I had to say, I was looking forward to earnings, and while I expect it to report impressive results, due to the current macro, I didn’t expect the price to follow. Tech sector this year reported impressive ER vs estimates, either meeting or exceeding estimates in nearly 80% cases. Despite that, most of the time, the stocks sold off after earnings.

It was no different with Credo last evening. The stock dropped from $206 at close to $185. Let me remind you that Credo hit its ATH in June and traded at $308.67.

Is it justified?

Let’s find out.

If you’re new to Credo, please check my deep dive (500k+ views on X).

Credo reported record financial results for its fiscal Q1 2027, massively exceeding both internal guidance and consensus estimates. Revenue reached an all-time high of $479.0 million (up 114.7% YoY and 9.6% QoQ), while non-GAAP diluted EPS reached $1.20, outperforming the $1.16 Street consensus. Furthermore, management confidently raised its full-year fiscal 2027 revenue guidance from 'greater than 80%' to 'greater than 85% year-over-year'. Despite this flawless fundamental execution, CRDO shares experienced a sharp post-market contraction, tumbling 8.65% in regular trading to close at $206.63, and sliding another 10.13% in after-hours trading to $185.00.

This report provides a deep analysis of Q1 results and validates the long-term investment thesis.

FISCAL Q4 2026 RECAP

In the previous quarter (Q4 FY2026 ended May 2, 2026), Credo established a robust baseline. The company delivered revenue of $437.0 million (up 157% YoY) and non-GAAP EPS of $1.16, beating estimates of $432.0M and $1.02. Gross margins came in strong at 68.3% (Non-GAAP), and the business generated $226.7 million in non-GAAP net income.

However, the Q4 print was characterized as a 'narrower beat' compared to prior quarters (exceeding guidance midpoint by only $7 million), which triggered a mild margin scare.

Q1 FY2027 PERFORMANCE VS. GUIDANCE & STREET ESTIMATES

Credo's performance in Q1 FY2027 represents a 'double beat'. The company surpassed the high end of its internal revenue guidance range by $4.0 million and outperformed consensus Street expectations.

Below is a detailed, line-by-line performance reconciliation.

The GAAP vs. Non-GAAP Margin difference:

The primary catalyst for short-term sentiment deceleration was the sequential drop in GAAP gross margin to 64.5% from 68.2% in Q4 FY26. However, this compression was entirely non-operational. As shown in the 8-K reconciliation tables, the closing of the DustPhotonics transaction introduced a $11.0 million amortization of acquired intangible assets and $5.7 million in stock-based compensation within cost of revenue. Adjusted for these non-cash expenses, the core Non-GAAP gross margin was stable at 68.0%, reflecting highly resilient unit economics and robust pricing power.

Q2 FY2027 OUTLOOK & EXPANDED FULL-YEAR GUIDANCE

Credo provided extremely bullish forward looking expectations for the upcoming fiscal Q2 ending October 31, 2026. The company guided Q2 revenue to $525.0 million – $535.0 million, representing 10.6% QoQ growth and crushing the Street consensus of $520.0 million. Non-GAAP gross margin is guided to 67.0% – 69.0% (68% midpoint), confirming stable unit economics. Q2 Non-GAAP operating expenses are guided to $100.0M – $105.0M.

FULL-YEAR FY2027 GUIDANCE RAISE:

Management formally raised its full-year total revenue growth expectation from 'greater than 80%' to 'greater than 85% year-over-year'. This translates to full-year revenue exceeding $2.47 billion. Furthermore, management expects its Non-GAAP net margin to remain near 50% for the full year, resulting in expected full-year adjusted net income of approximately $1.23 billion.

BALANCE SHEET & 8K

Analyzing the Q1 FY2027 balance sheet reveals major capital allocation shifts following the DustPhotonics closing on May 28, 2026:

• Cash and Liquidity: Total cash and short-term investments stood at $764.3 million (cash of $466.9M and short-term investments of $297.4M), down sequentially from $1.44 billion. This drop is directly attributed to the $750.0 million upfront cash payment for the DustPhotonics transaction.

• Goodwill & Intangible Assets: Goodwill skyrocketed from $92.8 million to $986.4 million, and net Intangible Assets jumped from $29.3 million to $378.8 million, collectively representing $1.365 billion (up $1.24B QoQ). This represents the acquired technological IP (Carmel PIC platform) and goodwill associated with the transaction.

