The combination of scale and efficiency explains the rapid recovery in lab gross margins.

Revenue continues to grow rapidly in 2027—2028, although growth begins to decelerate. The inference cost ratio remains at 42%, implying that subsequent efficiency gains are largely passed through in pricing or consumed by more complex workloads. The training cost ratio continues to decline from 35% to 30%, providing the principal source of operating leverage. If the actual inference cost ratio is below 42%, lab profits could exceed the forecast; if price competition intensifies, both revenue and gross margin will come under pressure.

Hyperscaler AI revenue is projected to rise from $11 billion in 2024 to $36 billion in 2025, $124 billion in 2026, $289 billion in 2027, and $502 billion in 2028. As a share of AI lab revenue, it declines from 153% to 73%. The ratio exceeds 100% in the early years because of training compute, prepaid commitments, and revenue not yet recognized by the labs; it gradually falls as end-market lab revenue expands.

This ratio trajectory matters more than any single revenue figure. It describes the AI value chain’s transition from “build compute first, generate application revenue later” toward a model in which application revenue covers infrastructure investment. The declining ratio does not mean cloud AI revenue is shrinking—it continues to grow. Rather, cloud spending per $1 of lab revenue gradually declines as lab efficiency and bargaining power improve.

8. The 2028 Inflection Point: Labs Will Move More Compute Beyond General-Purpose Public Clouds

Barclays views 2028 as a critical inflection point in infrastructure architecture. The report defines “floor-backed AI infrastructure” as compute projects built by labs, capital partners, or specialized infrastructure providers, with demand underwritten through long-term purchases, minimum-usage commitments, or financing arrangements. These projects can reduce labs’ dependence on general-purpose public clouds.

Scale is the key driver. When lab usage is limited, public clouds offer clear advantages in flexibility, deployment speed, and global reach. Once spending reaches billions of dollars, fixed procurement, dedicated clusters, and long-term power contracts may deliver lower unit costs. Labs also want greater control over chip configurations, network architecture, and scheduling software while reducing cloud distribution fees and strategic revenue sharing.

The migration will not happen all at once. Training, peak inference demand, regional compliance, and enterprise cloud marketplaces will continue to require public clouds. A hybrid structure is more likely: large, stable baseload workloads will move to dedicated facilities, while highly variable, geographically dispersed workloads—or those requiring enterprise distribution—will remain on hyperscale cloud platforms. Cloud providers’ share may decline even as their absolute revenue continues to grow.

Dedicated infrastructure also introduces new risks. Labs must assume longer asset cycles, financing costs, and fluctuations in utilization. If models or products evolve faster than data centers depreciate, hardware may become obsolete prematurely. If end demand falls short of expectations, minimum-purchase commitments will shift from revenue protection for cloud providers to fixed obligations for labs.

Cloud providers can respond by lowering unit prices, offering custom chips, strengthening model-marketplace distribution, integrating enterprise data and security capabilities, and participating in the financing of dedicated projects. Competition will shift from “who owns the most GPUs” to “who can consistently deliver usable compute at the lowest total cost while retaining the customer relationship.”

IX. Who Has the Edge: Focus on Contracts, Distribution, and Utilization—not a Single Model Leaderboard

AI labs’ competitive advantage will ultimately rest on three capabilities. First, inference efficiency: delivering equivalent performance with less compute. Second, product mix: enterprise APIs and workflow products are generally more likely to generate sustainable profits than subsidized consumer subscriptions. Third, distribution bargaining power: whether labs can retain customer ownership, reduce partner revenue sharing, and allocate workloads across multiple clouds.

Hyperscalers likewise have three critical capabilities. First, access to power, chips, and networking, which determines deliverable capacity. Second, utilization management, which determines whether heavy capital expenditure can generate adequate returns. Third, enterprise distribution: cloud marketplaces, identity, security, data, and compliance capabilities can reduce friction in model procurement.

Model leaderboards still influence near-term market share, but commercial outcomes depend more on consistent delivery. A leading model may struggle to sustain a revenue advantage if it is too costly, has unstable latency, or lacks enterprise controls. A slightly less capable model with better cost, reliability, and distribution may achieve stronger unit economics. Open models will accelerate capability diffusion, forcing labs to shift value creation toward tools, data, workflows, and customer relationships.

As accounting disclosures mature, the market may reassess both groups. Labs must demonstrate that rapid growth can translate into contribution profit sufficient to cover training and R&D; investment. Cloud providers must show that AI revenue can cover depreciation, financing, and power costs. Both may report substantial AI revenue, but returns on capital are much harder to disguise.

This report provides neither stock-specific price targets nor valuation multiples, so it cannot support precise valuation conclusions. Instead, it offers a profit-pool framework: labs capture model and product profits, while cloud providers capture compute, distribution, and strategic-partnership economics. As dedicated infrastructure matures, these profit pools will be redistributed. Any valuation judgment requires additional company-level data on revenue, capital expenditure, depreciation, contract duration, and cash flow.

X. The Strongest Counterevidence and Key Metrics to Track: Five Sets of Numbers Will Determine Whether the Framework Holds

The first is the inference cost ratio. If AI-lab revenue continues to grow while inference costs as a share of revenue begin rising again, model complexity, price competition, or inefficient usage is consuming the efficiency gains. The report assumes that the inference cost ratio remains stable at 42% from 2026—2028, making this the framework’s most directly falsifiable assumption.

The second is the training cost ratio. The report expects it to decline from 48% in 2026 to 30% in 2028. If spending on next-generation model training continues to outpace revenue growth, labs will struggle to generate operating leverage. R&D; expenses and capitalization policies must also be monitored to ensure that improving gross margins are not masking broader cost pressure.

The third is business mix. A rising share of enterprise and direct API revenue generally supports paid-inference profitability. If subsidized subscriptions account for too much of the mix, heavy users may erode gross margins. Investors should track API revenue, subscription revenue, usage by paying customers, and changes to plan limits in parallel.

The fourth is the alignment between cloud providers’ AI revenue and capital expenditure. Even with rapid revenue growth, returns on capital may decline if depreciation, lease obligations, and power commitments grow faster. Data-center utilization, transfers from construction in progress to fixed assets, depreciation periods, and operating cash flow can test whether a 35%—45% operating margin is genuinely creating value.

The fifth is infrastructure migration. Commissioning schedules, minimum purchase commitments, financing costs, and actual utilization for dedicated AI projects will determine whether 2028 becomes an inflection point. Delays would prolong hyperscalers’ period of elevated market share. On-time commissioning with inadequate utilization would shift risk to labs and financing providers. Only sustained high utilization would place meaningful pressure on hyperscalers’ share and bargaining power.

The framework’s greatest value is that it translates model competition into a verifiable P&L; chain. Customer spending first enters through subscriptions or APIs and is then allocated between labs and cloud providers. Inference efficiency determines contribution profit, training investment determines operating leverage, and capital expenditure and utilization determine cash returns. As disclosures become more transparent, the market will increasingly shift from narratives about model capability toward comparisons of profit and returns on capital.

Related Reading

404K SEMI-AI Evening Brief 2026-08-05 — Broadening Constraints: AI Demand Remains Strong, While Gross Margins and Cash Flow Begin to Differentiate Company Performance

404K SEMI-AI 2026-08-28 AI Model Weekly — Capability Premiums Narrow as Open Source Catches Up and API Prices Fall; Real-World Task Completion Becomes the Key Test

404K SEMI-AI Evening Brief 2026-08-12 — Bottlenecks Spill Over: AI Orders Push Scarcity into Memory and Optical Interconnects

CCL Price Hikes Enter the Profit Validation Phase: Elite Material and TUC Beat Q2 Expectations as Trainium 3 Ramps

SK hynix 2Q26 Earnings Deep Dive: Revenue and Profit Miss Market Expectations, Driving a Further Sharp After-Hours Selloff in Memory Stocks — KRW60.5 Trillion Profit, Early HBM4 Shipments and LTA Cont

• Post-Market Wrap

• AI & Semiconductor Full Supply Chain

• AI Models, Applications & CapEx

• AI Cloud & Data Center Operators

• GPU / CPU / ASIC

• HBM / DRAM / NAND / SSD / HDD

• Semiconductor Equipment, Testing & Advanced Packaging

• Internet / Platforms

• Software / SaaS

• Consumer Electronics & Smart Vehicles

• Jensen's Selected Portfolio

• Memory Index Performance

• US-Listed Memory Turnover Share

AI infrastructure bottlenecks are shifting down the stack from raw GPU supply to security, HBM, packaging/testing (OSAT), and grid constraints, leaving platform and software plays relatively stronger.

404K SEMI-AI | August 31, 2026

Post-Market Wrap

In the last US trading session (August 28), the S&P; 500 fell 0.23%, the Nasdaq 100 slid 0.65%, and the Russell 2000 declined 1.35%. Sector-wise, Communication Services gained 1.42%, while Tech dropped 1.55%. Rather than a broad-based market retreat, capital rotated away from volatile hardware supply chains.

Semiconductors were the primary source of drag: the Semiconductor ETF dropped 3.47%, NVIDIA fell 4.57%, Applied Materials declined 4.29%, and Lam Research dropped 5.24%. By contrast, the Software ETF dipped just 0.74% (still up 15.78% over the past 20 trading days), and Cloud Computing gained 15.48% over the last 20 days, widening sector dispersion.

While industry fundamentals remain intact, structural constraints are shifting downstream. HBM supply-demand dynamics could flip into deficit by 2027, packaging and testing capacity is near full utilization, and neoclouds face heightened multi-tenant security risks alongside financing pressures. Moving forward, key verification points center on order conversion, cost-per-token reductions, and sustained enterprise willingness to pay.

AI & Semiconductor Full Supply Chain

AI Models, Applications & CapEx

• OpenAI
  1) Following the Hugging Face security incident, OpenAI disclosed that it had paused model training to verify sandbox defenses against chained multi-vulnerability exploits.
  2) Risk exposure has now expanded beyond model outputs into research clusters, model weights, and the training pipeline itself. Going forward, the primary catalysts to track are sandbox hardening, weight security, and the pace of training resumption rather than pure model capabilities.

"We paused training. In a world of chained zero-day exploits, we had to figure out how to secure our sandbox."

• Anthropic
  Claude's weekly usage limits were reportedly cut by 17%, whereas ChatGPT indirectly raised its quotas. This divergence may stem from inference serving efficiency, capacity provisioning, or commercial strategy, though available evidence remains insufficient to isolate the exact driver. The most direct metrics to watch are whether quotas, response quality, and unit inference costs can improve simultaneously, alongside enterprise usage trends across complex workflows.

• AI Frontier Labs
  Barclays models AI labs as standalone commercial entities: for every $1 of revenue generated, approximately $0.90 is spent on leased compute. Under its framework, aggregate lab revenues are projected to surge from $7 billion in 2024 to $690 billion by 2028, with training costs falling from 96% to 30% of revenue—making margin expansion heavily dependent on economies of scale.

• Inference Workloads
  As workloads evolve from chatbots to advanced reasoning, single agents, and multi-agent systems, per-task token consumption increases by orders of magnitude. The demand side shows no clear ceiling; expansion is primarily bounded by constraints across GPU availability, power supply, and financing. Consequently, monetizing CapEx into deployable services matters more than headline model upgrades, with facility energization timelines and gross service margins serving as key checkpoints.

"As we evolve from chatbots to reasoning, to agents, and then multi-agent systems, each step scales the tokens consumed per task by an increasing order of magnitude."

• Hugging Face
  Public post-mortems indicate that OpenAI's internal privilege escalation step leveraged a known CVE Linux kernel vulnerability, though external accounts diverge on the specific models involved. The high-conviction takeaway: AI models can now chain known vulnerabilities, requiring platforms to validate isolation against real-world attack chains rather than relying solely on static security benchmarks.

AI Cloud & Data Center Operators

• IREN
  FY4Q26 EBITDA came in at $19 million (vs. Goldman Sachs estimate of $33 million), while AI Cloud revenue reached $71 million (beating the $56 million estimate). The company reported contracted ARR of $4 billion and guided IT capacity delivery targets of 0.3 GW in 2026 and 0.8 GW in 2027, with capital expenditure growth and refinancing requirements escalating concurrently.

• Neoclouds
  ClusterMAX 3.0 benchmarking across 25 cloud providers and 32 clusters revealed systemic vulnerabilities in shared control planes, monitoring privileges, InfiniBand keys, and DPU trust models. Neocloud offerings can no longer compete solely on raw GPU hours—tenant isolation, patch latency, and systems engineering are now the make-or-break criteria for hosting mission-critical workloads.

"The product is no longer just the chip. The product is gigawatt-scale tokens, which encompasses data centers, financing, software, operations, and security controls all in one."

• CoreWeave
  Security evaluations benchmarked CoreWeave alongside Azure for container breakout baselines, underscoring that tier-1 cloud providers must continuously validate component patch levels and isolation configurations rather than treating compliance as a one-time audit. For enterprise customers, continuous auditability across patch deployment cycles, container boundaries, and cross-tenant data access remains paramount.

• Microsoft Data Centers
  Reports indicate that a $19.4 billion Microsoft-backed AI data center project is facing scrutiny over unpermitted gas turbines and a 1.5-million-gallon LNG storage tank. This controversy highlights how expansion constraints are shifting from silicon availability to zoning permits, fuel logistics, and community acceptance. Project timelines must be cross-checked against permitting approvals, grid interconnection plans, and potential delivery delays.

• Data Center Infrastructure & Ancillary Equipment
  Supply-chain checks identify transformers, cabling, liquid cooling, large chillers, and complex networking fabrics as more acute operational bottlenecks than initial power procurement. Notably, BASX reported a 216% YoY surge in data center sales and a 185% increase in backlog. Capital continues to favor full turnkey system integrators over single-component suppliers.

GPU / CPU / ASIC

• NVIDIA
  1) Down 4.57% on August 28, but still up 8.37% over the past 20 trading days.
  2) Industry sources indicate NVIDIA projects hyperscaler CapEx to reach $1.3 trillion in 2027 (up from ~$800 billion in 2026), accompanied by ~70% top-line revenue growth.
  3) Security software stacks, validated reference architectures, and DPU designs are expanding NVIDIA's competitive moat from discrete silicon to full-stack datacenter operations.

• AMD
  Down 2.33% on August 28, bringing YTD gains to 108.34%. Security analyses rank AMD alongside NVIDIA as primary beneficiaries of sustained investments in enterprise security software and system-level capabilities. As procurement criteria shift from single-accelerator compute toward rack-level reliability, broad model support, and multi-tenant isolation, AMD must demonstrate these system capabilities through large-scale hyperscale deployments.

• AI Accelerator Startups
  As vendors like Cerebras, SambaNova, and Etched pivot toward token delivery and managed cloud services, they absorb operational burdens spanning rack stability, model portability, security operations, and debt service. Lacking liquid secondary markets for custom ASICs, collateral residual values remain difficult to underwrite, driving higher cost of capital and equity dilution relative to established GPU platforms.

• GPU Kernel Benchmarking
  Frameworks like KernelBench typically evaluate static shapes, restricted datatypes, and limited randomized trials, probing only a fraction of the total input space. As kernel generation outpaces verification capabilities, evaluations must account for numerical corner cases, cross-GPU portability, and benchmark blind spots to ensure synthetic performance gains translate into reliable production workloads.

"Kernel generation is advancing faster than kernel verification. The next tier of hardware/software evaluation must capture corner numerical cases, cross-GPU portability, and other issues missed by standard benchmarks."

HBM / DRAM / NAND / SSD / HDD

• HBM Supply & Demand
  UBS projects 2027 HBM demand to approach 60 billion Gb against shipments of ~54 billion Gb, tipping the market into a structural deficit with demand up nearly 77% YoY. Long-term supply agreements (LTAs) and hyperscaler expansions remain core drivers. As wafer allocation shifts toward HBM, commodity DRAM supply faces further crowding out, warranting close tracking of pricing, yields, and capacity rollout schedules.

"Then into 2027: demand surges to nearly 60 billion Gb, while shipments reach only around 54 billion Gb."

• Micron
  Down 0.27% on August 28, but up 13.34% over the past 20 trading days. UBS notes that while Micron, Samsung, and SK Hynix are all expanding HBM lines, greenfield fab construction takes years. For Micron, the key investment variable is whether pricing and margin premiums from HBM expansion will outweigh the opportunity cost of constrained conventional DRAM supply.

• Samsung Electronics
  Samsung is ramping up HBM lines, while combined 1H corporate tax payments from Samsung and SK Hynix reached KRW 11.21 trillion (+167% YoY). Because corporate taxes are lagging indicators rather than real-time earnings proxies, key operational signals to track are HBM capacity ramp-up, manufacturing yields, conventional DRAM output dynamics, customer qualifications, and unit production costs.

• SK Hynix
  SK Hynix remains at the epicenter of HBM capacity expansion. Its ADRs dipped 0.35% on August 28 (+12.04% over the past 20 sessions). Trading volume in South Korea's leveraged semiconductor ETFs has subsided to 4% of its June peak, pointing to retail deleveraging; however, the medium-term earnings trajectory hinges on the realization of the projected HBM supply deficit.

