Sales partner growth is generally measured by Consumption, which is highly correlated with Revenue, because SNOW primarily evaluates partners on Consumption. A few contacts use consulting revenue where Consumption data are unavailable.

U.S. Partner A

1. Consumption Performance

   1. 1Q25 growth accelerated and exceeded expectations

   2. 2Q25 growth accelerated and significantly exceeded expectations

   3. 3Q25 growth slowed and fell short of expectations

   4. 4Q25 growth accelerated and exceeded expectations

   5. 1Q26 growth was flat versus 4Q25

   6. 2Q26 exceeded expectations

U.S. Partner B

1. Consumption Progress

   1. 1Q25 accelerated versus 4Q24 and exceeded expectations by 15-20%

   2. 2Q25 accelerated versus 1Q25 and exceeded expectations

   3. 3Q25 slowed versus 2Q25 and fell short of expectations

   4. 4Q25 accelerated versus 3Q25 and exceeded expectations

   5. 1Q26 accelerated modestly versus 4Q25 and exceeded expectations by 8%

   6. 2Q26 slowed modestly versus 1Q26, in line with expectations.

2. 2Q26 Trends

   1. Customers are increasingly using Snowflake to consolidate data, and Snowflake is taking share from platforms such as Google BigQuery, AWS Redshift, and Microsoft Fabric.

   2. However, we also note that CSPs such as AWS are reducing sales resources dedicated to Snowflake and allocating more sales resources to Databricks and Anthropic, which could affect Snowflake sales going forward.

U.S. Partner C

1. Consumption Progress

   1. 3Q25 grew 24% YoY, exceeding the 22% expectation

   2. 4Q25 grew 28% YoY, at the high end of the expected 26-28% range

   3. 1Q26 grew 28-29% YoY, exceeding the expected 24-26%

   4. 2Q26 grew 32% YoY, exceeding the expected 27-28%, making it the strongest quarter over the past several quarters

   5. 3Q26 and 4Q26 YoY growth expectations have been raised from 30% to 34% to 35%.

2. Reasons for 2Q26 Outperformance

   Subscribe now

Read more

Optics are moving inside the tray, and the memory tier it targets is the largest bandwidth tier in the system, roughly ten times the scale-up bandwidth per GPU and about one hundred times the scale-out. The capital is already moving, with $2 billion each from NVIDIA into Lumentum and Coherent, $3.25 billion from Marvell for Celestial AI, and a revenue ceiling of up to $120 billion on the Google and Marvell warrant. Scale-in, in this note, means die-to-die and chip-to-memory optical links at millimeters to about a meter, which Lumentum’s CEO on August 27 called “the thing to watch over the next 12 to 18 months”. The word is spreading with split meanings since NVIDIA uses it for something else, and Section 3 sorts out who means what. The kicker is Section 7, where our checks point to the largest scale-in exploration running at Google, built around memory pooling, and the warrant’s disclosed scope, “memory interface controllers and near-memory compute”, is what we believe to be its public paper trail. The work is early, and first-generation products start well below the ceiling.

1. Scale-In’s Coming-Out Week

At an investment conference on August 27, Lumentum CEO Michael Hurlston said, “I think the thing that is going to be significant is the scale-in, and that’s the thing to watch over the next 12 to 18 months... Are you going to see it actually go in-tray and serve this high bandwidth connectivity between memory and between GPUs?” He was describing optical links inside trays. Scale-in was also featured in imec’s slides on the first day of Semicon Taiwan, in a talk on 3D integrated optics for chip-to-chip interconnect.

Also on August 27, Marvell said on its FY27Q2 earnings call that its fiscal 2028 revenue outlook for scale-up optics “has increased meaningfully compared to prior expectations”, against the roughly $300 million category framing it gave a quarter earlier, and that it is on pace to make about $1 billion of capacity prepayments to suppliers in fiscal 2027. The parts are the same since scale-up optics uses the lasers, optical engines, packaging, and test flows that scale-in will need, just at longer reach. Scale-up optics ships starting in 2027 while scale-in follows two to three years later, if not more, because those same parts still have to clear tighter thermal, reliability, test, and other requirements within the tray.

On August 19, a week before the earnings call, Google and Marvell announced a partnership that, in the filing’s words, “spans a comprehensive range of custom silicon programs that attach to the TPU ecosystem”. Marvell issued Google a warrant on up to 59 million shares. Most of it vests in 240 equal installments, one for each $500 million of qualifying revenue through fiscal 2033, which is where the $120 billion figure in the press comes from. None of that revenue is committed. The disclosed scope covers AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute. We believe the last two items refer to the pooled-memory tier behind scale-in, and Section 7 explains why.

At Hot Chips 2026, NVIDIA used Scale-In for a DPU services network, and Meta used it for chiplet-level data movement inside its MTIA 400 accelerator. Scale-in now has at least three different technical meanings.

2. The Definition Adopted in This Note: Optical Links Inside the Tray, From Millimeters to About One Meter

In this note, scale-in refers to optical links inside the tray that serve high-bandwidth connectivity between memory and between GPUs. This includes die-to-die connections between chiplets and compute die, and chip-to-memory connections such as XPU to HBM, over reaches from millimeters to about one meter. Links from the XPU to pooled memory belong to the adjacent pooled-memory fabric, which this note treats separately. The term also appears as scaling-in and, at Marvell, as scale-inside.