• Strategic Inventory Buffer: Inventory rose sequentially by $62.2 million to $313.1 million. CFO Dan Fleming confirmed this was a deliberate choice: 'Inventory levels increased... reflecting supply chain investments'. This mitigates foundry and capacity constraints ahead of the second-half volume ramp.

HYPERSCALER CONCENTRATION

A vital risk factor in the Credo investment thesis is customer concentration, which remained high in Q1. The top four domestic customers represented 84% of total revenue, structured as follows:

While 84% concentration represents a risk, some sell-side analysts view this “issue” as an infrastructure moat. High-speed networking involves multi-year qualification cycles in the world's most critical AI data clusters. CFO Dan Fleming highlighted that concentration is accompanied by deep product diversification within each account: “For sure, there is a strengthening and broadening across the board. It is not just AECs that these hyperscalers are consuming of ours.”

THESIS CHECK (VS EXPECTATIONS IN THE DEEP DIVE)

$600M+ Optical Revenue Target in FY27: Credo guided above $600M of optical revenue across ZF Optics, SiPho PICs, and optical DSPs, with each contributing $100M+.

On the call, CFO Dan Fleming reiterated this exact breakdown: “We continue to expect an inflection in the second half, driven by more than $600 million in optical revenue with ZeroFlap Optics, silicon photonics, PICs, and optical DSPs, each contributing more than $100 million.”

Furthermore, CEO Bill Brennan highlighted that ZeroFlap Optics is experiencing active customer ramps across both hyperscalers and neo-clouds, positioning it as the leading near-term volume contributor.

50/50 Copper/Optical Mix & Large Optical TAM: thesis highlights a 50/50 long-term mix and the potential for optical to exceed copper.

CEO Bill Brennan reiterated this long-term balance during Q&A: 'Long term, I think you'll see AECs grow, and as we continue to scale as a company, I think you'll see a really nice balance between copper solutions and optical.' Optics is growing faster simply because it starts from a smaller base. The addressable pluggable optical transceiver market is projected by forecasters to grow from 60 million units in 2026 to 175 million units by 2030, presenting an enormous TAM.

SiPho PIC Traction & DustPhotonics In Q1, Credo recognized its first-ever silicon photonics PIC revenue. More importantly, Brennan announced: 'We have got design wins with two major players for next generation ramps that will occur in FY 2028 and maybe starting in the late part of this fiscal year.' Critically, these first two wins do not include Credo DSPs, meaning Credo sold standalone PICs. This leaves a massive up-sell pathway once Credo begins bundling its DSPs and PICs together

Co-Packaged Optics (CPO) and Near-Port Optics (NPO) ahead of the FY2028 ramp. On the call, Brennan provided major updates on this front. Credo officially joined the Open CPX MSA consortium specifically to address the NPO scale-up interconnect market. Brennan said: 'We are seeing increasing customer activity around NPO for scale-up networks with confirmed design wins expected to begin ramping in our fiscal 2028... we'll lead with the silicon photonics PIC, but we'll also look at doing the complete optical engine long term as we add more functionality to our portfolio.' This establishes a firm operational pathway to monetization in scale-up GPU fabrics.

Is the H2 Inflection on Track?

A key concern for bears has been the heavily back-end-loaded nature of Credo's FY2027 revenue targets. In Q4, management guided to 'mid-single-digit percentage' sequential growth in H1. However, Q1 delivered 9.61% sequential growth, and the Q2 midpoint of $530M implies 10.65% sequential growth. This acceleration proves that H1 is already tracking significantly ahead of the historical run-rate, which materially de-risks the H2 inflection. To hit its raised full-year target of >85% growth ($2.47B), Credo needs to average roughly $730 million per quarter in H2. This will be supported by production ramps of ZeroFlap Optics, Blue Heron 200G/lane retimers, and early 1.6T optical DSP shipments.

DUSTPHOTONICS INTEGRATION

The acquisition of DustPhotonics represents a shift in Credo's competitive positioning. Before this deal, Credo was a high-performance analog silicon vendor. Now, by integrating DustPhotonics' Silicon Photonics (SiPho) PIC technology with its own DSP and SerDes IP, Credo is the only fabless player that owns the entire high-speed optical connectivity stack:

SerDes IP ➔ Optical DSP ➔ Silicon Photonics PIC ➔ Pluggable Transceiver (ZeroFlap) ➔ PILOT Software

This full-stack vertical integration provides three distinct structural advantages:

1. Cost Leadership: Eliminates margin stacking by cutting out third-party optical component providers. As port speeds transition from 800G to 1.6T and 3.2T, silicon photonics packaging yields become the dominant cost factor. Owning the PIC layer allows Credo to optimize manufacturing yields directly, lowering costs at volume.