• SanDisk
  Up 22.24% over the past 20 trading days. Market focus is shifting from spot NAND price elasticity to long-term contract structures: SanDisk's strategy appears to trade near-term ASP upside for a firmer price floor and predictable cash flows. Write endurance and actual enterprise workload intensity remain critical demand indicators, alongside contract duration and fulfillment rates.

Read more

• I. Net Profit Beat, but Why Did Operating Profit Remain Weak?

• II. How Much of the Gross-Margin Decline Can Be Explained by Upfront Investment?

• III. How Large an Operating Step-Up Is Required in the Third and Fourth Quarters?

• IV. Order Visibility Extends to 2027—but How Far Is It from Revenue?

• V. Expansion Consumes Cash First; Returns May Not Emerge Until 2027

• VI. The Price Target Was Cut from RMB455 to RMB367—What Changed in the Valuation?

• VII. Five Numbers to Watch Next

本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读

Victory Giant Technology’s second-quarter net profit exceeded expectations, but operating profit and gross margin both missed. Whether new-platform ramp-up can restore profitability in the third quarter will determine whether 2027 order visibility and capacity-expansion plans warrant a higher valuation.

I. Net Profit Beat, but Why Did Operating Profit Remain Weak?

The easiest mistake is to interpret net-profit growth as evidence that the core business strengthened in tandem. Victory Giant reported revenue of RMB6.11 billion, up 10.7% QoQ and 29.5% YoY. Growth remained strong, but revenue missed BofA’s forecast and pb consensus by 3.2% and 5.8%, respectively. Demand continued to expand, but the pace of volume growth was slower than previously expected.

Operating profit provides a better measure of core operating quality. Second-quarter operating profit was RMB1.532 billion, down 1.0% QoQ but up 8.3% YoY. This was 7.8% below BofA’s forecast and 16.4% below consensus. With revenue rising nearly 30% YoY but operating profit increasing only 8.3%, the incremental revenue generated materially less profit than a year earlier.

Net profit reached RMB1.568 billion, up 21.7% QoQ and 28.3% YoY, beating BofA’s estimate and consensus by 9.6% and 4.4%, respectively. BofA attributed most of the upside to higher fair-value gains. Such gains can lift reported earnings for the period, but they do not demonstrate improvement in orders, capacity utilization, or product mix. Operating profit and gross margin therefore provide more useful signals of an operating inflection than net profit.


Gross margin provides a second layer of evidence. Second-quarter gross margin was 32.1%, down 2.4 percentage points QoQ and 6.7 percentage points YoY. Revenue growth did not offset cost pressure, indicating that the company absorbed additional labor, depreciation, and ramp-up costs during the platform transition. The income statement already reflects the cost of capacity expansion, while the contribution from that capacity remains to be confirmed in subsequent quarters.


II. How Much of the Gross-Margin Decline Can Be Explained by Upfront Investment?

Management’s explanation follows a clear timeline. The company hired ahead of the new-platform ramp, increasing headcount by more than 30% QoQ in the second quarter. Depreciation on new capacity also began. Costs had therefore entered the income statement before revenue fully reflected the new platform’s contribution, creating a temporary margin trough.

The explanation depends on three conditions. First, the new workforce must reach stable production in the third quarter rather than continue depressing output per employee. Second, utilization of the new equipment must rise quickly so that depreciation is absorbed by a larger revenue base. Third, the share of higher-value PCB products—including high-layer-count and high-density interconnect boards—must increase sufficiently for mix improvement to offset ramp-up losses.

BofA expects the Vera Rubin-related platform to begin ramping in the third quarter. Vera Rubin is NVIDIA’s next-generation AI computing platform and requires more complex high-speed PCBs within its servers. Higher layer counts, faster transmission speeds, and greater manufacturing complexity increase PCB content per server. Victory Giant’s earnings leverage depends on shipment growth and higher content per server occurring simultaneously.

Competition could weaken this transmission mechanism. As more PCB manufacturers enter NVIDIA’s supply chain, customers gain more choice, potentially affecting pricing, share, and certification timelines. BofA therefore cut its 2026 earnings forecast by 7% and reduced its 2027–2028 forecasts by approximately 1%–3%. The larger near-term revision indicates that cost mismatch is the primary short-term explanation; the reductions to longer-term forecasts show that competitive pressure has not disappeared.


Upfront near-term costs and intensifying long-term competition can coexist. A third-quarter gross-margin recovery would demonstrate improved capacity utilization, but would not by itself prove that the company’s market share is secure. If revenue accelerates while gross margin remains near 32%, the earnings quality of incremental orders will need to be reassessed. The platform-transition thesis would receive stronger support only if revenue, gross margin, and operating profit improve together.

III. How Large an Operating Step-Up Is Required in the Third and Fourth Quarters?

BofA’s quarterly model is aggressive, making it particularly useful as a benchmark. It forecasts third-quarter revenue of RMB8.267 billion, up 35% QoQ; operating profit of RMB2.206 billion, up 44%; and gross margin recovering to 32.9%. Fourth-quarter revenue is projected to rise further to RMB11.561 billion, up 40% QoQ; operating profit to RMB3.44 billion, up 56%; and gross margin to 35.1%.

This path requires two consecutive step-ups. The third quarter must demonstrate that the new platform is beginning to contribute revenue, while the fourth must show continued improvement in capacity utilization and product mix. A delay at either stage could push full-year earnings below the model. BofA’s own 2026 net-profit forecast of RMB7.689 billion is already 7% below its previous estimate and 10.7% below consensus. The report retains a positive rating while using a more cautious earnings model than the market—a combination investors should note.

Read more

• Table of Contents

• What the Q2 Turnaround Truly Signals

• How Optical Module Equipment Replaced Photovoltaics as the Primary Growth Engine

• Can the 43.6% Quarterly Gross Margin Be Sustained?

• Multiple Balance Sheet Constraints Remain Even After Cash Flow Turns Positive

• What Does BofA's RMB 630 Price Target Imply for 2027–2028?

• Key Metrics to Watch Ahead

Robotechnik saw a simultaneous surge in revenue and profit in Q2, with optical transceiver equipment officially taking over as its core business. The next critical milestone is converting order backlog into recognized revenue, operating cash flow, and sustainable, high gross margins.

Table of Contents

• What the Q2 Turnaround Truly Signals

• How Optical Module Equipment Replaced Photovoltaics as the Primary Growth Engine

• Can the 43.6% Quarterly Gross Margin Be Sustained?

• Post-Cash-Flow Inflection: Ongoing Balance Sheet Constraints

• What BofA's RMB 630 Price Target Implies for 2027–2028

• Key Metrics to Monitor Next

What the Q2 Turnaround Truly Signals

Robotechnik’s 1H 2026 interim report presents headline numbers that appear unremarkable at first glance, but are striking upon closer inspection. First-half revenue reached RMB 608 million, up 145.03% YoY, while net profit attributable to parent company shareholders fell 80.51% YoY to RMB 6.56 million. The YoY profit decline was largely due to a Q1 net loss of RMB 38.82 million. Backing out Q1 from 1H results reveals standalone Q2 revenue of approximately RMB 445 million, net profit of approximately RMB 45.37 million, and net operating cash flow of approximately RMB 241 million.

In other words, Q2 contributed roughly 73% of 1H revenue and fully erased Q1 losses. Quarterly revenue grew approximately 193% YoY and 172% QoQ. Gross profit reached approximately RMB 194 million, translating into an implied gross margin of roughly 43.6%—a 7.3 percentage point expansion over Q1. The simultaneous convergence of revenue scaling, gross margin expansion, and positive operating cash flow demonstrates a much higher quality of earnings recovery than one driven merely by cost-cutting or non-recurring items.

The quarterly financial progression compiled in BofA's report underscores the magnitude of this turnaround. From Q4 2024 through Q1 2026, Robotechnik saw quarterly revenue linger below RMB 200 million for multiple consecutive quarters, logging net losses of approximately RMB 26 million in Q1 2025 and RMB 39 million in Q1 2026. In Q2 2026, revenue rebounded above RMB 400 million, gross margin climbed to a 10-quarter high, and net profit reached approximately RMB 45 million.

However, the Q2 jump should not be conflated with guaranteed, steady full-year hypergrowth. Equipment vendors are inherently subject to revenue recognition timing driven by customer acceptance test schedules; shifting a few major projects across quarter boundaries can dramatically swing quarterly top and bottom lines. This volatility was already evident when revenue reached RMB 533 million in Q4 2025 before pulling back to RMB 164 million in Q1 2026. As such, Q2 primarily validates the revenue- and earnings-generating capacity of the new business lines; sustained performance will hinge on subsequent delivery and acceptance cadences.

Structurally, Q2's P&L; improvement stemmed from a decisive shift in business mix rather than a sudden recovery in the legacy photovoltaic (PV) business. In 1H 2026, revenue from optical module and semiconductor equipment reached RMB 488 million, surging 952.17% YoY and accounting for roughly 80% of total revenue, whereas PV equipment revenue fell 53.93% YoY to RMB 83 million. Generating more optical module equipment revenue in just six months than the RMB 485 million recorded in all of 2025 confirms that the company has completed a substantive transition of its strategic focus.

This pivot fundamentally enhances operating leverage and profit elasticity. PV equipment caters to mature production lines characterized by cyclical capex downcycles and fierce price competition. In contrast, optical module equipment spans high-precision active optical alignment, die bonding, testing, and automation integration—demanding stringent precision, high yields, and custom engineering capabilities. As high-value equipment expands as a share of total revenue, each incremental unit of revenue generates substantially higher gross profit than legacy PV orders ever did.

How Optical Module Equipment Replaced Photovoltaics as the Primary Growth Engine

Robotechnik’s new growth curve is primarily powered by the consolidation of ficonTEC, a specialist in micro-assembly and testing equipment for optoelectronic devices catering to silicon photonics, optical transceivers, and co-packaged optics (CPO). As AI data centers drive surging demand for high-speed optical interconnects, optical transceiver manufacturers have accelerated capacity expansions, transitioning tool procurement from R&D; prototyping and pilot validation to high-volume commercial deployment. This has unlocked a significantly faster-growing demand window for Robotechnik than legacy PV automation could provide.

The company’s position in the supply chain dictates its revenue recognition cadence. Equipment vendors secure purchase orders, proceed through engineering design, procurement, assembly, commissioning, and on-site customer acceptance before booking revenue. Order intake leads reported financial growth, with production throughput and customer acceptance efficiency determining the lag between the two. Robotechnik’s ultimate valuation depends not only on broader optical transceiver industry tailwinds, but also on its ability to standardize bespoke engineering into modular platforms that can be replicated across diverse customer fabs and manufacturing sites.

At year-end 2024, the company’s total order backlog stood at approximately RMB 570 million. By year-end 2025, backlog had surged to RMB 2.187 billion, including RMB 1.105 billion in optical module and semiconductor equipment. By August 25, 2026, total backlog expanded further to RMB 3.386 billion, with optical module and semiconductor equipment reaching RMB 2.452 billion, or approximately 72% of the total. In less than eight months, optical module and semiconductor equipment backlog surged by RMB 1.347 billion, representing a 122% increase.

The expanding backlog provides strong visibility into future revenue. Based on 1H 2026 optical module and semiconductor equipment revenue of RMB 488 million, the RMB 2.452 billion segment backlog represents roughly 2.5 semiannual revenue periods. While this ratio cannot simply be extrapolated into the next twelve months of revenue due to varying delivery cycles and potential customer acceptance delays, it confirms that the primary operational bottleneck has shifted from "order acquisition" to "execution and on-time delivery."

Read more

• Table of Contents

• 1. Q2 Revenue of RMB 90.6B Confirms Resilient Demand

• 2. Consumer Electronics: Anchoring Cash Flow and Baseline Profitability

• 3. Communications Segment Accelerates, but Margins Lag

• 4. How to Interpret the 12-Million-Unit Capacity Target

• 5. Margin Inflection Driven by Multiple Structural Shifts

• 6. Key Assumptions Underpinning the RMB 86 Price Target

• 7. Cash Flow and Debt Profile Dictate Expansion Quality

• 8. Seven Key Signals to Monitor Over the Next Four Quarters

本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读

Bernstein’s RMB 86 price target rests on a clear premise: consumer electronics defends baseline profitability through 2026, while high-margin interconnect products take over as the primary growth engine in 2027. The 12-million-unit capacity represents only an upper ceiling for supply; ultimate execution depends on delivering 2.5 million shipments, recovering communications gross margins, and generating robust cash flows.

Table of Contents

• 1. Q2 Revenue of RMB 90.6B Confirms Resilient Demand

• 2. Consumer Electronics: Anchoring Cash Flow and Baseline Profitability

• 3. Communications Segment Accelerates, but Margins Lag

• 4. How to Interpret the 12-Million-Unit Capacity Target

• 5. Margin Inflection Driven by Multiple Structural Shifts

• 6. The Core Assumptions Behind Bernstein's RMB 86 Price Target

• 7. Cash Flow and Leverage Dictate the Quality of Expansion

• 8. Seven Key Signals to Watch Over the Next Four Quarters

1. Q2 Revenue of RMB 90.6B Confirms Resilient Demand

Luxshare Precision reported Q2 2026 revenue of RMB 90.616 billion (+8% QoQ, +44% YoY), beating consensus expectations by 5.7%. Net profit attributable to the parent company reached RMB 4.183 billion (+14.3% QoQ), missing institutional consensus by 3.5%. These figures reflect two distinct realities: robust order intake and shipment execution that have yet to fully translate into bottom-line profit expansion.

The 1H 2026 interim report provides a complete financial picture: revenue reached RMB 174.504 billion (+40.16% YoY), attributable net profit rose 18.04% to RMB 7.843 billion, while recurring net profit grew just 6.47% to RMB 5.962 billion. The company disclosed FX losses of RMB 1.986 billion from foreign currency balance sheet exposures, partially offset by RMB 1.297 billion in hedging gains classified under non-recurring items. This accounting presentation masked the net hedging benefit in recurring earnings, underscoring foreign exchange volatility as a primary variable governing earnings quality.

Figure 1: Divergence between top-line and bottom-line growth rates. Source: Luxshare Precision 1H 2026 Interim Report. Data period: 1H 2026.

Gross margins showed no broad-based deterioration. Blended gross margin for 1H 2026 expanded by 17 bps YoY to 11.78%. In Q2, gross margin softened by 30 bps QoQ to 11.6%, matching consensus expectations. The near-term margin pressure stems primarily from product mix, ramp-up costs in nascent divisions, and FX headwinds—rather than structural degradation in core manufacturing capabilities.

2. Consumer Electronics: Anchoring Cash Flow and Baseline Profitability

Consumer electronics generated RMB 122.476 billion in 1H 2026 (+19.27% YoY), accounting for 70.18% of total revenue, with segment gross margin easing 13 bps YoY to 10.25%. Amid muted global smartphone and PC demand, growth was driven by market share gains, new ODM/JDM programs, non-Apple customer expansion, and consolidation effects. While this business sustains plant utilization, supply chain bargaining power, and R&D; funding, it is unlikely to drive significant valuation multiple expansion on its own.

Bernstein forecasts Apple-related revenue to grow at a ~3% CAGR through 2028, compared to >20% for non-Apple consumer electronics. Additional profit upside stems from iPhone assembly at Luxsan, where net margins are projected to expand from 1.3% toward 1.7%, contributing nearly RMB 5.0 billion in adjusted operating profit to Luxshare by 2028. This margin trajectory does not rely on a broad end-market rebound, but rather on yield stability, program scale, and operational cost discipline.

Figure 2: Apple revenue contribution declines as communications and automotive expand. Source: Bernstein. Data period: 2021A-2028E.

As illustrated, Apple's revenue contribution is projected to decline from 74% in 2021 to 39% by 2028, while communications and automotive are expected to rise to 15% and 19%, respectively. This diversification moderates single-customer concentration and product-cycle risk, though it introduces execution challenges around automotive integration, data center certifications, and cross-border operations. The ultimate success of this diversification must be judged by segment margins and free cash flow generation, not merely revenue mix.

3. Communications Segment Accelerates, but Margins Lag

In 1H 2026, communications and data center revenue surged 49.66% YoY to RMB 16.609 billion, increasing its revenue share to 9.52%. However, segment gross margin contracted 248 bps YoY to 17.29%. This divergence between top-line expansion and gross margin indicates that current growth remains diluted by server assembly, lower-margin legacy products, and initial ramp-up costs on new manufacturing lines.

Bernstein outlines a clear re-rating roadmap for the communications segment. Deliveries of copper Cable Cartridge assemblies are slated to commence in Q4 2026, followed by volume shipments of 1.6T optical transceivers in early 2027. By mid-2027, the company aims to operate 10 automated production lines with an aggregate annual capacity of 12 million units. While server assembly provides customer access and operational scale, high-value components—copper interconnects, optical modules, thermal solutions, and power supplies—drive content per box and margin expansion.

Read more

• Table of Contents

• The 2Q Report Card: Growth Intact, with the Expectation Gap in Revenue and Gross Margin

• Why Profit Looks Better Than Gross Margin: Separating Non-Operating Income from Operating Quality

• How Next-Gen GPU and ASIC Raise PCB Dollar Content

• Inventory and capacity have been front-loaded—what gates remain before revenue is realized

• Cash flow and capacity expansion: the other side of growth preparation is capital lock-up

• What the RMB 550 target price depends on, and which numbers to watch over the next two quarters

本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读

Second-quarter revenue and gross margin came in below expectations. Next-generation GPU and ASIC orders, raw-material and work-in-process inventory build, and the ramp of new domestic capacity still provide verifiable support for a second-half acceleration.