AI systems use different interconnect media at different physical layers. Within the package, CoWoS, interposers, and RDL dominate. On the board, systems rely on PCB traces and retimers. Within the rack, copper DAC and AEC cables handle scale-up traffic. Within the data center, pluggable optics dominate, and between data centers, coherent optics dominate. As bandwidth demand per XPU rises, optics enters one layer at a time, and the electrical-to-optical boundary moves progressively closer to the chip.

This also explains why bandwidth requirements rise at each inward layer. Scale-out is measured by network-port bandwidth, scale-up by fabric bandwidth, and scale-in approaches memory bandwidth. The closer the link is to the chip, the higher the aggregate bandwidth per XPU and the more demanding the power and density requirements.

Comparing those tiers requires conversion: a byte is eight bits, so 1 TB/s equals 8 Tb/s, and a link can be quoted one way or both ways. Lumentum’s ECOC 2025 material puts per-GPU scale-up bandwidth at 7.2, 14.4, and 28.8 Tbps for Blackwell, Rubin, and Feynman, counted one way, while NVIDIA quotes NVLink 6 at 3.6 TB/s, counted both ways. The two agree once directions are matched, since Lumentum’s 14.4 Tbps is per direction while NVIDIA’s 3.6 TB/s is the both-ways total of the same link. This note quotes bandwidth one way throughout, and memory-bus figures are aggregates.

The optical boundary moves inward from scale-out to scale-in. Source: FUNDA.

3. Everyone Means the Tray, Except NVIDIA

The optics and memory camp uses scale-in for the tier inside the tray. Marvell’s solutions page positions its portfolio “across scale-in, scale-up, scale-out and scale-across architectures”, and its blogs call the inner tier scale-inside, meaning die-to-die interconnects that move data between compute and memory dies inside XPUs, extending to chip-to-chip links within a tray. Avicena defines it as “AI scale-in (die-to-die and die-to-memory connectivity)” with reach from a few millimeters to one meter. Keysight bounds it at “the silicon and package domain including pre-silicon design, chiplets, memory and die-to-die/package interconnects”. Eliyan frames its market as “AI scale-up and scale-in networks”. LightXcelerate, a Palo Alto startup, states that its optical chiplet supports scale-in, scale-up, and scale-out across UCIe, UALink, PCIe, CXL, and NVLink.

Executives use the word the same way. Coherent CTO Julie Eng, on a Celesta Capital TechSurge panel with Eliyan founder Ramin Farjadrad and Lumentum cloud CTO Matt Sysak, put scale-across and scale-out in deployment now, large-scale scale-up deployment next year, and scale-in after that, describing it as solving “inter-die interconnects with light” when copper runs out in the die-to-die space. She also noted renewed interest in NRZ signaling for short links, “especially scale-up and die-to-die scale-in”.

Eliyan founder Ramin Farjadrad, Coherent CTO Julie Eng, and Lumentum cloud CTO Matt Sysak on the deployment sequence for optics. Source: Celesta Capital TechSurge, Aug 2026.

The research world has picked the word up as well. On the opening day of Semicon Taiwan, imec presented a slide at its ITF Taiwan forum titled 3D integrated optics for chip-to-chip scale-in interconnect, showing wide-and-slow area I/O with low-speed SerDes and a step from pluggable optics to co-packaged optics to an optics layer stacked directly under the XPU and HBM. On imec’s chart, the 3D step pulls away from co-packaged optics by roughly two orders of magnitude in bandwidth density per unit of energy (Gbps/mm per pJ/bit) by 2035. That came just four days after the Lumentum and Marvell statements.

imec at ITF Taiwan, Semicon Taiwan: 3D integrated optics for chip-to-chip scale-in interconnect. Source: imec. Photo by Jukan (@jukan05 on X), August 31, 2026.

At Hot Chips 2026, NVIDIA used the same word for a different layer. Its five-network AI factory model pairs Scale-In with the BlueField-4 DPU, a per-node infrastructure network for agentic CPU capacity, storage, security, and orchestration, quantified at 800 Gb/s per Vera Rubin compute tray, plus 4x1.6 Tb/s of scale-out. NVIDIA also has the pooled-memory tier this note discusses, called Context Memory (CMX), built on BlueField-4, and does not call it scale-in. Meta’s MTIA 400 slide deck uses “Scale-in DMA & streaming reductions” for data movement between chiplets inside the package, over electrical links. Microsoft’s Azure documentation has long used scale-in for removing virtual machine instances in autoscaling. The table below explains each term.

Meta MTIA 400: scale-in DMA and streaming reductions inside a multi-chiplet design. Source: Meta, Hot Chips 2026.

Scale-in as defined in this note, other uses of the term and the adjacent pooled tier. Source: FUNDA, redrawn from public sources. MPU means memory processing unit, covered in Section 7.

NVIDIA sells the pooled tier this note describes, under the CMX name. It has invested $2 billion each in Lumentum and Coherent with purchase commitments, and licensed Groq’s technology for SRAM-first inference. Its current scale-up racks run on copper.

The five networking infrastructures of the NVIDIA AI factory. Scale-In paired with BlueField-4. Source: NVIDIA, Hot Chips 2026.