2. Telemetry and the Reliability Moat: By optimizing the DSP and PIC together, Credo can light up 'continuous telemetry' via its PILOT software platform. Brennan highlighted that PILOT monitors critical physical-layer metrics (eye height, SNR, post-FEC histograms) across all six segments of an optical link. This acts as a 'check engine light' for AI cluster operators, identifying fiber degradation, dust, or electrostatic damage before a link flap occurs, preventing expensive GPU idle time.

3. Strategic Up-sell: Standard transceivers compete in a crowded merchant market. By offering ZeroFlap Optics, which integrates hardware and telemetry SDKs directly into switch software, Credo creates a high-margin, sticky solution that rivals cannot provide.

VALUATION MULTIPLES (AT $185.00 SHARE PRICE)

The market's reaction to Q1 results was primarily driven by valuation multiple compression. Trading at over 90x trailing GAAP P/E prior to earnings, any minor increase in full-year spending triggers systematic multiple contractions. At the current after-hours share price of $185.00, we analyze Credo's valuation framework based:

For a technology leader that is growing revenues at a 114.7% YoY and maintaining a 49.3% non-GAAP net margin, a Forward FY2028 multiple of 20.2x represents an exceptionally compelling valuation.

FY2028 EMERGING PRODUCTS & TAM EXPANSION

While the optical ramp defines the FY2027 narrative, the earnings call highlighted three massive product pillars planned for production revenue in FY 2028, representing significant underappreciated upside for the bull thesis:

• OmniConnect AI Memory & 'Weaver' Gearbox: Designed to address the 'memory wall' and memory fanout bottlenecks in large-scale AI inference clusters. Bill Brennan revealed: 'We believe OmniConnect solutions can represent thousands of dollars of Credo content per GPU with revenue beginning in fiscal 2028.' Brennan confirmed their first customer, Positron, is deploying a 2TB Weaver-enabled system.

• Active LED Cables (ALCs): Utilizing microLED emitters, ALCs extend the reliability and power benefits of active copper cables to reaches of up to 30 meters. This bridges the gap between active copper (AECs, up to 7m) and traditional laser-based optics, creating a third differentiated product category. Real-world demonstrations will occur at OFC in October, with production scheduled for FY28.

• Blue Heron 200G/Lane & 1.6T AECs: The transition to 200G/lane architectures is a major growth vector for copper. Brennan confirmed that 1.6T AEC deployments will begin contributing in late FY27 and ramp in a much larger way in FY28, with 200G-per-lane serving as the volume crossover point during the fiscal year

CONCLUSION

I see no flags with my Credo thesis.

The story is getting better and better, and the stock became a lot cheaper.

Micron reported Q3 revenue of $41.5B with a gross margin of 85%, and guided Q4 revenue of $50B which is more than the revenue of all of FY2025 in a single quarter.

INTRO

Three distances from the processor, one company at all of them

For nearly two years, memory was the layer of AI infrastructure that no one could price correctly. It is also the layer that determines whether tens of thousands of GPUs (often even more) run at full capacity or wait for data to arrive. In inference, model weights must be constantly fetched to generate responses, making memory bandwidth the critical bottleneck.

The memory isn’t one single market. At its core, memory solves one basic problem - keeping data close to the processor - by offering three distinct solutions at varying distances from the die.

On the package - millimetres away

This is HBM (High Bandwidth Memory): stacks of DRAM dies connected through thousands of through-silicon vias and mounted beside the GPU or accelerator on the same package, typically through an advanced interposer. HBM exists because conventional motherboard traces cannot provide the bandwidth and energy efficiency required by modern AI accelerators. In many leading AI accelerators, achievable performance is constrained by both compute throughput and how quickly HBM can feed data into the chip.

Three companies on earth make HBM at scale: SK Hynix, Samsung and Micron.

On the board - centimetres away

Standard DDR5 server memory and LPDDR low-power modules. A lot of the market just assumed that this line would stay commoditised forever while HBM took all the margin. That’s not how it happened, and that is one of the most underreported facts in this story: Micron’s chief business officer told investors that non-HBM DRAM margins have in periods exceeded HBM margins. This tells us that memory scarcity is not confined to HBM.

In the drive bay, and (increasingly) across the rack.

NAND flash and the SSDs built from it. This is a slower storage that keeps data when the power goes off. Data centre SSD revenue passed $5B in FQ3 alone, more than doubling sequentially.