Table of Contents

• The 2Q Report Card: Growth Intact, with the Expectation Gap in Revenue and Gross Margin

• Why Profit Looks Better Than Gross Margin: Separating Non-Operating Income from Operating Quality

• How Next-Gen GPU and ASIC Raise PCB Dollar Content

• Inventory and Capacity Have Been Front-Loaded; What Gates Remain Before Revenue Is Recognized

• Cash Flow and Expansion: The Other Side of Growth Preparation Is Capital Lock-Up

• What the RMB 550 Target Price Depends On, and Which Numbers to Watch Over the Next Two Quarters

The 2Q Report Card: Growth Intact, with the Expectation Gap in Revenue and Gross Margin

Victory Giant Technology continued to grow in the second quarter; the market's disappointment stemmed from growth pace and unit profitability coming in below what had been priced in. Goldman Sachs' August 28 report shows 2Q26 revenue of RMB 6.11 billion, up 29% year over year and 11% quarter over quarter. That figure was 16% below Goldman's estimate and 6% below consensus. Gross margin was 32.1%, down 2.4 percentage points from 1Q26 and down 6.7 percentage points from the prior-year period, also below Goldman's 35.6% forecast and the 36.8% consensus.

Net profit held up better. 2Q26 net profit was RMB 1.568 billion, up 28% year over year and 22% quarter over quarter, 4% above consensus but still 11% below Goldman's estimate. Weak revenue, weak gross margin, and near-consensus net profit together indicate that the expense ratio and non-operating income cushioned pressure at the production level. They can support current-period profit, but they cannot substitute for order conversion, yield improvement, and utilization ramp.

This report card updates the prior thesis to "revenue direction confirmed, gross-margin recovery slower than expected." Earlier coverage of Victory Giant Technology framed 2Q26 as a shallow-ramp period for the new platform, with a re-acceleration expected in 3Q26 and key watch items being revenue slope, gross margin, new customers, inventory, and cash flow. Revenue did grow 11% quarter over quarter, indicating demand has not stalled; the continued decline in gross margin to 32.1% indicates that headcount, depreciation, and process ramp have already entered the cost base before new capacity has entered the revenue line.

Goldman attributes the revenue shortfall mainly to the manufacturing difficulty of AI-server PCBs. Higher layer counts, higher-order HDI, and stricter signal-integrity and material requirements mean new equipment takes longer to reach stable yields. The company's explanation on the August 27 earnings call was more granular: the second-quarter gross-margin decline was driven mainly by headcount and depreciation, with raw materials having a smaller impact. First-half headcount rose about 30% year over year, with training and staffing running ahead of orders; after new plants were capitalized, depreciation also entered costs first. The two explanations differ in emphasis. Goldman stresses materials and ramp; the company stresses headcount and depreciation. The operating fact they both point to is that capacity costs have already been incurred while revenue is still ramping.

The real test of the second quarter therefore falls on 3Q and 4Q. If headcount, depreciation, and inventory build represent front-loaded volume preparation, subsequent revenue growth will dilute fixed costs and gross margin should recover as utilization rises. If order delivery is delayed further, the front-loaded costs will remain on the income statement, and inventory and depreciation pressure will increase. The second quarter alone cannot resolve this judgment; revenue, gross margin, and inventory turnover must all improve simultaneously over the next two quarters.

Why Profit Looks Better Than Gross Margin: Separating Non-Operating Income from Operating Quality

The interim report shows Victory Giant Technology's core business still growing, while the income statement also contains a sizable non-recurring cushion. The company's unaudited 2026 interim report discloses H1 revenue of RMB 11.629 billion, up 28.77% year over year; net profit attributable to shareholders of RMB 2.857 billion, up 33.30%; and non-GAAP net profit attributable to shareholders of RMB 2.418 billion, up 12.50%. The 20.8-percentage-point gap between attributable and non-GAAP growth is already enough to affect judgments about earnings quality.

Cost-side pressure is clearer. H1 operating costs rose 34.83% year over year, faster than revenue, implying a consolidated gross margin of 33.22%, down 3.00 percentage points from 36.22% in the prior-year period. Main-business gross margin fell from 33.05% to 28.58%, a decline of 4.47 percentage points; parent-company gross margin also fell by about 3.23 percentage points. The simultaneous decline across all three measures indicates broad-based pressure on the core manufacturing business that consolidation-scope effects cannot explain away.

Reported net profit was boosted by financial-asset gains and dragged by FX losses. H1 fair-value gains were RMB 595 million, mainly from trading financial assets; total non-recurring items were RMB 439 million, about 15.4% of attributable net profit. Finance costs in the same period were RMB 297 million, up nearly 7x year over year, including FX losses of about RMB 222 million. Victory Giant Technology has high export revenue while material purchases are settled more in RMB, which tends to create large foreign-currency receivables and cash exposures. The company uses forward FX contracts and foreign-currency borrowings for partial hedging but does not apply hedge accounting, so quarterly profit can still swing.

R&D; spending also needs to be viewed within the expansion cadence. H1 R&D; expense was RMB 359 million, up 1.81% year over year, with the R&D; expense ratio falling from about 3.91% to 3.09%. The company is simultaneously advancing processes such as 100+ layer high-multilayer boards, 30-layer 10-order and 36-layer 14-order HDI, vertical power-delivery embedded capacitance, and COWOP. The decline in the R&D; ratio may reflect revenue growing faster than expense input, or it may reflect resources shifting more toward equipment, manufacturing engineering, and volume ramp. This should be assessed in conjunction with high-end product revenue, yields, and R&D; headcount changes; the expense ratio alone is insufficient to infer weakening technology investment.

Evaluating profit quality requires looking at three sets of numbers simultaneously. The first set is gross margin and non-GAAP net margin, which represent core operating efficiency; the second is fair-value gains, FX, and finance costs, which determine quarterly net-profit volatility; the third is operating cash flow and working capital, which test whether profit converts into cash. The resilience of 2Q26 net profit deserves recognition, but the current financials do not yet prove that unit profitability has bottomed.

Read more

• Overall Assessment

• Model Releases, Upgrades, and Capability Changes

• Reasoning, Training, Multimodality, and Agent Capabilities

• Open- vs. Closed-Source Competition, Benchmarks, and the Developer Ecosystem

• API Pricing, Commercialization, and Vendor Divergence

• Diverging Views, Disconfirming Evidence, and What to Watch Next Week

Model competition shifted this week from “who tops the leaderboard” to “who can complete the task at the lowest cost.” Narrowing capability gaps and lower API prices should support broader demand, but agent reliability, differences in revenue recognition, and customer concentration remain the principal risks.

Overall Assessment

• Frontier capabilities continue to improve, but leading an individual benchmark has less investment significance. Grok 4.6 scored 94.9% on GPQA Diamond, just 0.4 percentage points above Gemini 3.7 Flash, which has a lower cost per task. A narrow lead will make it difficult for model vendors to sustain a premium based solely on being “the smartest.”

• Open-source models are competing simultaneously on performance, price, and deployment flexibility. QWEN38270 ranked No. 1 among open-source models and No. 7 overall in the Image-to-WebDev Arena. Open-source models’ share of tokens on Vercel rose from 28% to 62% in 2 months. Although this reflects only one developer platform, it is sufficient evidence that routers are choosing “good enough and inexpensive” models more aggressively.

• Agents are moving into longer-duration tasks, but reliability has not crossed the threshold at the same pace. Claude unified memory across Chat and Cowork, Claude Code added isolated subagents, and OpenAI is testing a Codex mode that can continue working across sessions. Meanwhile, the best run on real-world e-commerce tasks completed only 66 of 107 tasks, while multi-agent discussions produced correct answers in just 17%–36% of runs in hidden-information experiments.

• The price war is now measurable at the task level. OpenAI cut both input and output prices for GPT-5.6 Sol. In one coding test, GPT-5.6 Luna completed more fixes in less time and at lower cost. The sample is small, but buyers will increasingly route workloads based on completed outcomes.

• Monetization is expanding beyond subscriptions and APIs into advertising, tool gateways, and enterprise workflows. OpenAI is testing sponsored units below responses for some Free and Go subscribers in India. Meta’s spending on Anthropic, while reportedly declining, is still said to amount to hundreds of millions of dollars per month. Model revenue is growing, but customer concentration, traffic monetization, and privacy boundaries will also enter valuation frameworks more quickly.

This week’s most important development can be summarized as a transmission chain: narrowing capability gaps make it easier for developers to switch models; API prices then decline, potentially expanding usage; and as agents take on longer tasks, enterprises increasingly require proof that the model actually made, saved, and submitted the requested changes. Value is no longer determined solely by parameter count and static benchmark scores, but jointly by cost per effective unit of work, success rate, and auditability.

The valuation implication is straightforward. Capability leadership still matters, but if the lead is insufficient to offset differences in price, latency, and integration costs, customers will direct more traffic to lower-cost models. To protect revenue, model vendors must continually enable new tasks that older models cannot perform, while achieving success rates high enough to deliver genuine labor savings. Using price cuts alone to drive volume will pressure margins first.

There was also no evidence this week of a winner-takes-all market. Different models stood out in scientific reasoning, image editing, speech synthesis, code repair, and robotic learning. A portfolio-based routing strategy therefore fits enterprise procurement logic better: flagship models for difficult tasks, lower-cost models for high-volume workloads, open-weight or privately deployed models for sensitive data, and a unified evaluation and permissions layer for governance. Model vendors are competing not only for inference volume, but also to become the default routing gateway.

Model Releases, Upgrades, and Capability Changes

Release activity remained high this week, but gains were distributed across scientific reasoning, imaging, speech, and embodied intelligence. No single model established an overwhelming lead across all dimensions; vendors instead appear to be pursuing leadership in the tasks they can monetize most readily.

Capability upgrades fall into 3 categories. The first pushes established task formats to higher performance, as with Grok 4.6 in scientific reasoning. The second raises generation quality to professional-work standards, as with Microsoft’s image editing. The third changes how tasks are specified and executed, as Skild S1 does by learning robotic actions from video. The first 2 categories can enter APIs and subscriptions more quickly. If the third proves replicable, it could generate greater value per customer, but it also entails longer deployment cycles, hardware variability, and greater safety liability.

Grok 4.6’s 94.9% shows that the ceiling for scientific question answering continues to rise, but Gemini 3.7 Flash trails by only 0.4 percentage points and has a lower cost per task. At the other extreme, the new high on ARC-AGI-3 is still just 4.58%. Taken together, the 2 benchmarks support a clear conclusion: models are progressing rapidly on familiar knowledge and reasoning formats but remain weak at tasks requiring them to discover new rules. Investors should not extrapolate a near-perfect score on one leaderboard into evidence that general intelligence has matured.

The divergence between these leaderboards also underscores the need to understand what each test measures. GPQA emphasizes difficult scientific question answering, whereas ARC-AGI-3 is closer to discovering unfamiliar rules from limited examples. They measure different capabilities. For products that only need to retrieve, analyze, and explain established knowledge, near-perfect models can already create value. For tasks requiring new-process discovery, operation in anomalous environments, or sustained self-correction, low-scoring benchmarks reveal the risks more clearly.

Microsoft’s MAI-Image-2.6-Preview is best understood through the lens of product control. It ranks No. 1 in image editing and No. 2 in text-to-image generation, including first place in 5 of 19 text-to-image categories. Microsoft’s continued iteration of proprietary image models means its enterprise AI platform need not cede every high-value multimodal workload to external model vendors. The decisive test will come after the private preview becomes broadly available: whether quality, latency, and pricing can all be sustained.

The monetization paths for multimodal models are also more concrete than for text chat. Image editing maps directly to marketing assets, product images, and design workflows; speech models to customer service, real-time companionship, and voice agents; and robotics models to physical tasks. The closer a use case is to an existing budget line, the easier it is for customers to quantify time savings. The closer it is to physical equipment, however, the higher the cost of a single failure—and the less adequate average benchmark scores become for acceptance testing.

Breeze TTS 2 illustrates another side of open weights. With an Elo score of 1215 on the Provider Voices leaderboard, it ranks No. 6 among all models and leads the next open-weight model by 90 Elo. Open weights provide the option of private deployment, but do not automatically make hosted inference less expensive.

It processes 45 characters per second, below Fish Audio S2 Pro’s 102 characters per second. Its hosted price is $34 per 1 million characters, versus $15 for Fish Audio S2 Pro. Buyers choosing self-hosting must also include compute, operations, and peak-capacity requirements in total cost.

Read more

• Overall Weekly View

• Cloud Revenue and AI Capital Expenditure

• Data Centers, Compute Procurement, and Custom Silicon

• Models, Software, and Application Monetization

• Diverging Models: M7 vs. Neocloud Providers

• Divergent Views, Disconfirming Evidence, and Next Week’s Watchlist

AI compute demand continues to spread from hyperscalers to neoclouds, with Amazon AWS placing a major incremental GPU order as Google improves TPU energy efficiency. The investment stance remains bullish, but cash conversion, guarantee exposure, and power delivery have become more important validation points.

Overall Weekly View

The most important development this week is that demand for AI infrastructure has not contracted as in-house chip development accelerates. Instead, it is expanding along two tracks. Amazon AWS plans to deploy approximately 2 million additional NVIDIA GPUs in 2027–2028, while Google continues to specialize TPUs for training and inference and scale up its clusters. Cloud providers want both the time-to-market of general-purpose GPUs and the unit economics of proprietary silicon, shifting procurement from an either-or decision to a dual-track strategy.

The two procurement tracks serve different business objectives. Proprietary chips give cloud providers control over silicon costs, software orchestration, and power efficiency, while externally sourced GPUs provide a mature development environment, early access to advanced products, and broader customer reach. As long as model iteration remains rapid, cloud providers are unlikely to commit all demand to a single architecture. For investors, actual share shifts should be measured by how much power each chip category consumes and how many paid workloads it runs—not procurement announcements alone.

Supplier-side demand indicators remain very strong. NVIDIA reported FY2027 Q2 revenue of $96.2 billion, including $89.0 billion from Data Center. Hyperscale customers contributed $48.7 billion, while ACIE—AI cloud, industrial, enterprise, and sovereign customers—contributed $40.3 billion. ACIE revenue grew 138% year over year and 25% sequentially, reaching 83% of the hyperscale business. This shows that incremental demand no longer depends solely on a handful of large cloud providers, but it also warrants caution: customer classifications are becoming less distinct, so supplier segmentation should not be treated as a direct proxy for individual cloud providers’ revenue.

Customer reclassification deserves particular scrutiny. One AI cloud customer was moved into the hyperscale category this quarter following a change in its business model, illustrating how the boundary between neoclouds and traditional CSPs shifts with customer scale and business mix. A lower hyperscale share does not necessarily indicate a genuine decline in concentration, while a higher ACIE share does not necessarily come entirely from new customers. Investors should track the largest end customers, tenant concentration at cloud providers, and accounts-receivable quality together.

AI capital expenditure should still be viewed as an expansion cycle rather than one that has already peaked. However, the counterevidence is now more concrete: NVIDIA’s operating cash flow fell sequentially from $50.3 billion to $24.1 billion, accounts receivable rose to $63.1 billion, and days sales outstanding lengthened from 45 days to 60 days. Neocloud providers, meanwhile, increasingly rely on customer prepayments, GPU financing, and asset refinancing. Looking only at orders overstates quality, while focusing only on debt understates the acceleration in underlying demand. Procurement, delivery, deployment, utilization, revenue, and cash collection now need to be verified sequentially.

Cloud Revenue and AI Capital Expenditure

There were no new like-for-like quarterly disclosures for M7 cloud revenue this week, so NVIDIA sales cannot be used as a proxy for revenue at Amazon AWS, Microsoft Azure, or Google Cloud. What can be confirmed is that cloud providers’ procurement appetite and supply readiness continue to strengthen. NVIDIA Data Center revenue grew 117% year over year, while FY2027 Q3 total revenue guidance is $108.0 billion, plus or minus 2%. These figures support continued growth in cloud capacity available for sale, but the ultimate conversion into cloud revenue will depend on customer deployment speed, hourly pricing, and inference utilization.

Cloud revenue typically appears at least two operating stages after chip orders. Equipment first enters construction in progress or fixed assets. Only after power-up, networking, and software acceptance does it become a saleable instance, after which customers must migrate workloads and raise utilization. Capital expenditure and depreciation appear first during construction; revenue and gross-margin improvement follow later. Mapping a supplier’s current-quarter revenue directly onto current-quarter cloud growth will therefore distort both timing and margins.

Amazon AWS has taken the clearest action. Its approximately 2 million additional GPUs span Blackwell Ultra, Rubin, and Rubin Ultra and are separate from the existing plan to deploy more than 1 million GPUs beginning in 2026. Amazon AWS is also continuing to invest in Trainium. This combination suggests that proprietary silicon primarily addresses controllable costs, energy efficiency, and specific workloads, while NVIDIA GPUs provide model compatibility, early access to advanced products, and customer choice. The two are likely to coexist for a considerable period.