4. Sizing It: The Memory Tier Is Roughly 10x Scale-Up

Subscribe now

Read more

This was evidenced by crowded short positioning on CRM into earnings, and it was exactly why we turned neutral from conservative in preview. We didn’t foresee the announcement of Claudeforce, which was a major bullish narrative turn in our view. Previously, enterprise SaaS like CRM has been trying hard to defend its UI value, which would likely be commoditized in the agent era. The shift started a few months ago, when CRM announced Headless 360, allowing LLMs to access its underlying data more directly and effectively giving up the UI layer. Claudeforce takes it a step further, as CRM confirms its strategy shift to defend the data and workflow layer without defending the UI. More importantly, the involvement of Anthropic demonstrates that LLMs have very low intention to fully replace SoRs, at least for now, while actively looking for a more realistic way of partnership. Therefore, a re-rating of enterprise SoRs makes sense here, while SMB-focused players remain more vulnerable.

On semi, NVIDIA did something this quarter it has never done: it put a number on the following year. FY2028 revenue is guided to roughly +70% vs. sell-side expectations of +45-50%; customer orders point to demand doubling; and the guide is still held back by supply rather than customer demand. A company only underwrites a year out when it can see the order book and knows the ceiling sits somewhere other than demand. Our AIDC deep dive this week is the same story told from the other end of the stack. North American demand in 2027 runs to roughly 35GW on an IT basis against 16.5-23.4GW that can actually be delivered, a gap of about 2x, and the binding constraint is power delivery rather than capital or construction. The 2027 grid path is already closed: anything not in the interconnection queue by mid-2025 does not energize next year, large gas turbine slots are sold out through 2031 on a four-to-five year order-to-COD cycle, and permitting killed 78% of the 43GW of stranded projects. Modular construction compresses the build and still cuts COD slip probability by only 10-20%, because what it cannot touch sits upstream of the build. Scarcity is already in the price: ARR per GW spans $8.3bn to $50bn across the four players we compare, and NBIS has signed at 3-4x its own installed-base unit price within 18 months.

This Week’s Reports

AI data centers — 2027 demand runs at roughly 2x what can actually be delivered, and power is the gate. North American demand of about 35GW IT meets 16.5-23.4GW of deliverable supply next year, with the grid path closed to anything not queued by mid-2025 and large turbine slots sold out through 2031.

Enterprise AI — Volume 1 of our adoption survey. Spend is still growing across the thirteen samples, seats are close to saturated at the companies already committed, and the increment is moving into API calls and production workflows.

Premium Report Snapshot

A portion of our research is reserved for Premium subscribers and is not distributed via Substack. Below is a snapshot of what Premium subscribers received this week beyond the Substack feed.

Subscribe now

Read more

Read more

North American AI data center demand in 2027 runs to roughly 35GW on an IT basis, against 16.5-23.4GW that can actually be delivered. Where our two independent supply methods overlap, that gap is 1.8-2.1x, and it is the number this whole report is built around.

Demand cross-checks two ways and lands in the same place. A CoWoS-capacity build gives 44-49GW of global chip-level power in 2027; a platform-by-platform BOM build gives 48.4GW. The market debates 40-60GW. We assume roughly 50GW globally, with North America taking about two-thirds.

Supply is already locked, and capital cannot move it. The grid path delivers 11-14GW IT, set entirely by who entered the interconnection queue before mid-2025; an application filed today does not make 2027. Behind-the-meter equipment adds 5.5-9.4GW IT net of overlap, but large gas turbine slots are sold out through 2031 on a four-to-five-year order-to-COD cycle, so a frame ordered now becomes supply in 2030-32.

Power delivery is the binding constraint, and it sits upstream of everything a sponsor can pay to accelerate. Permitting killed 78% of the 43GW of stranded projects. Power-side causes explain 60-70% of MW-weighted slippage over the past 12 months. Below that sits licensed labor: a journeyman electrician takes 10,000 hours to train, and only about 30% of MEP workers are in the regions holding 70% of projects. Modular construction compresses the build but not the queue, cutting COD slip probability by only 10-20%.

Scarcity is already visible in price. ARR per GW spans $8.3b to $50b across the four players we compare, and NBIS has signed at 3-4x its installed-base unit price within 18 months. The violent repricing of Q1 and Q2 is over; from 26Q4 we expect stabilization with modest increases, and rents down roughly 20-30% by generation as supply lands, with utilization staying high.

Four routes to power, four different risks. xAI buys speed with 2x power cost and permitting risk, on the shortest contracts of the four; CRWV converts power delivery risk into refinancing risk; NBIS runs on customer prepayments and is the clearest evidence of the price move; IREN sits on legacy energized capacity with the best cost of capital of the four, but its incremental capacity queues like everyone else’s.

I. Where Things Stand Today

The whole US market fits into a single funnel (Aterio project database, July 2026 edition):

Subscribe now

Read more

The headline from those thirteen: spend is still growing, and the interesting part is where the increment lands and where the trajectories start to split. Seats are close to saturated at the companies already committed, so the growth is moving into API calls and production workflows. One large global bank has gone from roughly $150-160mn a year to $200-300mn. A large European telecom operator is running $30-50k a month, up from $15k at the start of the year, and expects just under $100k within six months. A mid-to-large biopharma has moved from roughly 80/20 subscription-to-API at the start of the year to 70/30. Against that, a large European automaker is up only 10-15% year to date, guides to broadly flat in 2027, and a large global pharma expects to approach a plateau over the next 12-18 months. Our read is that the seat line is maturing on schedule and the API line is the one that compounds, which makes production workflow deployment the focus, not headcount. This number decides whether model vendor ARR reaccelerates.