Most companies at this layer pick one distance and focus on it alone. SanDisk and Kioxia are focusing on NAND. Nanya is focused on legacy DRAM. Micron shows up at all three.

This is the reason why every single of four Micron’s business units posted gross margins between 79% and 87%. This includes the automotive and embedded line that nobody would call an AI trade.

On the June 24 earnings call, Micron’s management said that DRAM and NAND industry demand continues to significantly exceed industry supply, and that they expect tight conditions to persist beyond CY2027.

SECTION 01

THE COMPANY

How a dentist’s-basement startup became the scarcest asset in AI infrastructure

The company is listed on Nasdaq, ticker MU.

Micron was founded on October 5, 1978 by four men working out of the basement of a dental office in Boise: twin brothers Ward and Joe Parkinson, Dennis Wilson and Doug Pitman. In 1978, starting a memory chip company was almost the worst timing possible because Japanese manufacturers were about to crush the American DRAM industry. Micron) survived because a potato billionaire, J.R. Simplot, invested early and became its biggest shareholder. Simplot was an eighth-grade dropout who made his fortune supplying McDonald's with frozen fries.

For the next four decades it was the textbook cyclical, buying distressed capacity at the bottom and getting punished at the top after the supply catches up. Revenue moved sharply from $30.8B in FY2022 to $15.5B in FY2023. That is the business most people still think they are buying today.

Micron is a 48 years old manufacturing company that kept pouring capital into leading-edge DRAM through cycle after cycle that punished it for doing so, but then found itself holding the one input the entire AI buildout cannot create on its own.

What happened next was impossible to miss. Quarterly revenue went $11.3B, then $13.6B, then $23.9B, then $41.5B. Micron crossed $1 trillion in market capitalization for the first time on May 26, 2026, becoming the third U.S. semiconductor company to join the trillion-dollar club after NVIDIA and Broadcom.

SECTION 02

THE BUSINESS MODEL

Where the revenue comes from, who pays it, and the contracts that changed the format

In April last year, Micron reorganized its reporting. The company moved from technology-based reporting to market-facing segments. Reorganising around customer markets rather than product technologies matters because it makes it easier to see where demand is coming from and which end markets are absorbing the current pricing increases.

Here is where the revenue comes from.

Every segment grew by triple digits, but the more useful observation is how evenly that growth was distributed across the portfolio.

The revenue base is spread across all four segments:

• Cloud Memory is 33.2% of revenue and 32.3% of the growth

• Core Data Center is 27.8% of revenue and 31.1% of the growth

• Mobile & Client is 27.8% of revenue and 25.7% of the growth

• Automotive & Embedded is 11.2% of revenue and 10.9% of the growth

No segment’s share of total growth differs from its share of revenue by more than roughly 3%. And the two lines with no AI story attached - Mobile & Client and Automotive & Embedded - are 39.0% of the total revenue, contributing 36.6% of the growth.

Every segment today sits between 79% and 87% gross margin, while twelve months ago the same four segments ranged from 24% to 39%.

That convergence is the most important information on this chart. If the shortage were confined only to AI-specific products, the data-center segments should have separated much more sharply from the rest of the portfolio; instead, pricing and margin expansion spread across all four segments.

Margins of a company do not converge because a product got popular. It happened because there is no supply left anywhere in the industry, so every customer pays scarcity pricing regardless of what they are using Micron’s products for. That is the bull case in short and, since scarcity is not a moat, it is also the bear case.

Under all four business segments, there’s a single constraint: wafer output. Micron does not have a demand problem but a capacity problem. Every fab it announced lands in FY2027 or later.

The contracts that broke the format

In Q3 this year Micron reported 16 strategic customer agreements carrying remaining performance obligations of approximately $100B. Sell-side estimates put the current agreements at roughly 20% of DRAM volume and around one-third of NAND volume, while Micron says it ultimately expects roughly half or more of company revenue to be covered by SCAs.

Take-or-pay contracts with price floors, in memory?

The memory industry spent decades relying heavily on spot and short-cycle pricing, a structure that amplified price declines whenever supply loosened. Management is now saying that the SCAs will “fundamentally transform” their business.

Who buys from Micron?

Hyperscalers, AI labs, PC and handset OEMs, and Tier 1 automotive suppliers. Micron does not name the SCA counterparties, and customer concentration is not disclosed.

The Anthropic agreement is the one to sit with, because it shows where this demand is actually coming from. When a frontier AI lab signs directly with a memory manufacturer rather than buying through a server vendor, it is telling you that the memory allocation has become strategic enough to negotiate at the top of the stack.