Capital-expenditure pressure has not disappeared; it is simply being distributed across more parties. NVIDIA inventory rose from $25.8 billion to $31.6 billion, while purchase and supply commitments increased from $119.0 billion to $279.0 billion, primarily to secure HBM and system supply. Suppliers’ advance capacity purchases can shorten cloud providers’ lead times for complete racks, but they also expose upstream vendors to inventory, gross-margin, and customer-performance risk. NVIDIA’s next-quarter gross-margin guidance of 74.0%, down 1 percentage point from this quarter, is one indication that costs are being recognized before the related revenue is realized.

Cash flow is a better indicator of cycle quality than procurement value. NVIDIA’s days sales outstanding lengthened to 60 days this quarter, while the company disclosed $108.5 billion of guarantee exposure. Operating cash flow remained positive but declined sharply sequentially. If cloud providers and model developers secure longer payment terms, suppliers may recognize revenue well before collecting cash. Over the next quarter, investors should monitor NVIDIA’s accounts receivable, cloud providers’ free cash flow, and neocloud financing rates together. Only a concurrent deterioration across all three would constitute firm evidence that capital-expenditure intensity needs to be revised downward.

This expansion cycle differs in another respect: the financing entity is no longer limited to the end cloud provider. Chip suppliers pre-purchase HBM and system capacity, model developers sign multiyear leases, neocloud providers obtain customer prepayments and equipment financing, and completed infrastructure assets are subsequently refinanced on a long-term basis. Each layer can accelerate construction, but it can also cause the same end demand to appear in the commitments and contracts of multiple entities. These figures therefore cannot simply be added to estimate total demand, nor can risk be assessed solely from any one company’s reported debt.

Returns on capital should ultimately be tested against unit economics. Cloud providers must demonstrate higher revenue per unit of power and rack capacity; chip suppliers must show that securing supply does not continuously depress gross margins; and neocloud providers must prove that cash generated by long-term contracts can cover interest and maintenance capital expenditure. If any layer depends on another financing round to fill the cash shortfall from existing projects, the quality of growth will deteriorate.

Data Centers, Compute Procurement, and Custom Silicon

The unit of competition in data centers has shifted from individual chips to complete systems delivered on schedule. Customers are buying usable training throughput, inference capacity, and reliability within a fixed power budget. A delay in any component—GPU, HBM, networking, power, cooling, or rack delivery—prevents contracted compute capacity from generating revenue. Amazon AWS is therefore purchasing at significant scale, but its actual contribution in 2027–2028 must still be recognized progressively as capacity is installed, energized, and utilized by customers.

In this chain, chips are typically secured first, but their revenue impact materializes last. HBM and networking components determine whether complete racks can be delivered; substations and grid-connection quotas determine whether equipment can operate; and software orchestration determines whether expensive compute capacity can sustain high utilization. Any delay leaves orders in backlog. Rising supply commitments indicate management’s willingness to bet on future demand, but do not by themselves prove that data centers are already generating sustainable cash flow.

Google’s TPU 8t addresses the costliest constraint: power. The disclosed 8t Superpod contains 9600 chips, accesses 2PB of shared memory, and delivers 121 exaflops of FP4 compute. Its real-world performance per watt is 2 times that of Ironwood. In commercial terms, the same gigawatt-scale data center can process more training tokens, allowing Google either to perform the same work with less power and fewer racks or to sell more compute within the same budget.

The next-generation TPU v10 supply chain is evolving toward multiple suppliers and more complex packaging. MediaTek continues to handle the I/O chiplet and system integration, while other design-service providers may participate in selected modules. For Google, a multi-supplier structure reduces single-point risk and makes it easier to scale custom-silicon production in line with data-center deployment. For external suppliers, share of an individual chip may fragment, but the value attached to I/O, HBM, packaging, and interconnects should rise with system complexity.

Custom silicon has not displaced Nvidia from cloud-provider racks. Amazon AWS plans to add NVLink Fusion and NVHBM support to future Trainium generations while also procuring Nvidia CPUs, networking, models, and software. Nvidia also plans to deliver Vera Rubin to Google Cloud, Microsoft Azure, Oracle, and Nebius. Competition is intensifying at the chip level even as cooperation expands across racks, interconnects, and software. Nvidia must defend its GPU share, but whether the rest of the system continues to be built around its interfaces will matter even more for the value of its platform.

Apple is pursuing a more closed architecture. One engineering assessment indicates that M5 is entering Private Cloud Compute servers in serviceable modular form, with unified memory and lower power consumption used primarily for Apple’s internal inference workloads. The design still appears closer to engineering validation or early production and does not support conclusions about large-scale training capacity or external cloud revenue. The investment implication is narrower: M7 companies that control devices, chips, and service distribution can retain part of their inference workloads within their own ecosystems, reducing incremental purchases from external clouds.

The 3 strategies represent 3 distinct moats. Amazon AWS emphasizes customer choice between general-purpose GPUs and Trainium; Google emphasizes coordination across models, chips, and data centers; and Apple emphasizes devices, unified memory, and private-cloud security. All 3 reduce dependence on any single external chip supplier, but differ in openness. General-purpose GPUs retain an advantage when cloud customers need portability across models and mature software. Custom silicon becomes more economical as workloads grow larger and more stable, allowing development costs to be amortized more effectively.

Supplier share should also be assessed by layer. Even if custom compute chips gain share, HBM, interconnects, networking, packaging, and system software may still come from external platforms. Nvidia’s expanding collaboration with Amazon AWS across these layers shows that lower GPU share does not necessarily translate into a proportional decline in platform revenue. Platform value would come under genuine pressure only if cloud providers progressively replace its interconnect and software interfaces.

Power remains the final gate to system deployment. Anthropic has reportedly signed a 6-year, US$45 billion compute agreement with Nscale covering 460 MW of power capacity, expected to use Vera Rubin and begin coming online by the end of next year. Such long-term contracts support data-center financing and lock in demand for longer, but construction delays, changes in model economics, and customer credit risk accumulate over a multiyear term. The contract value is substantial, but revenue can only be recognized progressively as capacity comes online.

Models, Software, and Application Monetization

This week’s developments show further divergence in monetization models, but no new revenue figures that can be compared directly across the M7. Amazon AWS is bundling GPUs, Vera CPUs, Spectrum networking, Nemotron models, and CUDA-X libraries, while extending the stack into Amazon’s robotics operations. It sells both compute time and an end-to-end deployment stack spanning models and physical workloads. When customers use networking, storage, models, and operations services on the same cloud, each AI workload generates more cloud revenue than a bare-GPU rental.

Meta is placing greater emphasis on the consumer-relationship layer. One member of its superintelligence lab has argued that model recipes may converge, leaving data collection, reward systems, compute, and consumer distribution as the main sources of differentiation. This should be treated only as a strategic signal, not as formal company guidance. It suggests that Meta is more likely to monetize its investment first through advertising, recommendations, and consumer products than through standalone cloud-computing sales. The key question is whether AI features improve advertising conversion, engagement time, and revenue per unit of compute.

Google and Apple are more focused on using custom silicon to protect the economics of their own applications. Google optimizes training and inference chips separately to reduce the die area and power consumed by unnecessary circuitry, while Apple is extending the M-series unified-memory architecture into private-cloud inference. The immediate benefit for both companies is lower internal service costs. External revenue must emerge through Google Cloud’s marketable TPU capacity, Search and model products, or Apple’s device and services experience. Until those revenue metrics appear, silicon advances should be treated primarily as cost and supply advantages.

Model developers are also reshaping cloud infrastructure. Anthropic’s long-term compute contract directly secures power and future chips, while Amazon AWS is bringing models, software libraries, and robotics workloads into the same partnership. As models require more persistent inference, cloud providers gain an opportunity to shift from one-off training demand toward recurring compute revenue. However, falling inference prices may offset token growth. Higher usage alone is insufficient validation; simultaneous improvement in revenue per unit of power, GPU utilization, and cloud margins would be a more meaningful signal.

Inference economics is the link that determines whether application monetization can fund capital expenditure. Faster chips reduce the cost per query but may also drive prices lower. If demand is sufficiently elastic, token growth will outpace price declines, lifting cloud revenue and utilization together. If customers merely use cheaper inference to replace existing workloads, total revenue may not improve. Next week and in subsequent earnings reports, investors should prioritize evidence of “higher volume, lower pricing, stable margins” over isolated performance records.

Diverging Models: M7 vs. Neocloud Providers

Neocloud providers compete on price, while M7 cloud providers command a premium through broader services, vendor choice, and balance-sheet strength. This week’s market data show CoreWeave and Nebius pricing comparable GPUs below Amazon AWS, Microsoft Azure, and Oracle. Newer GPUs can command higher hourly rates by delivering greater throughput while reducing cost per token. Price competition has not simply depressed revenue; it has shifted pricing power from merely owning GPUs to deploying the latest GPUs fastest and maintaining high utilization.

What M7 cloud providers ultimately sell is optionality. Customers can combine multiple chips, regions, storage products, databases, and software tools while relying on cloud providers’ balance sheets for stable, long-term capacity. Neocloud providers sell speed and focus: faster access to the latest GPUs, simpler products, and lower prices. The former may charge more per unit, while the latter can still generate attractive returns at lower prices if utilization remains high.

Closing the gap requires neocloud providers to control both power and financing. GPUs without grid capacity cannot determine deployment timing; power without low-cost capital cannot support sustained purchases of next-generation equipment. IREN is expanding from power, land, and data centers into cloud operations, while CoreWeave is using large-scale debt to accelerate construction. For both models, the central question is whether revenue can grow faster than the cost of capital.

IREN represents the ideal neocloud expansion model. The company is targeting more than US$4 billion in contracted annual recurring revenue from capacity scheduled to enter service by year-end and approximately 800 MW of IT capacity by 2027. Customers can currently prepay roughly 45%–55% of GPU capital expenditure, while the company plans to finance approximately 90% of GPU costs; typical contract terms are 3–5 years. If power, customer prepayments, and asset refinancing all materialize as planned, equity funding requirements should decline and power resources can be converted into compute revenue more quickly.

The risks are embedded in the same structure. Most new capacity is scheduled to come online late in the quarter, so a meaningful revenue contribution is not expected until the following March quarter. Market estimates suggest a funding gap of approximately US$5.5 billion may remain beyond committed capital, GPU financing, and prepayments. Contracted annualized revenue, revenue from operating capacity, and cash collected are three different figures. Investors should prioritize capacity commissioning, utilization, refinancing rates, and changes in share count rather than focusing solely on total contract value.

CoreWeave illustrates the stressed version of the model. Second-quarter revenue increased 115% year over year to US$2.6 billion, while the company’s loss for the same period totaled US$626 million. Debt stood at US$35.6 billion, first-half interest expense alone reached US$985 million, and service prices rose 25%. Revenue is growing rapidly, but so are interest expense and construction funding needs. The price increase may reflect strong demand, or it may indicate that previous pricing failed to cover the cost of capital. The next indicators to watch are customer retention after the increase, gross-margin improvement, and debt growth—not headline growth in isolation.

Nvidia has paused some transactions with smaller AI cloud providers that exchanged credit support for revenue sharing, covering approximately US$36 billion of commitments. The reasons included concerns about potential antitrust scrutiny. Broader customer support has not stopped, but the move establishes a boundary for neocloud financing: chip suppliers can help customers secure compute capacity, but direct participation in customer revenue creates greater regulatory and conflict-of-interest risks. If the original arrangements are replaced with conventional debt, guarantees, or asset financing, neocloud funding costs and equity returns could both change.

This also explains why traditional software multiples are insufficient for valuing neocloud providers. These companies simultaneously assume data-center construction, equipment depreciation, energy procurement, cloud operations, and customer credit risk. Contracted revenue provides visibility, but the financing structure determines how much return accrues to shareholders. Higher debt costs, faster equipment-refresh cycles, or excessive customer concentration can make seemingly stable multiyear contracts fragile.

Divergent Views, Disconfirming Evidence, and Next Week’s Watchlist

The bullish evidence still predominates: Nvidia reported strong data-center revenue and next-quarter guidance, Amazon AWS increased purchases, Google expanded its TPU clusters, and neocloud providers signed multiyear contracts. Bearish evidence, however, has moved beyond the abstract concern that capital expenditure is simply too high. It now includes weaker operating cash flow, longer accounts-receivable days, declining gross margins, and expanding debt and guarantees. The greatest analytical risk is assuming that strong demand will generate equal returns for every participant.

The next quarter can be framed through three scenarios. In the bullish case, Rubin ships on schedule, cloud-provider utilization rises, Nvidia’s receivables days decline, and neocloud providers complete low-cost refinancing; capital expenditure would continue converting into revenue. In the base case, demand remains strong, but power and HBM constraints delay delivery, pressuring gross margins without causing order cancellations. The bearish case requires several signals to emerge simultaneously: customers extending payment terms, more expensive financing, insufficient utilization, and gross margins falling below the validation threshold.

No single metric is sufficient to identify an inflection point. Lower gross margins may reflect a new-product ramp, higher receivables may result from quarterly delivery timing, and neocloud equity issuance may fund the acquisition of scarce power. Only when these changes corroborate one another and persistently impair cash collection should the risk assessment escalate from “requires monitoring” to “deteriorating returns on capital expenditure.”

Next week’s validation should proceed sequentially from hardware delivery to cash collection

From an investment perspective, the companies fall into three categories. The first comprises M7 cloud providers with cloud revenue, customer relationships, and multi-chip choice, whose advantages are financing capacity and value-added services. The second includes neocloud providers that control power and land and can deploy new GPUs quickly, benefiting from speed and scarce capacity. The third consists of platform suppliers providing chips, networking, and financing support across the ecosystem; they offer the greatest revenue sensitivity, but their off-balance-sheet risks are also rising. All three benefit from broadening demand, but their disconfirming indicators are entirely different.

It is still too early to conclude that capital expenditure has become excessive, just as large orders alone do not validate returns. A more measured assessment is that demand is broadening, supply is being committed earlier, and capital responsibility is spreading with it. As long as delivery and utilization continue to improve, current investment can still convert into revenue. If collection periods, financing costs, and gross margins deteriorate simultaneously, however, the cycle will shift from an expansion challenge to a capital-recovery problem.

Related Reading

404K SEMI-AI 2026-08-28 Semiconductor Equipment Weekly — Demand Broadens: WFE Forecasts Rise, HBM Bottleneck Shifts to Packaging, Testing, and Yield Validation

404K SEMI-AI Morning Brief, 2026-08-12 — Financing Broadens: $500 Billion Platforms Drive Compute Buildout, While Power and Utilization Determine Returns

404K SEMI-AI Morning Brief 2026-08-10 — Demand Broadens: Agentic Storage, AI Cloud Financing, and Optical Interconnects Heat Up in Tandem

404K SEMI-AI Morning Brief 2026-08-01 — Diverging Returns: AWS Revenue Accelerates, Storage and Optical Communications Continue Raising Prices and Expanding Capacity

404K SEMI-AI Evening Brief 2026-07-28 — Expansion Shifts Gears: AI Hardware Demand Persists as Financing Costs and Supply Yields Enter the Validation Phase

• Overall View This Week

• Front-End Wafer-Fabrication Equipment

• Metrology, Inspection and Test

• Advanced Packaging Equipment and Critical Materials

• Company, Order, and Capex Validation

• Debates, Disconfirming Evidence, and Next Week’s Watchlist

Support for the equipment cycle is broadening from advanced logic and front-end memory into packaging, metrology, and testing. Order visibility is improving, but hybrid-bonding costs, packaging and testing capacity execution, and customer capital expenditure remain the key validation points for next week.

Overall View This Week

Semiconductor-equipment demand forecasts continue to rise. The broadening scope of spending is this week’s most important development, outweighing improvements in any single equipment category. Wafer-fabrication equipment (WFE) expectations continue to move higher, supported simultaneously by DRAM, NAND, and advanced logic. In the back end, HBM stacking, advanced packaging, and high-value chiplet systems are increasing investment intensity across bonding, alignment, metrology, inspection, and testing. Equipment vendors now have two growth drivers: more wafers and more process steps per wafer and per package.

Three sets of evidence now support this thesis. First, baseline WFE forecasts for 2026–2028 have risen across all 3 years, with the largest increase in 2027. Second, revenue or shipment plans are improving simultaneously across lithography, deposition, and memory equipment, indicating that demand is not confined to a single process. Third, OSAT capital expenditure, HBM stacking complexity, and the migration of testing upstream are all increasing, showing that equipment value is extending from front-end fabrication into packaging and yield protection.

Higher forecasts do not mean that all orders have already materialized. Some market expectations are approaching or even exceeding the bull-case scenario, reducing valuation tolerance for positive surprises. Investors must distinguish among the 5 stages of customer capital-expenditure announcements, equipment orders, installation and acceptance, production ramp, and revenue recognition. Focusing only on capex announcements can overstate near-term revenue, while focusing only on quarterly revenue can understate the medium-term visibility of long-lead-time equipment.

By earnings driver, 2026 equipment revenue reflects customer capacity additions, supply availability, and equipment price increases. Growth in 2027 depends more heavily on new fabs, memory-process transitions, and the concentrated ramp of advanced-logic capacity. The quality of growth therefore differs between the 2 years. If vendors raise prices without a corresponding increase in system shipments, revenue will still improve, but the resulting service and consumables stream will be less durable than one supported by genuine installed-base expansion.