The other thread this week was optics, and it keeps getting better. Supply-chain chatter has Google taking the training-side TPU 9t from a 3D torus to a 6D torus for 2027-2028 deployment, which sounds incremental but isn't: wrap-around optical ports per chip go from 1.5 to 6, because at edge length 2 every chip becomes a boundary node. That is a 4x lift in OCS ports and optical module content per TPU before the 800G-to-2.4 T Coherent Lite upgrade is layered on top. Our OCP APAC takeaways point the same way from the other end of the stack, with copper running out of road at 400 Gbps per lane and the whole power path being redesigned from grid to chip. Next week is NVDA, where the only question that matters is whether management puts more shape on the cumulative $1T Blackwell/Rubin revenue path.

This Week’s Reports

Optics — a 6D torus would quadruple optical content per TPU, and the derivation is simple enough to check. Wrap-around optical ports per chip rise from 1.5 to 6 because edge length falls from four to two and every chip becomes a boundary node, which fills in the 2:1 to 10:1 scale-up range Lumentum showed at OFC without spelling out the architecture behind it.

OCP APAC — copper keeps losing ground on reach, and the power path is being redesigned end to end. Nvidia and Broadcom both claim CPO in production, while ASE says the ecosystem isn't ready; on the power side, the industry is converging on 800VDC with solid-state transformers as the endgame, and vertical power delivery is reshaping MLCC, SPS, and inductor demand.

Premium Report Snapshot

Below is a snapshot of what Premium subscribers received this week beyond the Substack feed.

Subscribe now

Read more

Supply-chain rumors suggest Google’s training-oriented TPU 9t might raise the torus dimensionality to 6D while retaining the torus architecture, with deployment targeted for 2027–2028. If realized, wrap-around optical ports per TPU on the scale-up side would rise from 1.5 (3D) to 6, a fourfold increase in OCS ports and optical module content per TPU; combined with the 800G→2.4T Coherent Lite speed and modulation upgrade, the value uplift would exceed 4x. This direction aligns with the two evolution paths we laid out in our earlier report.

I. The Two Evolution Paths from Our Earlier Report

In Deep|AI Infra 2026, we argued that, in the agent era, the system bottleneck is shifting from single-chip compute to system-wide communication efficiency. MoE shifts the communication pattern from All-Reduce toward bisection-bandwidth-hungry All-to-All. Per the GShard paper, as experts scale from 128 to 2,048, All-to-All time for MoE dispatch/combine rises from 16% to 36% of MoE + Transformer layer execution time. On that basis, we expected Google to push along two lines:

1. Scale-out expansion: the number of in-rack TH6 switches rises from one to four–six per rack, DCN bandwidth up 4–6x, the scale-out OCS port ratio moves from 0.2:1 to 1:1, and the optical module ratio increases from 1:1–1:2 to 1:3–1:4;

2. Scale-up dimensionality: a 3D torus’s bisection bandwidth grows only as N^(2/3); assuming the torus architecture is retained, adding dimensions is the main practical route to higher bisection bandwidth, with the OCS Port-to-xPU ratio rising from 1.5:1 toward 2:1–10:1.

II. TPU 8t/8i and Virgo: How the Thesis Is Playing Out, Layer by Layer

Subscribe now

Read more

A clear theme at the conference was scale-up networking, and whether it stays on copper or moves to optics once a domain grows past one rack. Like OFC 2026 and OCP EMEA before this one, the industry keeps landing on the same answer: copper where you can so far, but optics remains the structural future direction.

The second theme was co-packaged optics, where views on the adoption timeline diverged. Nvidia and Broadcom both assert that they have CPO in production. ASE, one of the OSATs that assemble and test CPO packages, said the ecosystem is not ready.

On the non-optics side, our takeaway is that the industry is redesigning the entire power path end to end, from grid to chip. At the facility level, data center build-out is shifting from component stacking toward vertically integrated, modular designs that scale to GW-class deployments. In power architecture, the industry is converging on 800VDC as the next standard, with solid-state transformers emerging as the endgame for cutting conversion loss as rack power heads past the MW level. At the package level, power delivery is moving underneath the chip via vertical power delivery, reshaping component demand and driving higher-capacitance MLCC, smaller form-factor SPS, and a materials transition in inductors along the way.

We expand on these points below, then turn to our power takeaways, and close with the more technical optics sections.

Subscribe now

Read more

Memory is where we most disagree with the consensus read. The flattening in ASP growth is being interpreted as the top of the cycle. We think it is normalization, and the mechanism is not hard to see: consumer devices cannot absorb further price increases, and CSPs are signing LTAs at prices below spot, so blended ASP growth decelerates by construction while vendors gain upfront payments and penalty protection. Kioxia told us at FMS it can fill only 40-50% of demand and expects the rest to roll into next year, which is what Micron has been saying. SanDisk’s investor day then put a floor under the argument with a mechanism that holds non-GAAP gross margin near 80% even at minimum contract prices. Higher for longer, not peak and collapse.