SECTION 03

THE PRODUCT PORTFOLIO

What each product actually is, what job it does, and why it matters

This section explains the products because memory is often discussed as if it were one commodity with one price; in reality, Micron turns related DRAM and NAND technologies into products that serve very different workloads and carry different margins.

Where the $41.5B came from, by product

This is another view of the same quarter: Section 2 split revenue by end-market segment, while this section splits it by product family.

Two details are worth pointing out. Micron is still overwhelmingly a DRAM company with 76% of revenue, with HBM living inside that number rather than beside it.

Most of the sequential growth came from pricing rather than bit-volume growth.

• DRAM bit shipments rose in the low single digits while prices rose in the low 60s percent range.

• NAND bits rose mid-single digits while prices rose in the mid 80s.

This means that Micron sold roughly the same quantity of memory and charged a lot more for it.

Products

DRAM - Dynamic Random-Access Memory

DRAM is the working memory in every computer on. It is in our phones, our laptops, and also in every server in every data centre. It's where a processor keeps whatever it's actively working on, and nothing computes without it. It's Micron's single largest product, and the foundation everything else here is built from: HBM, server memory, mobile memory are all this same technology packaged for a different job.

Each DRAM bit is stored as electrical charge that leaks away and therefore must be refreshed regularly; DRAM is orders of magnitude faster than NAND storage because its cell architecture allows much lower-latency electronic access, not because of the refresh process.

• HBM - High-Bandwidth Memory

  • HBM is memory that sits beside an AI accelerator and feeds it data at extremely high bandwidth; for many AI workloads, that memory bandwidth is one of the key constraints on achievable performance. It's made by stacking eight to sixteen DRAM chips, drilling tiny channels through them so every layer can talk to the processor at once, then mounting the stack immediately beside the chip instead of on the motherboard.

• DDR5 server memory

  • DDR5 is the standard memory in practically every server. It’s the same DRAM as above, but sold as a plug-in module instead of a stack. It's Micron's highest-volume product, and it was supposed to be the boring, commoditised half of the business. But it hasn't been. Micron's management said that non-HBM margins have in periods exceeded HBM's. It's slower than HBM but much cheaper and holds far more data per module.

• LPDDR and SOCAMM

  • LPDDR is DRAM designed to reduce power consumption through lower operating voltages and aggressive power-saving features, trading some peak performance for substantially better energy efficiency. Normally these chips are soldered permanently onto the phone’s circuit board. SOCAMM2 takes that same low-power memory and puts it on a small removable module instead, closer to how memory works in a regular computer — so a server can plug it in and swap it out, rather than having it fixed in place.

  • LPDDR was originally designed for battery-powered devices, where reducing memory power consumption materially extends battery life.

NAND flash - Non-volatile storage

NAND is the memory that holds the data when the power is off. It's inside every SSD, USB stick and memory card. It has historically been the more volatile of Micron’s two core memory markets because the supplier base is broader and commodity pricing can reset sharply when capacity outruns demand. NAND revenue grew 361% YoY in FQ3, while sequential bit shipments increased only in the mid single digits and sequential ASPs rose in the mid 80% range, and the latest quarter’s step-up was therefore overwhelmingly pricing driven.

• Data-centre SSDs

  • A data-centre SSD is a finished drive built from Micron's own NAND where Micron adds the same chips mentioned above, wrapped in a controller and sold as complete hardware. That's where Micron keeps the assembly margin instead of handing it downstream.

Automotive and embedded - DRAM and NAND qualified to automotive reliability standards

It’s the same DRAM and NAND above, but built to survive extreme heat, cold, vibration and fifteen years of use which is the bar for anything installed in a car. I mention this not because of its revenue size, but because it shows that pricing and margin expansion are extending beyond the data-center segments.

SECTION 04

THE FINANCIALS

The quarter the cycle stopped looking like a cycle

Fiscal 2023 was the trough of the last memory downcycle, with $15.5B of revenue and heavy losses. Fiscal 2026 will land somewhere near $129B on company guidance.

Here’s the cleanest way to see what happened here: a year ago, now, and what’s coming up next.

Micron’s non-GAAP gross margin rose from roughly 39% to 45%, then to 74.9% and 84.9% over the last four reported quarters, with roughly 86% guided for fiscal Q4. Micron’s current earnings inflection is not primarily a fixed cost absorption or volume-growth story; it is overwhelmingly a pricing story. In FQ3, DRAM bit shipments rose only in the low-single-digit percentage range QoQ while average selling prices rose in the low 60s; NAND bit shipments rose in the mid-single digits while prices rose in the mid-80s. Micron shipped only slightly more memory and charged dramatically more for it.