Investors should focus on 2 types of companies. The first possesses scarce process capabilities that customers cannot readily bypass when expanding capacity, including advanced lithography, critical deposition and etch, and process control. The second operates in segments where equipment value is accelerating, including hybrid bonding, precision alignment, metrology and inspection, probes, and automated testing. The former offers order certainty; the latter provides upside from higher equipment value and testing intensity. The primary risk is the pace of adoption: before a technology roadmap is finalized, equipment qualification may lead revenue by several quarters.

Front-End Wafer-Fabrication Equipment

Front-end demand combines early strength in memory, catch-up growth in logic, and easing lithography supply. In the June quarter, Lam Research’s DRAM systems revenue doubled year over year, while Tokyo Electron’s DRAM sales increased 66% year over year. As NAND manufacturers transition to higher layer counts, WFE demand associated with new NAND fab space is projected to increase to 3x its prior level during 2026–2028. Higher foundry capital expenditure is also extending incremental demand from memory into advanced logic.

The implications for equipment vendors are straightforward. Lam Research and Tokyo Electron both benefit from the increasing number of memory-process steps; Applied Materials benefits from deposition, materials engineering, and advanced-packaging processes; and ASML is seeing simultaneous improvement in lithography demand from advanced logic and DRAM. Process-transition-driven equipment demand is generally more durable than demand from capacity expansion alone because changes in layer counts, materials, and structures increase the number of repeated process steps.

Front-end equipment exposure cannot be ranked solely by customers’ aggregate capital expenditure. For the same investment amount, high-layer-count NAND, advanced DRAM, and gate-all-around (GAA) logic require substantially more deposition, etch, cleaning, and metrology steps than straightforward mature-node capacity expansion. Investors should continue examining vendors’ system revenue by application and the respective contributions from new tools, chamber upgrades, and services to determine whether growth is being driven by unit volume, process intensity, or pricing.

Lithography demand is broadening, but the 3 companies occupy different product and profit positions

The table uses different currencies, so revenue scale is not directly comparable as a measure of competitiveness. ASML’s core advantages are scarcity and pricing power in advanced lithography, while Nikon and Canon have greater exposure to ArF, KrF, and specialty-process demand. An improving lithography cycle does not imply synchronized recovery across every product line: Canon’s flat-panel-display lithography business remains notably weak, indicating that semiconductor processes—not a broad recovery in optical equipment—are driving the upturn.

The challenge in advanced logic is shifting from whether a structure can be fabricated to whether it can be controlled consistently. In gate-all-around (GAA) nanosheets, geometric differences among stacked channels alter mechanical stress, affecting mobility, threshold voltage, and leakage. Process simulations show that increasing top-channel thickness from 4 nm to 7 nm can significantly narrow the stress differential between layers; under high lateral-etch-ratio conditions, the nFET stress differential can decline from approximately 900 MPa to approximately 500 MPa.

These challenges increase the value of computational process modeling, 3D metrology, and process control. As advanced nodes become more complex, customers can no longer wait until final electrical testing to identify systemic deviations. Equipment demand is therefore expanding beyond the core etch and deposition tools into process modeling, film-thickness and morphology metrology, defect inspection, and feedback control. Tool throughput remains important, but the ability to shorten customers’ process-learning cycles increasingly influences purchasing decisions.

Read more

• Pre-Market Highlights

• Full AI/Semiconductor Value Chain

• AI Models, Applications and Capital Expenditure

• AI Cloud and Data-Center Operators

• GPU/CPU/ASIC

• HBM/DRAM/NAND/SSD/HDD

• Foundries and Semiconductor Equipment/Testing

• Optical Communications and Components

• High-Speed Interconnects, Connectors, Cooling and Power

• Space/Satellites

• Internet/Platforms

• Software/SaaS

• Consumer Electronics/Smart Vehicles

Overview

404K SEMI-AI | 2026-08-28

Pre-Market Highlights

The pre-market technology narrative continues to broaden toward infrastructure delivery. Nvidia says next year’s capacity is already sold out, but incremental compute deployment now depends on more than GPU supply: power, grid connections, storage, packaging, networking and optical components must all arrive on schedule. Order visibility remains high, but the lag to revenue recognition is widening. Power-on schedules now matter more than headline budgets.

Storage and interconnects are today’s two most active incremental themes. Samsung Electronics is advancing 8-layer NVHBM, while SK hynix has set Q3 2029 as the mass-production target for HBM4E in the US. Lumentum and Marvell Technology are connecting stronger orders and higher revenue forecasts with CPO, NPO and CXL ramp-ups. Qualification and yield still separate announced capacity from usable supply, so new-factory announcements will not immediately relieve shortages.

Application-layer monetization and risk-control boundaries are becoming clearer. OpenAI is testing contextual advertising, while Cognition’s annualized revenue is rising rapidly. Whether enterprise agents can move from demos into production will depend on task-completion rates, access-control accountability and insurance coverage. Application revenue is growing, but enterprise acceptance standards are rising in parallel. Commercialization and liability frameworks must develop together, with retention and claims ratios the next metrics to watch.

Full AI/Semiconductor Value Chain

AI Models, Applications and Capital Expenditure

• AI infrastructure delivery: Conventional data-center counts are losing explanatory value. The US has roughly 5,400 buildings classified as data centers, about 10 times the number in its nearest competitor, yet the US, China and Europe accounted for approximately 45%, 25% and 15%, respectively, of 2024 electricity consumption. Many legacy colocation facilities cannot support rack densities above 100 kW, direct liquid cooling or massive east-west traffic. The real determinants of AI-factory revenue recognition are energized capacity, accelerator deployment and construction completion. For Stargate UAE, investors must also distinguish between the initial 200 MW phase in 2026, the 1 GW cluster and the longer-term 5 GW campus plan.

“Investors should stop counting buildings and start tracking energized megawatts, accelerator deployments, grid-connection approvals and construction completion rates.”

• Hyperscaler capital expenditure: A common estimate for major cloud providers’ 2026 capital expenditure is approximately $800 billion. Goldman Sachs expects global AI investment to exceed $1 trillion over the same period, including approximately $581 billion in the US. Google’s TPU, Amazon’s Trainium and Inferentia, and Microsoft’s Maia show that budgets are also shifting toward internally developed chips. The former figures also include overseas and non-AI spending and therefore cannot be treated entirely as procurement. Projects generate supplier revenue only when they connect to the grid and come online on schedule.

AI Cloud and Data-Center Operators

• Hut 8: The company has contracted 949 MW of IT capacity, representing $27 billion in base-term contract value and expected average annual net operating income above $1.75 billion, with a long-term project pipeline of approximately 8.7 GW. Long-term contracts, customer prepayments and locked-in costs provide cash-flow visibility, but delivery at scale remains unproven; last quarter’s net result was a loss, with the reported amount totaling $177.1 million. Beacon Point financing, a potential River Bend transaction and on-schedule project delivery are the 3 milestones for converting contract value into cash flow.

• Cipher Mining: The company has contracted 700 MW of capacity, corresponding to approximately $12 billion in revenue contracts. Estimated 2028 project net operating income is approximately $727 million, with a long-term project pipeline of approximately 4.1 GW. Black Pearl’s early delivery has strengthened management credibility, but the 2028–2035 cash flows remain long-duration and the company reported a net loss of $267.5 million last quarter. Commissioning schedules for Odessa, Reveille, Ulysses and other projects, together with interest rates and refinancing costs, will continue to materially affect value realization.

• Galaxy Digital: The company is leasing 800 MW to CoreWeave for a base term of at least 15 years. Once Helios is fully delivered, expected annual net operating income is approximately $1 billion, with potential power capacity of approximately 5.7 GW. Galaxy has demonstrated the full cycle from customer contracting and financing through construction and delivery. The next milestone is a potential Helios II lease in Q1 2027. Data-center assets remain housed alongside the crypto business, so earnings volatility and conglomerate discount risk cannot be ignored. The value of the remaining power capacity depends on whether the lease model can be replicated.

“3 colocation providers as anchors + 2 neoclouds to capture higher value per megawatt.”

• IREN
  1) AI cloud is targeting annualized recurring revenue above $4 billion from year-end operating capacity, 85% of which is already contracted. The company targets approximately 800 MW of IT capacity in 2027. Much of that capacity is scheduled to come online late in the quarter, with most revenue expected to appear in the following March quarter.
  2) Customers can prepay approximately 45%–55% of GPU capital expenditure, while the company can finance approximately 90% of GPU costs.
  3) The company has secured 5.8 GW of power, and its 5-year Microsoft agreement is worth $9.7 billion. Whether the approximately $3 billion Horizons data-center investment can be refinanced successfully will determine its equity-funding requirement.

• Nebius: Revenue increased 454% year over year to $582 million, while the core AI cloud business reached a 50% adjusted EBITDA margin. Contracted power has risen to 5 GW. Growth and execution are leading the sector, but expansion requires substantial capital expenditure. Nebius plans to add more than 1 GW annually from 2027 and bring most of its existing portfolio online during 2029–2030. Contracting power is only the first step: electricity supply, equipment, software utilization and customer contracts must align to determine how quickly capacity becomes revenue. A delay in any link will push revenue into later quarters.

• CoreWeave: High compute utilization can generate strong revenue per megawatt, but CoreWeave does not own the underlying power assets. Its debt costs and lease structure leave it more dependent on GPU residual values, customer renewals and financing markets. Relative to IREN and Nebius, which own power-backed land, CoreWeave is more sensitive to external assets and the cost of capital. Nvidia’s suspension of certain AI-cloud revenue-sharing arrangements also reminds investors to distinguish among chip-supply support, customer financing and long-term control of the business.

GPU/CPU/ASIC

• Rack power density: H100 SuperPOD is approximately 41 kW, GB200 approximately 120 kW, GB300 approximately 150 kW, Vera Rubin approximately 227 kW, and Rubin Ultra is expected to reach 660 kW. Deploying 90,000 NVL72 racks at approximately 225 kW each in 2027 would require approximately 23 GW of total power and 204 TWh of annual electricity at a 1.15 PUE. Even if half were deployed in the US, 102 TWh would equal 82% of projected incremental US electricity generation in 2026–2027—and that excludes Blackwell, AMD accelerators, TPU, Trainium, Maia, CPU and networking equipment.

• Nvidia
  1) Jensen Huang says next year’s capacity is already sold out and 70% growth remains supply-constrained. Following memory price increases, next year’s gross-margin guidance has been reduced from 75% to 72%–73%. Vera Rubin is expected to ramp faster than Blackwell.
  2) The company has suspended certain credit-support and revenue-sharing transactions for smaller AI cloud providers, covering approximately $36 billion of commitments, although broader support for access to compute has not stopped.
  3) AWS plans to add 2 million GPUs during 2027–2028. Approximately half of Nvidia’s revenue comes from roughly 6 hyperscalers, and most of that compute will be subleased to developers, requiring investors to track customer concentration alongside end-user utilization.

“Constraints include TSMC CoWoS packaging, wafers, memory, networking, silicon photonics, power generation and ‘high-quality land with power,’ compounded by a shortage of skilled labor in the US.”

HBM/DRAM/NAND/SSD/HDD

Read more

• Table of Contents

• What 2Q Results Answered

• Why Profit Grew Faster Than Revenue

• How AI PCBs Turn Compute Investment into Orders

• Why IC Substrates Are Becoming the Second Growth Curve

• How High-Speed Optical Module PCBs Are Upgrading

• How Capacity Expansion Will Reshape Profit and Cash Flow

• What the Earnings Forecasts and Target Price Imply

• Which Public Numbers to Track in the Coming Quarters

本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读

Shennan Circuits' second-quarter revenue, gross margin, and profit all accelerated simultaneously, with AI PCB and IC substrates now forming a dual engine. Whether the valuation can continue to move higher will depend on whether capacity expansion, working-capital absorption, and cash recovery can keep pace with profit growth.

Table of Contents

1. What 2Q Results Answered

2. Why Profit Grew Faster Than Revenue

3. How AI PCB Converts Compute Capex into Orders

4. Why IC Substrates Became the Second Growth Curve

5. How High-Speed Optical-Module PCB Upgrades Along the Path

6. How Capacity Expansion Will Reshape Profit and Cash Flow

7. What Earnings Forecasts and the Target Price Imply

8. Which Public Figures to Track in Coming Quarters

What 2Q Results Answered

Shennan Circuits' second-quarter results confirmed a key shift: high-end PCB demand and rising IC substrate utilization are simultaneously lifting revenue and margins. Morgan Stanley's single-quarter data show 2Q26 revenue of RMB 8.701 billion, up 32% QoQ and 53% YoY, beating the firm's estimate and consensus by 12% and 14%, respectively. Single-quarter gross margin reached 31.7%, up 3.0 ppts QoQ and 4.8 ppts YoY. Operating profit was RMB 1.781 billion, up 64% QoQ and 108% YoY; net profit attributable to shareholders was RMB 1.401 billion, up 65% QoQ and 61% YoY. (Morgan Stanley, August 27, 2026, p. 4)

The significance of these numbers goes beyond "orders increased." Strong YoY and QoQ revenue growth indicates the demand expansion is not a single YoY effect driven by a low base. Operating profit growth clearly outpaced revenue, showing that utilization, product mix, and cost control jointly produced operating leverage. The 31.7% gross margin was 2.1 ppts above Morgan Stanley's estimate and 1.2 ppts above consensus — earnings elasticity was the quarter's most important expectation gap.

The first-half official statements provide a more complete accounting base for the single-quarter changes. The company reported revenue of RMB 15.296 billion, up 46.33% YoY; net profit attributable to shareholders of RMB 2.251 billion, up 65.55% YoY; and ex-nonrecurring net profit of RMB 2.239 billion, up 76.94% YoY. Nonrecurring items were only about RMB 12 million, so attributable and ex-nonrecurring net profit were very close — profit growth came mainly from the core business. Basic EPS was RMB 3.32, and weighted-average ROE was 12.43%. (Shennan Circuits 2026 Semi-Annual Report, pp. 8, 49–50)

Bridging Morgan Stanley's 2Q data with the company's first-half statements yields roughly RMB 6.595 billion of 1Q revenue and about RMB 850 million of 1Q attributable net profit. Second-quarter revenue rose by about RMB 2.106 billion and net profit by about RMB 551 million, implying an incremental net margin of about 26%. This rough bridge is subject to single-quarter classification and rounding, but the direction is clear: the profit content of 2Q incremental revenue was significantly higher than the first-half average.

The market's two biggest prior concerns were whether AI-related PCB demand would remain stuck at sampling and small-batch stages, and whether IC substrate capacity expansion would drag on profits for an extended period. Second-quarter data provided positive evidence on both questions. Data-center-related order revenue has already exceeded RMB 2 billion, packaging substrate revenue doubled YoY, and gross margin rose sharply. Current evidence still covers only one strong reporting period; sustainability needs to be confirmed by subsequent orders, utilization, and cash flow.

The profit-acceleration transmission can be summarized in six links: AI compute investment rises, high-end PCB and packaging substrate orders grow, line utilization and product mix improve, gross margin drives profit acceleration; the company then uses more capex and working capital to capture demand, and cash returns determine the level at which the valuation can be sustained.

Why Profit Grew Faster Than Revenue

The most direct explanation for first-half profit growth outpacing revenue is the improvement in consolidated gross margin. Based on segment data, first-half consolidated gross margin was about 31.01%, versus about 26.28% a year earlier — an increase of 4.73 ppts. On revenue of RMB 15.296 billion, that margin improvement corresponds to a gross-profit increase of nearly RMB 720 million. That magnitude explains most of the roughly RMB 891 million YoY increase in attributable net profit; the remainder comes from revenue scale expansion, the expense structure, and other operating items.

All three main businesses grew, but contributed differently. PCB revenue was RMB 8.552 billion, up 36.32% YoY, with gross margin of 36.85%, up 2.43 ppts YoY; packaging substrate revenue was RMB 3.667 billion, up 110.76% YoY, with gross margin of 31.35%, up 16.20 ppts YoY; electronic assembly revenue was RMB 1.976 billion, up 33.70% YoY, with gross margin of 17.30%, up 2.32 ppts YoY. (Shennan Circuits 2026 Semi-Annual Report, pp. 13, 16)

PCB remains the core of revenue and profit. It has the highest gross margin and accounts for roughly 60% of the combined revenue of the three disclosed segments, so its 2.43-ppt margin improvement carries substantial absolute contribution. Packaging substrate revenue is only about 43% of PCB's, but its revenue increased by about RMB 1.9 billion YoY and its margin improvement was far larger than the other two segments, making it the strongest source of marginal profit growth. Electronic assembly maintained steady expansion; its lower gross margin means it primarily plays a customer-service, system-assembly, and revenue-supplement role.

The advantage of this mix is that the two growth curves are at different stages. High-end PCB already has a substantial revenue base, and product upgrades driven by AI servers, switches, and high-speed optical modules can convert directly into scaled profit. IC substrates previously bore the pressure of capacity ramp-up, yield, and depreciation; after utilization improved, margin recovery accelerated. The simultaneous improvement in both businesses gave the company a compounding effect of revenue growth and gross-margin expansion in the second quarter.