Last, on Google. The market read Jeff Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals leaving as a brain drain. We read it closer to the opposite. Most of those names had already drifted away from the frontier-LLM execution line, and Google’s problem after Gemini 3 was never talent or compute; it was too many senior leaders and two centers of gravity between London and Mountain View. We believe concentrating authority under Koray is the fix, not the damage.

One piece of housekeeping, and it is the thing we would most encourage you to sign up for. We are running FUNDA Product Day this coming Monday, 17 August at 10:00 AM ET. Investors have been asking to see the new Plays in detail, so this session is a live walkthrough of four of them: Trending Topics, which clusters social-media discourse across AI, space and adjacent themes into influence-weighted narrative topics with tickers attached; the Top and Bottom Indicator, a three-layer framework for market turning points combining a daily short-term signal, a weekly contrarian composite across credit spreads, positioning, flows, breadth, volatility and sentiment, and a daily overheat/washout gauge on semis; AI Labs Monetization, our monthly estimate of Anthropic and OpenAI annualized revenue with the range and implied growth around it; and the SpaceX Revenue Forecast, built bottom-up across Space Services and Connectivity off disclosed actuals and live pricing and subscriber inputs. We will show how each one is built and what data sits underneath it.

Register for FUNDA Product Day →

We published thirteen reports across our channels this week; the rundown is below.

This Week’s Reports

Subscribe now

Read more

Source: SanDisk Investor Day 2026

We currently model a mild pricing pullback in CY2028 and a mid-to-high-teens bit shipment CAGR for FY2028–FY2030, in line with the company’s long-term bit supply growth target. Under this base case, we project CY2027/28/29 EPS of $292/$334/$388, implying forward P/E multiples of 5.2x/4.6x/3.9x at the current share price.

Subscribe now

Read more

FUNDA Product Day is our recurring live product briefing for clients covering new FUNDA capabilities, research workflows, and platform updates.

This session will feature four new research Plays launched over the past few weeks, with live demonstrations of how they work, the research frameworks behind them, and the data powering each workflow.

Agenda

1. Trending Topics

Aggregates social-media discourse across key market-moving sectors such as AI, space, and adjacent themes, then clusters it into distinct narrative topics. Each topic carries a directional label, reach and engagement figures, contributor counts and affected tickers. Contributions are influence-weighted, with recognized KOLs assigned greater weight to improve signal quality.

Read more

Nebius delivered an overall positive Q2.

The biggest pre-earnings risk—whether the Vineland 2 delay would force a cut to FY26 revenue guidance—did not materialize. At the same time, GPU demand, pricing, and Token Factory all came in stronger than expected.

Q2 group revenue reached $582Mn, up 46% QoQ, while annualized run-rate revenue increased from $1.9Bn in Q1 to $3.0Bn. Adjusted EBITDA margin improved from 32% to 41%. The company maintained its FY26 guidance of $3.0–3.4Bn in revenue, $7–9Bn in ARR, and an adjusted EBITDA margin of approximately 40%.

Relative to Q1, we see three key incremental developments:

• Demand and pricing were stronger than expected. Management said it could sell out its entire 2027 capacity today if it chose to do so, while the company has begun using short-duration contracts and capacity auctions to drive higher monetization per MW.

• Token Factory has moved from product-market fit to P&L contribution. The CFO explicitly cited Token Factory as a contributor to both Q2 revenue growth and margin expansion.

• The Vineland 2 risk has shifted from a potential guidance cut to revenue timing. The company maintained its YE26 connected-power target of 800MW–1GW. Still, it clarified that it takes a few months for connected power to translate into revenue generation, implying that a meaningful portion of this capacity will only become active progressively through 1H27.

As a result, we still see a reasonable possibility that FY26 revenue ultimately comes in below the $3.2Bn midpoint of guidance, although an outright guidance cut is no longer our base case. More importantly, following Q2, we see greater upside to 2027 revenue/MW and margin potential than to FY26 estimates.

Demand Remains Strong; Pricing Is Better Than Expected

Subscribe now

Read more

“Expanding inferencing and training applications are driving full-rate connectivity between data centers, while political and regulatory constraints favor smaller, more modular builds. These 2 factors, among others, are substantially increasing the demand for our pump laser solutions.” –LITE, FY26Q4 earnings call, August 11, 2026.

Scale-Across Moved Into Design Wins and Order Books

A year ago, scale-across described an architecture concept. AI clusters that can no longer grow inside one building must link data centers across campuses and metros, because power, land, and compute are constrained at every individual site. On the latest round of earnings calls, the same term appeared attached to design wins, order intake, and TAM commentary.

Exhibit 1: The three connectivity domains of an AI cluster. Source: FUNDA.

Subscribe now

Read more

A clear theme at the conference was the need to reduce dependence on DRAM and HBM, and to look to solutions that address every part of the memory and storage hierarchy (now often referred to by Nvidia’s G1 to G4 labels).

This was echoed by Micron’s comment at the KeyBanc forum yesterday that the number one constraint faced by their customers is DRAM.

A version of the memory and storage hierarchy with CXL pooling added at G2.5 and Nvidia CMX at G3.5. Source: FUNDA

Feedback on DRAM and NAND markets

Subscribe now

Read more

On the surface, this is clearly a significant exodus of senior talent. Jeff Dean, Sanjay Ghemawat, Quoc Le, Oriol Vinyals, and several others among the most senior and influential technical leaders in Google’s history are leaving. Demis Hassabis is also moving from CEO of Google DeepMind to Chair and Alphabet Chief Scientist, with Koray Kavukcuoglu taking on more of the day-to-day operating responsibility. Understandably, the market sees this as a major brain drain.