The expense side did not show across-the-board contraction. R&D; expenses were fully expensed, administrative expenses were affected by personnel compensation and equity incentives, and financial expenses were dragged by a swing in FX from gains to losses. Rapid revenue growth diluted some expense ratios, but management, financing, and share-based payments still consumed part of the gross-profit increase. (Shennan Circuits 2026 Semi-Annual Report, p. 14, p. 52, and related notes)

Expense growth remained concentrated in management, financing, and share-based payments

R&D; growth was slower than revenue growth, so the R&D; expense ratio was diluted by rapidly expanding revenue. The company holds 1,031 authorized patents, including 614 invention patents and 128 PCT patents; backplane R&D; samples have reached up to 120 layers, and mass-production products up to 68 layers. These metrics do not directly equate to commercial revenue, but they show that the technical barriers and certification accumulation in high-end PCB have tangible substance. (Shennan Circuits 2026 Semi-Annual Report, pp. 11–14)

Profit quality still has two deduction items that require ongoing monitoring. First-half credit impairment losses were about RMB 137 million, and inventory write-downs plus contract-asset impairments were about RMB 207 million; the two together totaled about RMB 344 million. As inventory and accounts receivable grow rapidly, impairment charges may continue to rise. If future revenue growth slows, the positive operating leverage from gross margin could be reversed by inventory, bad debts, and fixed costs.

How AI PCBs Turn Compute Investment into Orders

The core business logic of AI PCBs is that customers pay higher unit prices for more layers, lower loss, higher density, and more reliable delivery. As server, switch, and optical module transmission rates upgrade, the requirements for signal integrity, thermal management, processing precision, and reliability that PCBs must meet rise significantly. Unit value and manufacturing difficulty for high-end products increase in tandem, and qualified suppliers must also pass a lengthy certification cycle, so order growth translates more readily into product mix and gross margin.

The company's data-center-related order revenue exceeded RMB 2 billion in the first half, already reaching roughly 13% of total first-half revenue. This remains a "related orders" measure and cannot simply be treated as pure AI revenue. It at least demonstrates that data-center demand has moved from R&D; validation into meaningful revenue scale and is genuinely occupying high-end PCB capacity. (Shennan Circuits 2026 Semi-Annual Report, pp. 11–14)

Morgan Stanley noted in its Q2 discussion summary that AI-related PCB business continued to grow, with overall PCB revenue up roughly 20% quarter-over-quarter. The company's Wuxi high-end PCB project targets AI server and switch demand, with a project budget of RMB 4.278 billion, cumulative investment of RMB 962 million, and progress of 22.48%. The low cumulative progress means most capacity and depreciation are still to be released, so revenue scale and execution risk will both increase over the next two years. (Morgan Stanley, August 27, 2026, p. 1; Shennan Circuits 2026 Semi-Annual Report, pp. 18–19)

Read more

• After-Hours Summary

• U.S. Stocks at New Highs

• Three U.S. Stock Rankings

• Full AI and Semiconductor Value Chain

• AI Models, Applications, and Capital Expenditure

• CSP and Cloud Capital Expenditure; AI Clouds

• GPUs, CPUs, ASICs, and High-Speed Interconnects

• Memory, Wafers, Packaging, and PCBs

• Internet/Platforms

• Software/SaaS

• Consumer Electronics / Smart Vehicles

• Jensen Huang’s Curated Portfolio

• Memory Index Performance

• Memory’s Share of US Equity Trading Turnover

Overview

404K SEMI-AI | 2026-08-28

After-Hours Summary

The S&P; 500 rose 0.51%, the Nasdaq 100 gained 1.10%, and the equal-weighted S&P; 500 fell 0.32%, as megacap technology stocks continued to drive the broader market. Technology ETFs rose 2.77% and semiconductor ETFs gained 2.33%, while communication-services ETFs declined 0.72%. The clearest market signal was the return of capital to technology following Nvidia’s earnings.

Single-stock dispersion was even more pronounced than the index moves. Nvidia rose 7.78%, Salesforce gained 22.58%, and CrowdStrike climbed 20.50%; Amazon, Meta, Google, and Netflix fell 1.54%, 0.80%, 0.36%, and 1.99%, respectively. Investors are again willing to pay for AI demand and software bookings, but remain selective on internet platforms and capital-intensive businesses.

Nvidia raised its FY2028 revenue-growth outlook to approximately 70%, temporarily easing concerns about demand and shifting the hurdle to execution. Procurement commitments, HBM, wafers, packaging, power, and customer financing are all scaling simultaneously. On the software side, cRPO, ARR, renewals, and margins must demonstrate that AI can generate real revenue rather than merely support a product narrative.

U.S. Stocks at New Highs

Three U.S. Stock Rankings

Full AI and Semiconductor Value Chain

AI Models, Applications, and Capital Expenditure

• OpenAI
  1) Its revenue run rate has increased from $13 billion last August to more than $40 billion, indicating that paid demand continues to expand.
  2) Codex is testing a persistent-work mode that can create follow-up tasks across sessions, although there are no plans for an immediate release. The next tests are whether growth can remain strong against a higher base and whether long-running Agent tasks materially increase inference calls.

“OpenAI and Anthropic are growing as fast as, or faster than, virtually any companies of comparable scale in history.”

• Anthropic
  1) Its revenue run rate reportedly reached $65 billion at the end of July; the combined OpenAI and Anthropic run rate increased from $30 billion to $105 billion between the start of 2026 and August.
  2) Its Vera Rubin lease with Nscale runs for 6 years, extending the contract duration from 4 years in the Hopper era.
  3) Model Hardware Standard enables Claude to operate physical devices through a shared driver layer, expanding the commercial opportunity from software into laboratory hardware.

• Frontier-model revenue: The generative-AI industry reached a deduplicated annualized run rate of approximately $175 billion in June 2026, with OpenAI and Anthropic accounting for most of the total. Further rapid growth may depend on coding Agents crossing a capability threshold, while the risk is that coding demand plateaus over the coming months. Quarterly revenue growth is a better test of paid utility than any single benchmark.

“Revenue—and the investment it supports—also feeds the compute flywheel: better AI generates more revenue, more revenue funds more compute, and more compute produces better AI.”

• Agent applications: Agents turn a single task into many inference calls, shifting the relevant cost metric from cost per token to total cost per completed task. Even small changes in browsing order or memory can alter a shopping Agent’s preferences, requiring platforms to manage outcome stability, inference costs, and auditability simultaneously.

• Inference architecture: Prefill is primarily compute-bound, while decode depends more heavily on memory bandwidth and the KV cache. Running a 70-billion-parameter FP16 model on an H100 requires reading approximately 140GB of weights per token; a 128K context can generate around 40GB of KV data. The bottleneck has not disappeared—it has shifted from weight movement to long contexts, batching, memory, and networking.

“The fact that AI inference was never a single workload has finally been etched into the hardware itself.”

CSP and Cloud Capital Expenditure; AI Clouds

• Microsoft: New Azure AI backlog was slightly above $50 billion, all from customers other than OpenAI and Anthropic, indicating that enterprises and other model providers are also expanding capacity. Demand will continue to exceed supply this year; FY2027 capital expenditure is expected to grow year over year, with first-quarter capital expenditure exceeding $50 billion. M365 Copilot has more than 30 million paid seats.

• Amazon
  1) AWS has received its first Vera CPU server and Vera Rubin GPU. Vera features 88 custom cores and memory bandwidth of up to 1.2TB/s.
  2) Annapurna Labs will be the first to work with NVHBM, while TRAINIUM4 will support NVLink Fusion. The relationship between proprietary XPUs and Nvidia’s rack-scale platform is shifting from substitution toward hybrid deployment.

• IREN: Contracted ARR tied to 2026 capacity has reached $4 billion, of which approximately $1 billion is already operational, leaving capacity effectively sold out. Recent 3-year contracts generate more than $20 million per megawatt, with negotiated prices approaching $25 million and payback periods of approximately 2 years. Customer prepayments cover 45% to 55% of GPU capital expenditure, making the financing structure critical to the equity dilution required for expansion.

• CoreWeave: Second-quarter revenue rose 115% year over year to $2.6 billion, but the company posted a quarterly loss of $626 million, incurred $985 million of interest expense in the first half, and carried $35.6 billion of debt. A 25% increase in service pricing can improve unit revenue, but also exposes the margin pressure embedded in legacy pricing. The real test is whether growth can outpace interest and refinancing costs.

“The key issue to validate is earnings quality: operating cash flow fell sequentially from $50.3 billion to $24.1 billion, accounts receivable rose to $63.1 billion, and DSO lengthened from 45 days to 60 days.”

• Nebius: The company has contracted more than 2 GW of power, but only approximately 170 MW of early revenue-generating capacity, leaving a substantial gap between contracted power and actual deployment. Emerging cloud providers must control substations, power distribution, construction schedules, and their balance sheets—not merely secure GPUs. Key metrics are revenue per megawatt and the conversion rate from contract signing to energization.

• Oracle: GPU utilization is close to 97.5%, demonstrating that operational capacity can be converted rapidly into revenue. Relative to Nebius and CoreWeave, Oracle benefits from high utilization and a mature balance sheet. The risk remains that grid connections, transformers, and power-distribution timelines delay new capacity, making GPU purchase volumes an insufficient standalone indicator.

• AI data-center supply: North American demand is expected to reach approximately 35 GW in 2027, versus only 16.5 to 23.4 GW of deliverable supply, implying a gap of approximately 1.8 to 2.1 times. Grid interconnection waits can reach 4 to 5 years, while lead times for high-voltage transformers exceed 160 weeks. Power constraints accounted for 60% to 70% of megawatt-weighted delays over the past 12 months, and capital alone cannot compress these timelines quickly.

GPUs, CPUs, ASICs, and High-Speed Interconnects

• Nvidia
  1) Quarterly revenue reached $96.2 billion, up 106% year over year, including data-center revenue of $89.0 billion. The company’s FY2028 revenue-growth outlook is approximately 70%.
  2) Purchase and supply commitments rose to $279.0 billion, primarily to secure HBM, while gross-margin guidance declined to 74.0%.
  3) Relative to standard HBM4E, NVHBM can increase bandwidth by up to 30%, reduce power consumption by 15%, and free up as much as 25% of the XPU compute-die area. Demand has been validated; inventory, cash collection, and guarantee exposure are now the key quality metrics.

“Gross margin is expected to decline from 75% this quarter to 74% in F3Q, potentially bottom at 71% to 72% in F4Q, and stabilize at 72% to 73% in FY28, reflecting pricing pressure from HBM and other memory.”

• Marvell Technology: Second-quarter revenue was $2.74 billion, up 37% year over year, including data-center revenue of $2.17 billion, up 46%. Third-quarter revenue guidance of $3.15 billion exceeded the $3.03 billion consensus. Adjusted gross margin was 58.9%, down 50 basis points year over year, indicating that AI bookings remain strong while product mix and costs continue to constrain margins.

• Groq: Nvidia’s approximately $17 billion non-exclusive licensing transaction with Groq pairs GPUs for prefill with SRAM-based LPUs for decode. An LPX rack contains 256 LP30 chips, 128GB of SRAM, and 12TB of DDR5. The vendor claims throughput per megawatt can reach up to 35 times that of GB200 under specific conditions, although GPUs retain an advantage for large batches and workloads with looser latency requirements.

• AMD: Cerebras, Nvidia, and Groq are all pursuing workload-specific disaggregation of inference. AMD’s Venice server CPU provides a test of CPU-side growth, while the GPU-side test is task-level cost when operating alongside specialized SRAM architectures. If unified engines prove more effective in scheduling and fault recovery, the benefits of heterogeneous architectures may be lower than currently expected.

• Lumentum: Management reported a substantial increase in backlog and set a target for earnings power of more than $40 per share in FY2028, versus $8.67 in FY2026 and the current sell-side estimate of approximately $33. OCS revenue is expected to exceed $100 million next quarter, while the NPO timeline has moved forward to late 2027 or early 2028. Laser qualification and lead times for substrates and equipment remain the key volume-production constraints.

• Arista: AI back-end networks, campuses, and core data centers are jointly driving networking demand, with scale-across already accounting for approximately one-third of AI revenue. Investors must distinguish between cross-rack networking and within-rack scale-up, which have different chip, optical-interconnect, and software economics. The next test is whether scale-up can reach meaningful scale in 2027.

Memory, Wafers, Packaging, and PCBs

• SK hynix
  1) HBM is expected to account for more than 50% of revenue in 2027, 2028, and beyond—a higher share than at Micron and Samsung Electronics.
  2) The company has broken ground on a $4 billion packaging facility in Indiana, with mass production of next-generation HBM planned for the second half of 2029; wafers will continue to be produced in South Korea.
  3) Management expects strong demand to persist through the end of 2030, although additional investment will depend on incentives, operating conditions, and customer collaboration.

• HBM industry: Bank of America expects the market to grow from approximately $77 billion in 2026 to approximately $280 billion in 2030, representing a compound annual growth rate of roughly 38%. HBM4E’s wafer-displacement ratio versus conventional DRAM could rise to 4 times. Value is being created simultaneously through bandwidth, logic base dies, and stacking yields, further crowding out conventional DRAM supply.

• Kioxia: The company plans to build a new memory facility in Iwate Prefecture, targeting mass production in FY2029. Together with its Mie Prefecture facility, total investment over the next 6 years will reach ¥5 trillion. The order and expansion trajectory is clear, but new capacity remains years from operation; returns will depend on NAND pricing, demand for high-speed AI storage, and actual utilization.

• TSMC: Advanced-packaging projects generated virtually no business from 2011 until around 2020; AI demand finally made heterogeneous integration a critical capacity requirement. TSMC rose 2.30%, as advanced packaging shifted from a long-term R&D; project into a mass-production bottleneck for GPUs, HBM, and networking chips. The next tests are CoWoS capacity expansion and yields on new architectures.

• PCB materials: At 10GHz, PTFE has a dielectric constant of approximately 2.1 and dielectric loss of approximately 0.0003, enabling longer high-speed copper interconnects. However, drilling, hole-wall quality, dimensional stability, and repeated lamination have yet to achieve stable mass production. Rubin Ultra may use a hybrid structure combining a PTFE core with M9 prepreg, or may shift to M10, which remains at an early sampling stage.

“The key issue is not whether PTFE can deliver low loss, but whether it can be processed into qualified PCBs consistently, repeatedly, and at scale.”

• Memory hierarchy: Rubin uses 288GB of HBM4 for training, prefill, and large batches; LP30 uses 500MB of SRAM for low-latency decode; and Crescent Island emphasizes capacity economics with up to 480GB of LPDDR. Value is spreading beyond the standalone GPU into memory tiers, packaging, rack-level networking, and scheduling software.

Internet/Platforms

• Meta
  1) Hatch Agent reportedly runs continuously in the background and connects to email, calendars, and consumer apps; pricing as high as $200 per month was considered.
  2) Meta’s annual spending on Anthropic may at one point have reached $10 billion and remains several hundred million dollars per month. External procurement is proceeding alongside in-house initiatives such as Muse Code and Watermelon. Shares fell 0.80% as the market continues to assess product execution and the returns on heavy spending.

“Overall market volume was quite light, making it difficult to read too much into the price action.”

• Netflix: One model generates “Because You Watched” recommendation copy, while another scores it, with humans reviewing drift weekly. During a 5-week test, members shifted slightly toward unwatched content and more often ended browsing by starting playback. Shares fell 1.99%; the key metric is whether better recommendation explanations consistently improve viewing conversion.

“A judge running in production has a lifecycle and must be maintained like any other model.”

• Google: Shares fell 0.36% but gained 2.13% over 20 days. TPU 8i represents an inference architecture optimized for Google’s own workloads. The platform’s value depends on whether its models, chips, and cloud services can jointly reduce per-task costs; if customers increasingly favor multi-vendor deployments, the lock-in power of a single architecture will decline.

• Amazon: Shares fell 1.54% but gained 8.83% over 20 days. AWS’s adoption of Vera Rubin and integration of TRAINIUM4 with NVLink Fusion indicate that proprietary chips and Nvidia’s platform will coexist over the long term. Near-term momentum in the internet business remains weak, while cloud capacity delivery will depend on power availability, utilization, and customer contracts.

• Shopify: Shares rose 2.69% and gained 26.09% over 20 days, making Shopify one of the few platform stocks maintaining strong momentum. The rally did not extend to most advertising and local-services platforms, suggesting investors still favor companies with clearer growth trajectories. Sustainability must next be validated by GMV, merchant services, and margins.

• Reddit: Shares fell 0.91% and declined 13.53% over 20 days, with the near-term recovery clearly lagging Shopify and leading software names. Both AI data licensing and advertising must demonstrate growth quality. Without new operating data, relative price performance signals market disagreement but cannot substitute for revenue and user metrics.

• Uber: Shares fell 1.96% but gained 9.35% over 20 days. Avride has completed 100,000 autonomous trips through Uber in Dallas, still with safety operators onboard, and targets fully driverless operations in the first half of 2027. Whether the platform can convert autonomy into lower fulfillment costs will depend on fleet scale, intervention rates, and commercial miles.

“A safety operator remains in the seat in this footage, but fully driverless operations are expected in the first half of 2027.”

• DoorDash: Shares fell 2.13% but gained 17.39% over 20 days. Medium-term strength alongside a daily pullback shows that local-services platforms did not participate in the broad AI and software rally. Attention should return to order frequency, per-order fulfillment profit, and automation investment rather than inferring demand changes from one day’s share-price move.