But we think the actual impact may be considerably smaller than the headlines suggest - and over the medium term, the changes could even prove constructive.

One important reason is that, despite their enormous historical contributions, many of the senior people now leaving had already become increasingly peripheral to the core Gemini / frontier-LLM effort over the past year or two. They remain world-class researchers, but there is an important distinction between the people who were historically most important to Google AI and the people actually driving frontier-model execution today.

In fact, we think Google’s bigger problem may have started after the success of Gemini 3.

Gemini 3 established itself as arguably the leading multimodal model, while Google already had one of the strongest proprietary compute infrastructures in the industry. At the same time, DeepMind’s work, including AlphaFold, had achieved Nobel-level scientific recognition. It would have been easy for the organization to conclude that the frontier-model race had largely been caught up - or even temporarily won - and that the highest-value use of its most exceptional talent was to go after the next generation of major scientific breakthroughs.

That helps explain why Demis and several other senior researchers increasingly shifted their attention away from the more engineering-heavy work of frontier-model scaling and toward AI for Science, drug discovery,y and other longer-horizon, potentially Nobel-level problems. For researchers of that caliber, the attraction is completely understandable.

The Strategic Misread: RSI Is the Path, AI for Science Sits on Top

Subscribe now

Read more

Software was the axis. PLTR grew 93% YoY with US Commercial up 149% to $764mn, which is a direct rebuttal to the pre-earnings sell-side channel checks arguing US Commercial was rolling over; management then raised FY26 by $496mn, and only a quarter of that is the beat flowing through. NET accelerated to 36% and, more importantly, raised the full year by more than the beat — about $26mn of genuine second-half revision, with DBNRR back to 120%. DDOG did nothing wrong: 35.6% growth, core non-AI accelerating to high-20s, a renewed nine-figure OpenAI contract. It fell anyway, because expectations were full and positioning was crowded into the print, exactly the risk-reward our preview flagged. Fundamentals and setup are two different trades. The other pair worth sitting with is AppLovin and Unity: APP’s gaming revenue grew just 2% QoQ ex-commerce, the slowest in years, while Unity Ads put up 23% QoQ on the D28 rollout. We do not think that is a coincidence.

The piece we would most want read this week is our LLM deep dive. July took more than $1tn of AI-linked market value off the table on an argument about applications — whether coding is the last killer app. We think that is the wrong lens. The variable is recursive self-improvement, and the early signals are already in public data: roughly one month from the end of GPT-5.5 pre-training to release, an up-to-80% price cut three weeks after GPT-5.6 shipped, and AI’s measured contribution to internal R&D velocity going from 5% to 15-20% in six months. If RSI crosses its threshold, distillation-based catch-up gets harder, not cheaper, and compute becomes the least compressible input in the stack. The reference TAM stops being a software budget in the hundreds of billions and starts being a labor market in the tens of trillions.

Away from software, the physical bottlenecks kept printing. AXTI beat consensus by nearly 40% on revenue with record InP, and the COHR six-inch LTA plus the $87mn LITE deposit agreement turned what had been our qualitative call into signed contracts. ATI posted its strongest quarter since 2007 and raised the year for the second time, with AA&S margin up 930bps — scarce nickel and titanium capacity converting straight into price. And SPCX got its first print out of the way; from here the stock is a function of Flight 14 catching Starship on the tower in late August and whether the AI compute number revises from 5GW toward 10GW. We published sixteen reports across our channels this week; the rundown is below.

This Week’s Reports

LLM — the market is arguing about the wrong variable. The ladder that matters runs coding → continual learning → recursive self-improvement, and the early RSI signals are already visible in release cadence, pricing, and internal R&D velocity. If RSI crosses its threshold, compute is the deepest and least compressible moat in the stack.

Optics — the FCC rule is coming, and the definition of a “Chinese transceiver” is the whole trade. Whether origin follows CBP substantial-transformation standards or pierces ownership and control decides who keeps market access; two CBP precedents put the line at where the main PCBA is done.

APP & U 2Q26 — D28 IAA is now the core battleground. AppLovin grew 53% YoY but landed 1% below the high end of guidance, and ex-commerce gaming revenue rose only 2% QoQ — the slowest sequential rate in several years — while Unity Ads accelerated 23% QoQ as Day 28 campaign spend nearly tripled and adoption passed 25% of advertisers. Management calls it model-cadence timing; we read the pattern as a meaningful share shift, which makes the 3Q recovery a question about competition rather than AppLovin’s own model iteration.

PLTR 2Q26 — the US Commercial slowdown thesis did not survive the print. Revenue grew 93% YoY to $1.94bn with US Commercial up 149% to $764mn, and FY26 guidance rose $496mn — roughly $362mn of that a genuine second-half revision rather than a roll-forward of the beat.

NET 26Q2 — the security pivot is now visible in the numbers. Revenue accelerated to $696.1mn, up 36% YoY, DBNRR recovered from 118% to 120%, and FY26 guidance was raised ~$58mn with about $26mn added to the second half. Our preview called it from the channel: Zero Trust moving from ~15% to ~25% of one partner’s Cloudflare business, with enterprise platform consolidation replacing standalone CDN/WAF as the driver.

DDOG 26Q2 — unchanged fundamentals met with too-high expectations. Our preview said the print would be solid but that full 3Q expectations and crowded positioning left poor risk-reward into the event; the quarter grew 35.6% with core non-AI accelerating to high-20s, and the guide came light only because of OpenAI optimization.

Premium Report Snapshot

A portion of our research is reserved for Premium subscribers and is not distributed via Substack. Below is a snapshot of what Premium subscribers received this week beyond the Substack feed.

Subscribe now

Read more

• 2Q26 revenue grew 53% YoY, 1% below the high end of guidance. Our revenue estimate was already below the bogey and below AppLovin’s typical 3% beat, but actual execution still came in below our estimate.

• The high end of 3Q26 revenue guidance implies 48% YoY growth, below the bogey. Assuming a typical 3% to 4% beat, implied 3Q26 revenue growth is roughly 53% YoY, broadly in line with 2Q26. Still, because 2Q26 did not deliver its typical beat, investors may question whether the company can return to that level.

Unity Performance Snapshot

• 2Q26 Unity Ads revenue grew 23% QoQ, modestly above our 20% expectation. Based on guidance and the typical beat, Unity Ads growth could remain near this pace in 3Q26.

AppLovin Earnings Call Takeaways

Why AppLovin Missed

• In our preview, we noted that 2Q26 could fall short of AppLovin’s typical beat. However, gaming advertising improved meaningfully from early July. This is fully consistent with management’s commentary on the call.

• Management repeatedly emphasized that the miss was not demand related. It reflected the timing of model improvements: “We did not meet our own standards this quarter. The key point is that we know what happened, the issue has been resolved, and the third quarter is off to a strong start.”

• “Gaming remains the majority of our revenue, and model performance is the largest single driver of its growth. This quarter came down to timing. The cadence of meaningful model improvements was lighter than usual during the quarter, and the next step up in model performance arrived just after quarter end.”

• When asked what model breakthrough had not materialized, management said: “Looking back over the past 12 quarters, every quarter other than this Q2 delivered very strong growth. Every quarter in which we outgrew expectations was supported by model improvements. This time, Q2 did not receive the magnitude of uplift we saw in prior quarters, and that uplift arrived just after the quarter ended.”

• On why the expected uplift did not arrive during the quarter, management stressed that this is inherent to R&D: “That is R&D. There is no guarantee that we will have an improvement every three months. The system is essentially a large set of A/B tests looking for gains. There will be periods when we do not find material improvements, and periods when we see very large gains that drive 12%, 13%, or 15% QoQ growth.”

• Management said demand and the competitive environment had not deteriorated: “Importantly, we have not seen any indication of weaker advertiser demand or a change in the competitive environment. Publisher revenue on MAX grew at a double digit sequential rate, and our share of publisher waterfalls remained stable.”

• “The guidance incorporates the model improvements already launched and the compute cost increases we have observed.”

Consumer Business Progress

• The consumer business performed exceptionally well, with spend up 28% even versus Q4, the seasonal peak: “Advertiser spend reached another record and was 28% above 4Q25. Remember that Q4 is the seasonal peak for these advertisers. Exceeding peak season levels by this much in a seasonal trough shows how steep the curve is.”

• Growth was driven primarily by existing advertisers rather than new logos: “New advertisers are helpful, but the underlying business is already quite strong. Growth of this magnitude in Q2 shows that existing customers on the platform are seeing substantial success.”

• On advertisers remaining well below their ROAS ceiling: “These companies manage budgets one to four quarters in advance. Most of their budgets are allocated to social and search. We are viewed as a new budget category, essentially a test category, and moving up takes time. These dynamics compound over quarters and years.”

Why EBITDA Came in Below Guidance

• EBITDA came in modestly below guidance, but management said all incremental investment was directed toward technology and compute, with returns as the gating factor: “The incremental investment is going exactly where we think it should go, into technology. We have been investing in architectural changes that allow us to build more sophisticated models that benefit more from additional training compute. When additional compute drives meaningful revenue upside through better model performance, that is a trade we would make every day.”

• Compute intensity guidance: roughly $0.10 of compute spend for every incremental $1 of revenue. “For every incremental dollar of revenue, we spend about $0.10 on compute. That is the level embedded in guidance. We do not expect that to change, although it may fluctuate over time.”

• Management said it would not delay a more complex model simply to manage quarterly cadence: “If someone finds a way to build a much more complex model with far more parameters, train and run that larger model, and generate a material revenue uplift, we will not hold it back.”

SEC Investigation into AppLovin Has Concluded

• “It was a voluntary request for information, and we never considered it material. The SEC recently informed us that its investigation has concluded and that it does not recommend any action. We are pleased to have the matter resolved.”

AppLovin Ads Manager Is Generally Available

• The platform became publicly available under the AppLovin Ads Manager name at the end of June, but the rollout is deliberately phased, beginning with the midmarket: “We launched the platform to the public as AppLovin Ads Manager. We are deliberately taking it step by step, starting with midmarket advertisers, where the platform is performing best today. As the data compounds, the long tail will gradually open up as it did in gaming. We will bring this to market through partners.”