• Airbnb: Shares fell 1.95% but gained 21.25% over 20 days, broadly mirroring DoorDash. Travel platforms faced near-term pressure from sector rotation, but the 20-day trend remains strong. Booking growth, average nightly rates, and margins are the relevant testable variables; current data are insufficient to interpret the pullback as fundamental deterioration.

• Pinterest: Shares fell 1.84% and declined 6.34% over 20 days, underperforming among advertising platforms. Research on agentic shopping shows that small changes in page-view order and memory can materially alter AI preferences. Platforms will need both to compete for user attention and to accommodate machine agents whose purchasing paths cannot be predicted reliably.

“Even small differences—page-view order and memory—change AI preferences in unpredictable ways.”

• Snap: Shares fell 1.66% but gained 13.65% over 20 days. Near-term performance has been stronger than Pinterest and Trade Desk, although there is no new operating anchor. A recovery must still be demonstrated through advertising pricing, user engagement, and augmented-reality monetization, rather than platform-sector rotation alone.

• AppLovin: Shares rose 1.47% but declined 22.59% over 20 days, marking a one-day rebound within a still-weak medium-term trend. The advertising-algorithm company must reverse that trend through improved advertiser returns and renewed platform-revenue growth. If AI tools only improve bidding efficiency without expanding budgets, valuation recovery will remain constrained.

• Trade Desk: Shares rose 2.99% but declined 26.59% over 20 days, the clearest oversold rebound among internet platforms. Whether programmatic-advertising demand has bottomed still requires validation from subsequent revenue and customer retention; the daily gain does not erase the stock’s relative weakness over the past 20 days.

• Spotify: Advertising showed no signs of accelerating in the second half, leaving the platform still needing to demonstrate that user scale can translate into greater ad monetization. As capital rotates from internet stocks toward software and Nvidia, companies lacking an advertising inflection are more vulnerable. The next key indicator is ad growth, not any single product launch.

“Meta is using Anthropic models to help support and test its forthcoming personal AI agent, Hatch, while developing competing products internally.”

• Agentic Shopping: Research shows that differences in page order and memory are sufficient to produce unpredictable agent choices, potentially undermining traditional human-oriented attribution and marketing methods. Platforms will need machine-readable product information, authorization controls, and outcome audits; commercial value will depend on converting unstable choices into controllable conversions.

Software/SaaS

• Salesforce
  1) Revenue was $11.345 billion, up 10.8% year over year; excluding Informatica, organic growth was approximately 6.4% to 7%. cRPO reached $33.5 billion, up 14% in constant currency.
  2) Agentforce ARR exceeded $1.5 billion, up 240% year over year, with renewals and expansions accounting for roughly half of AI bookings.
  3) Claudeforce offers 37 prebuilt sales skills, with public beta planned for September. As Claude takes over the interaction layer, Salesforce must defend revenue through data, permissions, workflows, and headless APIs.

“The core improvement in CRM came from high-quality bookings rather than current-period revenue: revenue reached $11.345 billion, up 10.8% year over year, but organic growth excluding Informatica was only approximately 6.4% to 7%.”

• CrowdStrike: Revenue was $1.471 billion, up 25.8% year over year; ARR reached $5.841 billion, with net new ARR of $332 million. Fiscal 2027 guidance for net new ARR growth was raised to 34%. Flex ARR reached $2.29 billion, up 101% year over year, and AIDR nearly doubled quarter over quarter, although endpoint security still accounted for approximately 90% of the growth improvement.

“AI is initially supporting bookings indirectly through identity modernization and platform consolidation, with material revenue potentially emerging only after FY28. Counterevidence includes DBNRR of just 107%, a mature core access-management market, and competition from Microsoft’s bundling.”

• Okta
  1) Revenue was $805 million, up 10.6% year over year; cRPO grew 14.1%, and adjusted operating margin was 28.1%.
  2) New products contributed approximately 30% of bookings, while deals incorporating new products were approximately 40% larger. One customer expanded from roughly 50 Claude Agents to 1,500 within several weeks.
  3) Jefferies raised its price target to $200, but direct AI revenue remains small. The key test is whether cRPO can continue to accelerate.

• SentinelOne: Second-quarter revenue was $292 million, up 21% year over year; ARR reached $1.22 billion, up 22%, with 1,715 customers generating more than $100,000 in ARR. Adjusted operating margin increased from 2% to 10%, but free cash flow remained negative at $13.2 million. Growth and cash-flow improvement must progress together.

“This appears more like a disappointment in expectations and positioning than a material fundamental problem.”

• Elastic: Shares rose approximately 17% after hours after revenue of $478 million and EPS of $0.70 both exceeded expectations. Elastic Cloud and sales-led subscriptions stabilized, while full-year operating-margin guidance increased to 19.4%. Cloud stabilization was the principal driver; the next test is whether cRPO and renewals can sustain the improvement, rather than a one-time margin increase.

• Workday: EPS of $2.75 and an operating margin of 31.1% exceeded expectations, while AI contributed more than 25% of new contract value. However, subscriptions, cRPO, and remaining performance obligations were relatively flat. Remaining performance obligations totaled $27.4 billion, below expectations of approximately $28.5 billion. AI-driven pipeline is emerging, but revenue conversion remains slow.

• Autodesk: Quarterly operating margin reached 41.0%, but third-quarter EPS guidance was weak. The increase in full-year revenue guidance includes the consolidation of MaintainX, while organic billings showed no clear inflection. Margins are outperforming revenue; investors must distinguish acquisition contributions from demand in the legacy business.

• Rubrik: Revenue reached $427 million and operating margin was 7.8%, with the company also raising its full-year outlook. However, the shares remained constrained by the higher expectations established by CrowdStrike and Okta. The divergence between fundamentals and price reaction shows that the benchmark for software earnings has shifted from consensus expectations to the strongest peer performance.

• MongoDB: Atlas has grown at least 29% for 4 consecutive quarters, with net ARR expansion of approximately 121%, as the database layer continues to benefit from increasing AI workloads. Shares rose 8.49% and gained 35.20% over 20 days. The next question is whether AI projects drive higher consumption rather than merely adding trial accounts.

• Snowflake: Shares rose 4.36% and gained 10.40% over 20 days, as institutions raised forecasts on accelerating AI usage. The key test for the data cloud is whether usage translates into faster product-revenue growth while preserving gross margin. If model calls generate only short-term compute consumption, sustainability will require confirmation over additional quarters.

“Software valuations now depend not only on quarterly results, but also on bearish positioning, whether AI generates incremental revenue, the competitive and cooperative dynamics between model providers and application platforms, and whether full-year guidance increases are sufficient relative to quarterly beats.”

• Veeva: Revenue growth accelerated to 17.6%, and full-year guidance was raised by 1.2%. Vault CRM added several large pharmaceutical customers in the second quarter, with 12 of the top 20 pharmaceutical companies now committed to adoption. Falcon is beginning to deploy agents in drug discovery and clinical workflows, while the horizontal CRM platform Aspen remains in early development.

• ServiceNow: Shares rose 10.04% and gained 25.77% over 20 days, placing ServiceNow among the stronger names in the software recovery. Salesforce says its IT service-management offering already has 450 customers and is competing for some ServiceNow accounts. Competition will ultimately center on workflow depth, incremental AI revenue, and renewals—not which company launches more agents.

• Palo Alto Networks: Shares rose 12.83% and gained 17.55% over 20 days, joining CrowdStrike’s strength in a broader cybersecurity rally. Without new operating data, sustainability still depends on platform consolidation, retention, and AI-security revenue. The sharp gain raises the expectations bar for the next earnings report.

• Cloudflare: Shares rose 8.19% and gained 8.77% over 20 days as investors renewed their focus on the edge-cloud and AI-networking thesis. The commercial test is whether AI inference, networking, and security traffic can translate into faster revenue growth. A one-day sector-wide rally cannot substitute for customer expansion and margin improvement.

• Datadog: Shares rose 6.70% but declined 9.54% over 20 days, leaving the near-term recovery insufficient to reverse medium-term weakness. Rising AI workloads support observability demand, but investors must still assess whether log, metrics, and tracing consumption recover in parallel and whether cloud-cost optimization re-emerges.

“The software sector is recovering from extreme pessimism, while partnerships between model providers and enterprise-software platforms have eased concerns about AI-driven disintermediation.”

Consumer Electronics / Smart Vehicles

• Apple: Shares rose 0.36% but fell 5.65% over 20 days, as on-device AI and consumer electronics failed to participate in the sharp rally in GPUs and enterprise software. Higher memory prices and premium-component costs could still squeeze the hardware mix. The next catalysts are product demand, memory configurations, and gross margin—not treating a broader market rebound as evidence of a replacement-cycle recovery.

“Recent memory-spec reductions have been concentrated in PCs, smartphones, and lower-end GPUs/ASICs.”

• Tesla: Shares rose 2.60% and gained 14.88% over 20 days, maintaining relative strength within autonomous driving. With no new product or delivery catalyst that day, the price action reflects investor preference rather than fundamental validation. The next proof points remain commercial autonomous-driving mileage, intervention rates, and robotics progress.

• Avride
  1) Its sidewalk robots have completed more than 600,000 deliveries, with 99% requiring no human assistance. The company plans to double its fleet to 1,000 units this fall and expand into 25 US campuses.
  2) It has completed 100,000 autonomous trips in Dallas through Uber and plans to achieve fully driverless operations in 1H27. Scaling is clear; unit economics and the removal of safety operators are the next tests.

“This is a volume product: walking speed, in-compartment food delivery, Uber Eats, and university campuses.”

• Unity Software: A potential world-model launch raised disintermediation concerns, with investors worried that content-generation and simulation tools could weaken the traditional game-engine gateway, although mid-quarter channel checks remain strong. Key proof points include developer retention, advertising monetization, and whether new tools continue to operate within Unity Software workflows.

• Duolingo: User growth continues to lead, while monetization still lags. Unless AI features in its education app improve paid conversion and bookings, usage growth will be difficult to translate fully into profits. The next metrics to watch are daily active users, CURR, and bookings—not merely the rollout of AI features.

“Edge-computing revenue was $7.2 billion, up 27% year over year, with limited impact on the current profit structure.”

• Qualcomm: Shares rose 0.65% and gained 8.69% over 20 days, showing relative resilience among consumer and edge-AI chips. On-device inference can increase semiconductor content value, but smartphone demand and customer inventories still determine the revenue trajectory. The next indicators are AI-smartphone penetration, average selling prices, and the share of non-handset revenue.

• NXP Semiconductors: Shares rose 1.21% but fell 7.87% over 20 days, as automotive and industrial chips continued to lag the AI data-center supply chain. Rising semiconductor content in smart vehicles must offset vehicle-demand weakness and the inventory cycle. Key metrics are automotive revenue, channel inventory, and power-management orders.

• onsemi: Shares rose 2.16% but fell 10.68% over 20 days; the one-day rebound has not reversed the medium-term weakness. Power semiconductors are sensitive to EV production, silicon content per vehicle, and pricing. If vehicle demand slows, AI data-center power opportunities are unlikely to fill the automotive shortfall immediately.

• STMicroelectronics: Shares rose 3.38% but fell 3.24% over 20 days, outperforming most automotive and analog-chip names. A recovery in automotive MCUs and power devices still requires confirmation through orders and inventory digestion. For now, the move looks more like a sector rebound than evidence of stronger end demand.

• Mobileye: Shares rose 0.82% and gained 8.99% over 20 days, outperforming Aurora and Aptiv among autonomous-driving chip names. Production vehicle programs, system shipments, and commercial autonomous-driving progress are the true revenue anchors. Without new orders, relative share-price strength does not imply higher customer production schedules.

“Many people do not realize this, but Avride is Nebius’s autonomous-driving company.”

• Aurora: Shares fell 1.22% and declined 8.81% over 20 days, as autonomous trucking remained under pressure. The next metrics are driverless operating mileage, fleet utilization, and safety validation. If operations still require costly safety and remote support, scaling may not improve cash flow in parallel.

• Aptiv: Shares fell 1.79% and declined 21.78% over 20 days, making it one of the weakest names in the smart-vehicle group. In-vehicle electronics demand, automaker production schedules, and product mix jointly determine margins. Without new orders or guidance, investors still need fundamental data to confirm an inflection point.

• Symbotic: Shares rose 1.31% but fell 4.19% over 20 days, marking a modest recovery in warehouse robotics. Project-based revenue is sensitive to customer acceptance and deployment schedules. The next metrics are backlog conversion, the number of systems brought online, and gross margin—not a direct extrapolation from AI-robotics enthusiasm to current-period revenue.

• Intuitive Surgical: Shares fell 0.90% but gained 3.99% over 20 days, as surgical robotics diverged from industrial automation and autonomous-driving names. System installations are only the starting point; procedure volumes and consumables revenue determine long-term returns. AI capabilities must translate into clinical efficiency and higher utilization.

• Rockwell Automation: Shares rose 0.29% but fell 8.07% over 20 days, as industrial automation failed to participate in the rebound in AI software and GPUs. Enterprise agents are beginning to control laboratory equipment, but factory-scale deployment must still overcome safety, determinism, and equipment-compatibility hurdles. The next indicators are orders, backlog, and the automation investment cycle.

“Inference has not simply shifted from the GPU market to the ASIC market. It is evolving toward stage-specific allocation of compute and memory, subsequently reintegrated through racks, fabric, and software into heterogeneous systems.”

Jensen Huang’s Curated Portfolio

Memory Index Performance

Memory’s Share of US Equity Trading Turnover

Related Reading

404K SEMI-AI Evening Brief 2026-08-05 — Broadening Constraints: AI Demand Remains Strong, While Gross Margins and Cash Flow Begin to Differentiate Company Performance

404K SEMI-AI Evening Brief, August 27, 2026 — Bottlenecks Spill Over: Nvidia Raises Demand Outlook as Memory, Packaging, and Power All Tighten

404K SEMI-AI Morning Brief, 2026-08-12 — Financing Broadens: $500 Billion Platforms Drive Compute Buildout, While Power and Utilization Determine Returns

Rubin Ultra HBM Capacity Cut to 192GB: Supply and Power Constraints Drive Roadmap Downgrade

404K SEMI-AI Morning Brief, 2026-07-30 — Hardware Under Pressure: Cloud Software Relatively Strong, Memory Long-Term Contracts and Silicon Photonics Expansion Continue

• Table of Contents

• I. Q2 Re-steepens the Growth Trajectory

• II. Data Center Drives Virtually All Incremental Growth

• III. How Custom Silicon and High-Speed Interconnects Flow into Revenue

• IV. Operating Expense Leverage Drives Margin Expansion

• V. Cash Flow Has Yet to Match Profit Acceleration

• VI. SBC, Dilution, and Post-M&A Balance Sheet Structure

• VII. Key Valuation Drivers and Monitoring Framework for Marvell

本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读

Marvell Technology delivered accelerating Q2 revenue alongside strong Q3 guidance, with Data Center expanding to 79% of total sales. While top-line growth is visibly materializing, gross margin trajectory, cash conversion, and equity compensation costs will dictate the true earnings quality of this expansion.

Table of Contents

• [I. Q2 Re-steepens the Growth Trajectory](#i-q2-re-steepens-the-growth-trajectory)

• [II. Data Center Drives Virtually All Incremental Growth](#ii-data-center-drives-virtually-all-incremental-growth)

• [III. How Custom Silicon and High-Speed Interconnects Flow into Revenue](#iii-how-custom-silicon-and-high-speed-interconnects-flow-into-revenue)

• [IV. Margin Expansion Driven Primarily by Operating Leverage](#iv-margin-expansion-driven-primarily-by-operating-leverage)

• [V. Cash Flow Lags Earnings Acceleration for Now](#v-cash-flow-lags-earnings-acceleration-for-now)

• [VI. SBC, Dilution, and the Post-Acquisition Balance Sheet](#vi-sbc-dilution-and-the-post-acquisition-balance-sheet)

• [VII. Valuation Framework: What Metrics Should Price Marvell Next?](#vii-valuation-framework-what-metrics-should-price-marvell-next)

I. Q2 Re-steepens the Growth Trajectory

In Q2 FY2027, Marvell Technology (NASDAQ: MRVL) generated revenue of $2.7393 billion, up 36.5% YoY and 13.3% QoQ, coming in approximately $39 million above the midpoint of prior guidance. While the magnitude of the beat itself was modest, the critical takeaway is that growth is not slowing down even against a rising baseline. Following double-digit sequential growth in Q2, management guided Q3 revenue to a midpoint of $3.15 billion, implying an additional 15.0% QoQ increase. If achieved at the midpoint, revenue will have surged from approximately $2.417 billion in Q1 to $3.15 billion in Q3, adding nearly $733 million in incremental quarterly revenue over just six months.

Q2 non-GAAP operating profit reached $1.0032 billion, up 43.6% YoY and 18.5% QoQ, yielding non-GAAP EPS of $0.94. Operating profit outgrew revenue on both a YoY and QoQ basis, demonstrating clear operating leverage as volume growth diluted operating expense ratios. On a GAAP basis, operating profit was $459.7 million with an operating margin of 16.8%, and EPS stood at $0.33. While both GAAP and non-GAAP metrics improved, their trajectories and underlying drivers diverge substantially, as detailed later in this report.