• On larger Fortune 500 brands, management said it will eventually pursue the segment, but first needs a sales force and a reputation for having proven the midmarket: “To win large customers, we need salespeople. We also need a reputation that proves we have made the midmarket work. If we can build a truly strong reputation here, we will absolutely move into the top end of the market.”

• Management sees attribution and analytics partners, including commerce analytics platforms, as a targeted route to the right customers: “If we go to the source that works with these companies, that is a more targeted way to bring the right advertisers onto the platform. That creates a structure in which everyone wins.” When an analyst asked whether this was similar to Triple Whale, Adam replied: “You have basically got it.”

• On whether advertisers will rush in as they did at Meta or Google, management said the platform remains concentrated among a small number of customers: “It took us 14 years to fully penetrate gaming. Building this system takes time, but we are seeing very rapid growth because the customers on the platform are having substantial success.”

Unity Earnings Call Takeaways

Why Unity Beat

• In our preview, we expected Unity to deliver a significant beat, primarily because of D28 IAA. D28 IAA drove nearly a doubling of net new budget. This is our most constructive quarter on Unity since we initiated coverage.

Subscribe now

Read more

26Q1 was characterized by “strong revenue, but relatively conservative and less transparent forward KPIs and guidance, leaving the market concerned about a potential growth slowdown.” 26Q2 reversed that setup: “revenue growth accelerated further, while RPO, DBNRR, large customers, and guidance all improved simultaneously.”

26Q1 revenue growth itself did not decelerate, but RPO, DBNRR, gross margin, and full-year guidance were relatively conservative, leading investors to question the quality of growth. 26Q2 saw a broad-based recovery: revenue growth accelerated from 33.5% to 36%; DBNRR recovered from 118% to 120%; large-customer growth improved from 25% to 27%; total RPO grew ~7% q/q; 26Q3 guidance still implies ~31% y/y growth; and the magnitude of the FY26 revenue guidance raise was meaningfully greater than simply flowing through the 26Q2 beat.

The results also validated the core thesis in the 26Q2 preview: growth drivers are shifting from traditional CDN/WAF toward broader enterprise platform expansion, with Workers, Zero Trust/SASE, the partner channel, and large-customer consolidation all becoming increasingly important. At the same time, Workers monetization is progressing faster than anticipated in the preview; the preview’s concern around H2 IT-budget/financing constraints has not yet shown up in 26Q2 results or 26Q3 guidance.

1. 26Q2’s Biggest Highlight: Growth Accelerated Further, with a Meaningfully Better-Quality Guidance Raise

Subscribe now

Read more

Result highlights: an expectation miss

Subscribe now

Read more

July 2026 produced the sharpest drawdown of this cycle for the AI trade. Semiconductors posted their worst month since 2002, and more than $1 trillion of AI-linked market value came off. The concern behind the move is straightforward: coding looks like a special case that is hard to replicate, enterprise adoption is running behind expectations, and the “next coding” that would justify hundreds of billions of dollars of capex has not yet appeared.

We think the debate is framed around the wrong question. The market is evaluating AI through an application lens, looking for the next killer use case and checking whether current revenue covers capital expenditure. The industry’s own objective, however, is AGI, and the mechanism most labs point to on that path is recursive self-improvement (RSI). A more useful way to think about it is as a sequence of capability milestones, each with its own commercial payoff. Coding capability has already been monetized, in the form of this year’s first-half revenue inflection. The next milestone is continual learning, which lets models accumulate knowledge in deployment the way an employee does; combined with deployment mechanisms such as forward-deployed engineering, it will open up most of the enterprise opportunity. Full RSI, in which models participate in and accelerate their own development, sits furthest out and will determine the competitive structure that follows. Progress is visible at each stage, and in several places it is accelerating.

We make three claims.

First, the “context problem” the market worries about is a capability question rather than a use-case question, and it is what continual learning is designed to address. Once models can accumulate tacit organizational knowledge the way employees do, the search for a single “next coding-like” use case dissolves: enterprise use cases open up progressively. The TAM reference then shifts from a software budget measured in hundreds of billions to a labor market measured in tens of trillions.

Second, early signals of RSI are already observable. Model iteration has moved from annual to monthly. GPT-5.5 took roughly one month from the end of pre-training to release. OpenAI cut prices by up to 80% three weeks after GPT-5.6 shipped, against a history in which price cuts followed months of inference optimization. Lab code is now written almost entirely by models, and AI’s measured contribution to internal R&D velocity rose from 5% to 15–20% over six months.

Third, if RSI crosses its threshold, catch-up strategies built on distilling frontier model outputs are likely to become progressively harder, and compute becomes the least compressible constraint and the deepest competitive moat. When the frontier iterates monthly, a follower may still hold a fixed time lag, but it cannot readily replicate a frontier lab’s full learning, data, and experimentation stack. Native self-improvement capability requires proprietary experimental infrastructure and the compute to run it. This is also the view inside China’s leading labs: China has not hit a scaling wall, but does not yet have enough compute to reach one.

Capex, on this reading, should not be dismissed as a bubble. The latest earnings season shows real current returns from AI-driven cloud growth, and beyond that, capex is what ensures a player has sufficient compute when the RSI threshold arrives. Strategic necessity is not the same thing as shareholder return, and we treat that distinction explicitly in a scenario framework. The variable that matters most is less “what is the next application” than when RSI arrives and who gets there first.

1. The Debate: Has AI Hit a Wall?

Subscribe now

Read more