The Q3 guidance elevates this print from a solid single quarter to an inflection point of accelerating multi-quarter expansion. Revenue is guided to $2.9925B–$3.3075B (midpoint $3.15B), with non-GAAP EPS guided to a midpoint of $1.10, representing approximately 17% sequential growth over Q2. Consequently, market focus will shift from whether Q2 beat expectations to whether Marvell can defend its earnings quality as top-line growth ramps sharply in Q3.

However, one critical set of margin figures warrants close scrutiny. Q2 non-GAAP gross margin came in at 58.9%, flat QoQ but down approximately 0.5 percentage points YoY. For Q3, non-GAAP gross margin guidance is set at 57.5%–58.5% (midpoint 58.0%), down roughly 0.9 percentage points from Q2. Accelerating revenue growth coupled with declining gross margin guidance typically signals shifts in product mix, customer mix, or early-stage ramp costs. While focusing solely on EPS creates an impression of broad-based operational acceleration, incorporating gross margin dynamics calls for a more measured conclusion: revenue visibility has strengthened, but unit profitability remains to be proven through volume production.

Management also noted that AI-related bookings were "exceptionally strong," revenue growth is expected to accelerate over the remainder of the fiscal year, and full-year revenue outlooks for both FY2027 and FY2028 were revised upward. However, the disclosures did not provide specific full-year revenue figures before and after the revisions. As such, while the directional improvement is confirmed, qualitative commentary should not be conflated with undisclosed annual growth rates or specific valuation targets.

II. Data Center Drives Virtually All Incremental Growth

Q2 Data Center revenue reached $2.1715 billion, up 45.7% YoY and 18.5% QoQ, representing 79% of total revenue—up from ~76% in the prior quarter and ~74% in the year-ago period. Notably, the sequential increase in Data Center revenue exceeded the company's total sequential top-line gain, as Communications and Other revenue totaled $567.8 million—up 10.1% YoY but down 3.0% QoQ. In other words, Data Center drove virtually 100% of sequential growth in Q2, while non-data-center segments acted as a minor drag.

An eight-quarter historical view illustrates a structural transformation in Marvell’s revenue mix. Total revenue expanded from approximately $1.516 billion in Q3 FY2025 to $2.739 billion this quarter, with Data Center revenue surging from ~$1.101 billion to $2.172 billion. Over the same timeframe, Communications and Other revenue hovered between $400 million and $600 million. The upward shift in the company’s revenue baseline has been driven almost exclusively by Data Center, lifting its revenue contribution from the low-70% range to nearly 80%.

This structural shift carries two major implications. First, Marvell has transitioned from a diversified multi-market semiconductor supplier into a pure-play vector on hyperscale data center capex. As cloud hyperscalers scale AI clusters, demand across high-speed interconnects, switching, electro-optics, and custom compute magnifies the company's operating leverage and revenue elasticity. Second, revenue concentration is rising. If ramp schedules for major custom programs slip or a single key customer adjusts procurement, non-data-center segments lack the scale to cushion the resulting volatility.

A 79% Data Center revenue share should not be conflated with 79% pure AI exposure. Marvell’s reported Data Center segment encompasses general compute, standard connectivity, and storage workloads beyond AI training clusters. Furthermore, management did not provide a granular breakdown of AI revenue, custom ASIC revenue, customer concentration, or specific product lines. Treating the entire Data Center segment as AI would overstate verifiable AI exposure while obscuring traditional enterprise and cloud data center dynamics.

Current disclosures support the conclusion that AI has become the primary catalyst behind Data Center acceleration, as substantiated by Q2 results and Q3 guidance. However, investors still cannot quantitatively isolate the respective contributions of AI interconnects, custom ASICs, and traditional data center products. The top-line acceleration is real, but revenue attribution remains coarse.

III. How Custom Silicon and High-Speed Interconnects Flow into Revenue

Management characterized the two AI growth vectors as being at distinct developmental stages. Data center connectivity remains robust today, indicating that high-speed optical and copper interconnects already contributed meaningfully to Q2 revenue. In contrast, the custom silicon (ASIC) business is expected to "accelerate significantly" in the second half of FY2027, with the bulk of its volume ramp slated for future quarters. This distinction is critical: connectivity delivers immediate near-term growth, while custom silicon is positioned to extend the duration of the cycle.

The lifecycle of custom silicon—from design win to revenue recognition—involves co-architecture, tape-out, silicon validation, manufacturing ramp preparation, and customer deployment. Strong early bookings or design activity do not immediately translate into concurrent quarterly revenue. While management’s emphasis on robust AI bookings provides a positive demand signal, the guided 15% sequential revenue growth in Q3 provides initial financial validation of booking-to-revenue conversion. Continued monitoring will be required around ramp timing, silicon yields, supply chain assurance, and customer acceptance milestones.

The thesis for high-speed interconnects is more direct. As AI clusters expand in scale, east-west and north-south data traffic between compute engines, racks, and datacenters surges exponentially. Because bandwidth, latency, power consumption, and reach dictate overall system efficiency, interconnects have evolved from auxiliary components into the primary physical bottleneck of computing architectures. Marvell’s revenue opportunity across electro-optics, PAM4 DSPs, switching, and custom interconnect solutions is driven by a dual multiplier: higher port density per accelerator and continuous port-speed upgrades.

However, top-line acceleration and gross margin compression may stem from the exact same dynamic. When large custom programs transition to high-volume manufacturing, segment revenue scales rapidly, but early yield curves, hyperscaler pricing terms, and product mix can dilute corporate average gross margins. A shifting mix between merchant connectivity products and custom silicon can create a divergence between revenue growth and gross margins. The sequential dip in Q3 gross margin guidance likely offers an early indicator of this production mix shift. Without a granular gross margin bridge from management, this remains a working thesis to monitor.

This growth trajectory faces four key constraints: first, heightened customer concentration amplifies the revenue impact of individual project schedules; second, hyperscalers developing internal chip design capabilities require suppliers to continuously defend their performance moat and execution value; third, packaging capacity, advanced node availability, and component supply may constrain ramp velocity; and fourth, export controls and trade policy shifts could alter the serviceable addressable market (SAM). While none of these factors necessarily derails the multi-year secular trend, any one of them could introduce quarterly lumpiness and deviation from a linear growth path.

Management has positioned its October 6, 2026 Investor Day as the next major communication milestone. The most valuable incremental data points will be custom silicon program counts and ramp schedules, connectivity product mix, customer concentration, and a quantitative revenue bridge for FY2027–FY2028. Absent granular disclosure, market consensus ranges for out-year earnings power will remain wide.

IV. Operating Expense Leverage Drives Margin Expansion

In Q2, non-GAAP gross margin was flat sequentially at 58.9%, while non-GAAP operating margin rose to 36.6%. The divergence stems from operating expenses growing more slowly than revenue. Non-GAAP OpEx came in at $610.8 million, up 24.0% YoY and 5.9% QoQ, notably lagging revenue growth of 36.5% YoY and 13.3% QoQ. As scale expanded, R&D; and SG&A; expenses were diluted across a larger revenue base, unlocking clear operating leverage.

This margin improvement carries genuine economic value. For semiconductor companies, substantial R&D; investments in design platforms and customer engagements are incurred upfront, well ahead of revenue recognition. Consequently, as mass-production programs ramp concurrently, incremental revenue does not require a proportional increase in period OpEx. In Q2, non-GAAP operating profit grew 43.6% YoY and 18.5% QoQ—both outpacing top-line growth—demonstrating this operating leverage model in action.

However, operating leverage cannot fully neutralize gross margin risks. While the midpoint of Q3 revenue guidance implies an increase of approximately $411 million QoQ, the midpoint for non-GAAP gross margin dips to 58.0%. On a static calculation, a 90 bps margin contraction offsets a portion of the incremental gross profit from top-line expansion. While Marvell could still deliver EPS growth via continued OpEx ratio reductions, this earnings structure becomes more sensitive to top-line scale and increasingly reliant on Q3 revenue landing near or above the midpoint.

The substantial divergence between GAAP and non-GAAP metrics also warrants scrutiny. In Q2, GAAP gross margin was 53.1% (5.8 percentage points below non-GAAP), GAAP operating margin was 16.8% (19.8 percentage points lower), and GAAP net margin was ~11.2% (~20.4 percentage points lower). These variances primarily reflect stock-based compensation (SBC), intangible asset amortization, and acquisition-related integration costs. While non-GAAP figures highlight underlying operational momentum, GAAP metrics reflect the full accounting burden borne by shareholders.

Consequently, Marvell's Q2 income statement reveals a three-tiered dynamic: top-line revenue is accelerating; adjusted operating income is expanding even faster thanks to economies of scale; and GAAP profitability remains depressed by SBC and M&A-related; charges. As long as top-line growth remains elevated, both profit tiers can improve. However, should revenue growth decelerate, the buffer from OpEx leverage will diminish, and the wide GAAP/non-GAAP gap will re-emerge as a source of valuation discount.

V. Cash Flow Has Yet to Match Profit Acceleration

Q2 operating cash flow was $605.5 million, up 31.2% YoY but down 5.2% QoQ. With capital expenditures at $126.7 million, the analytical proxy for free cash flow (operating cash flow less CapEx) stood at approximately $478.8 million, up 15.6% YoY but down 0.9% QoQ. Note that this FCF proxy is calculated for analytical clarity and is not a management-defined metric in company disclosures.

The cash flow lag relative to profit growth was primarily caused by working capital absorption. In Q2, while non-cash items added approximately $746.2 million back to operating cash flow, working capital changes consumed roughly $448.7 million. Specifically, an increase in accounts receivable resulted in a cash outflow of ~$346.6 million, and prepaid expenses and other assets drained ~$305.2 million, partially offset by other working capital items.

Accounts receivable surged 52.8% YoY, notably outpacing the 36.5% revenue growth. While this divergence likely reflects back-end-loaded quarter-end shipments, customer payment schedules, or payment terms typical of rapid growth phases—and a single quarter does not signify structural collection deterioration—it indicates that cash conversion has not kept pace with earnings growth. If Q3 revenue expands another 15% while receivables continue to outstrip sales significantly, cash flow pressure will transition from a timing anomaly into a structural trend requiring justification.

Verifying cash flow quality hinges on three key metrics: first, whether operating cash flow rebounds relative to non-GAAP net income; second, whether accounts receivable growth realigns with revenue growth; and third, whether the FCF proxy inflects upward as Q3 earnings reach new highs. If these three indicators improve, the current working capital drag can be justified as front-loaded investment for growth. Conversely, if revenue and adjusted profits continue to rise while cash flow stagnates, the market will discount the multiple applied to accounting earnings.

VI. SBC, Dilution, and Post-M&A; Balance Sheet Structure

Q2 stock-based compensation (SBC) totaled $326.2 million, up ~112% YoY and representing 11.9% of quarterly revenue. The diluted weighted-average share count reached 921.2 million, up 5.8% YoY from 870.4 million in the prior-year period. While SBC is excluded from non-GAAP net income, it dilutes future EPS via share count expansion and reduces existing shareholders' proportional claim on company growth.

This is precisely why EPS trajectory cannot be assessed solely through net income totals. Although Q2 non-GAAP EPS was strong at $0.94, per-share growth would have been materially higher had share count remained flat. Going forward, evaluating whether operating leverage genuinely accrues to shareholders requires tracking non-GAAP net income alongside SBC as a percentage of revenue and diluted share growth. The ideal scenario combines rapid revenue expansion with declining SBC intensity and moderating share dilution; if all three expand in tandem, per-share value creation will be partially diluted.

Marvell closed the quarter with $3.9328 billion in cash and cash equivalents against $4.9629 billion in long-term debt, yielding a net debt position of approximately $1.0301 billion. While liquidity appears ample on the surface, the elevated cash balance was bolstered by financing activities, including $2.0 billion in convertible preferred stock. As such, the cash increase cannot be attributed entirely to Q2 operating free cash flow.

Goodwill stood at $13.8739 billion and intangible assets at $2.3466 billion, totaling $16.2205 billion and accounting for approximately 58.9% of total assets. This asset profile reflects Marvell’s historical M&A; strategy to expand its technology and product portfolio. However, it also means GAAP earnings will continue to bear long-term amortization drag and carry impairment risks should acquired assets underperform. Together with SBC, acquisition-related integration costs represent the primary driver of the persistent GAAP vs. non-GAAP earnings spread.

Days sales of inventory (DSI) stood at ~96 days with an inventory turnover of ~3.17x, showing no severe imbalance for now; however, inventory must be monitored alongside accounts receivable during this steep revenue ramp. While building inventory ahead of demand and accumulating receivables post-shipment are standard dynamics during rapid expansion, the crucial question is whether these working capital assets convert into revenue and cash as scheduled. Should customer programs face delays, working capital drag and balance-sheet risks will compound simultaneously.

VII. Key Valuation Drivers and Monitoring Framework for Marvell

While current disclosures support upward earnings revisions, they remain insufficient to justify a rigid point price target. Without disclosures on prevailing share price, consensus estimates, segment long-term margins, or quantified custom silicon pipeline revenue, applying an arbitrary P/E multiple would project false precision. A more rigorous approach is to identify which fundamental validation trajectory Marvell follows, thereby determining whether its growth justifies multiple expansion, maintenance, or compression.

The revenue validation track is the most straightforward. The Q3 revenue guidance midpoint of $3.150 billion serves as the baseline benchmark for the upcoming quarter. The high end of the range ($3.3075 billion) would signal that AI connectivity and custom silicon ramps are exceeding plan, while the low end ($2.9925 billion) would indicate ongoing growth but slower project execution. Monitoring whether Data Center revenue holds at or above ~79% of total sales will confirm whether top-line expansion remains decisively driven by AI infrastructure.

The gross margin track determines the quality of top-line expansion. Q3 non-GAAP gross margin guidance is 57.5% to 58.5%. If revenue tracks toward the high end while gross margin approaches 58.5%, it will demonstrate that scale is ramping without compromising unit economics. If revenue increases while gross margin slides toward 57.5%, the market will scrutinize custom ASIC dilution, customer pricing power, and product mix headwinds. If both revenue and gross margin miss guidance ranges, the growth narrative will face material downward revisions.

The cash conversion track distinguishes temporary working capital buildup during expansion from chronic earnings quality lag. Q3 operating cash flow, accounts receivable growth, and the FCF proxy must be evaluated in tandem. If revenue hits record highs while cash conversion improves, Q2's cash drag will be confirmed as a timing issue. However, if revenue hits record highs while receivables continue to outpace sales, valuation multiples should reflect a discount for collection and working capital uncertainty.

The equity dilution track determines how effectively top-line growth translates into per-share value. SBC has reached 11.9% of revenue, and diluted share count is up 5.8% YoY, alongside a capital structure carrying substantial goodwill, intangibles, debt, and convertible preferred shares. While continued non-GAAP EPS growth is essential, moderating SBC intensity and slowing share dilution are equally critical; otherwise, operating leverage will be partially absorbed by dilution and financing costs.

This framework yields three distinct scenarios. In the Bull Case, Q3 revenue lands near or above $3.3075 billion, gross margin approaches 58.5%, cash conversion rebounds, and the upcoming Investor Day provides quantified evidence for custom silicon ramps and FY2028 growth, prompting dual upgrades in earnings estimates and valuation multiples. In the Base Case, revenue hovers near $3.150 billion, gross margin remains around 58.0%, and cash flow steadily catches up, with share price returns primarily driven by earnings growth. In the Bear Case, revenue tracks toward the low end of guidance, gross margin falls below 57.5%, or receivables consistently outgrow revenue, leading the market to reprice risks around customer concentration, program timing, and earnings quality.

The most significant takeaway from Q2 is that Marvell’s reported revenue and forward guidance confirm AI demand has transitioned from bookings and design wins into scaled financial monetization. With Data Center revenue reaching 79% of total sales, Q3 guidance midpoint projecting another 15% QoQ revenue growth, and custom silicon scheduled to accelerate in 2H, these catalysts collectively raise the company's growth ceiling over the next two years.

Yet the same quarterly report clearly delineates operational boundaries: gross margins are not expanding alongside revenue and guidance implies a sequential dip at the midpoint; operating cash flow fell QoQ amid substantial working capital absorption; and SBC and share count continue to climb, alongside a balance sheet heavily weighted toward goodwill and intangibles. Marvell’s investment narrative has pivoted from "whether demand exists" to "whether growth converts into high-quality earnings." Going forward, every incremental revenue beat will require corroboration from gross margin resilience, cash flow conversion, and per-share value creation.

Related Reading

404K SEMI-AI Evening Brief 2026-08-03 — Bottlenecks Spill Over: Cloud Acceleration Pushes Incremental Demand Toward Memory, Packaging, and Optical Interconnects

404K Tech Morning Brief — August 18, 2026 — Repricing the Bottlenecks: Memory Strengthens as Compute Expansion Shifts to Power, Packaging, and Optical Interconnects

404K SEMI-AI Evening Brief, 2026-08-24 — The Bottleneck Shifts: HBM, Packaging, and Optical Interconnects Become the New Tests of AI Expansion

Arm FY2027 Q1 Deep Dive: Data Center Royalties More Than Double, AGI CPU Demand Exceeds $2 Billion

404K SEMI-AI Evening Brief, 2026-08-10 — Efficiency Takes the Baton: AI Expansion Shifts Toward Coordination Across Memory, Power, and Interconnects