The margin expansion has been driven mainly by pricing, not by volume growth.

The bars above show the scale-up. The line shows the repricing in memory industry. Revenue roughly quadrupled from FQ4 FY2025 to the FQ4 FY2026 guide, and gross margin more than doubled over the same period, and it was climbing every single quarter with no drop in between. Guidance implies the margin line will keep growing into the fifth quarter, which is the detail worth watching most closely at the next print (green bar).

SECTION 05

THE INDUSTRY

Supply and Demand

The current AI expansion has triggered a serious imbalance in memory supply and demand. Before, during temporary demand spikes we had competitive overbuilding, but the current market is restricted by a physical, technological siphon: the production of HBM.

HBM stacks multiple DRAM chips vertically on top of each other and places them directly next to advanced AI processors. This design is essential to feed massive streams of data into high-performance GPUs, solving the data speed limits (or “memory wall”) in AI training and inference.

However, this architecture comes with a heavy production cost on raw silicon wafers:

• On a comparable bit-output basis, HBM3E requires roughly three times as much DRAM wafer capacity as conventional DRAM because of larger die sizes, stacking requirements, yield loss and advanced packaging.

• As the industry moves to HBM4 and taller stacks, the trade ratio is expected to worsen further; some industry estimates put it near 4:1, although the exact penalty depends on die size, stack height and yield.

• In 2026, HBM-related production is expected to consume roughly 23% of global DRAM wafer capacity, diverting nearly a quarter of DRAM wafer output toward high-bandwidth memory rather than conventional DDR and LPDDR products.

That shift has reduced the wafer capacity available to conventional DRAM, contributing to higher prices across PCs, smartphones and enterprise servers. HP Inc. said memory and storage costs rose from roughly 15–18% of its PC bill of materials to about 35%, illustrating how sharply component cost inflation has moved through the PC supply chain. PC makers have responded to higher memory costs with a mix of price increases, lower memory configurations, supplier diversification and tighter product segmentation.

Hyperscaler Capex

The Competitive Landscape

Unlike custom logic chips, which rely on proprietary designs and architectures, DRAM depends on manufacturing scale, process technology and complex engineering. This has created a tightly consolidated global market. In Q2 2026, a dominant triopoly controlled roughly 90% of all DRAM revenue: Samsung Electronics (39%), SK Hynix (26%), and Micron Technology (25%). The fastest-growing outside challenger to this group is China's state-backed CXMT, which currently holds a 7% market share.

The Memory Triopoly

SK Hynix
SK Hynix is the top HBM supplier, with a 58% share of HBM revenue as of Q1 2026. Its edge comes from its proprietary MR-MUF packaging process, which gives better thermal performance than rivals. SK Hynix sits at the center of NVIDIA’s HBM supply chain alongside TSMC, and some sell-side estimates put its share of next-generation HBM4 demand for Rubin at roughly 70%.

Samsung Electronics
Samsung holds 21% of global HBM revenue, using its dominant 39% share of conventional DRAM to fund HBM expansion. Samsung was the first supplier reported to ship HBM4 to NVIDIA and also began shipping 12-high HBM4E samples in 2026, while its in-house foundry gives it the option to manufacture custom logic base dies internally.

Micron Technology
Micron held 21% of global HBM revenue in Q1 2026 and has qualified HBM3E on NVIDIA’s Blackwell platform. Its main advantage is geopolitical — it’s the only memory maker headquartered in the West. With billions of dollars of CHIPS Act support, Micron is building major fabs in New York and Idaho and has disclosed roughly $100B of remaining performance obligations under SCAs spanning data-center, consumer and automotive customers.

CXMT & YMTC
China’s state-backed memory players CXMT (DRAM) and YMTC (NAND) pose the biggest long-term competitive threat in memory. Micron is especially exposed, having been largely cut off from selling to key Chinese infrastructure since a May 2023 CAC ruling.

These Chinese companies operate under several constraints:

• Export controls prevent leading Chinese memory producers from obtaining EUV lithography systems, forcing them to rely more heavily on older ArF immersion tools and multi-patterning for advanced nodes, which creates additional yield and cost challenges relative to the leading-edge suppliers.

• Despite the tech gap, CXMT has grown fast, reaching 7% of global DRAM revenue by Q2 2026. With state subsidies and no real cost-of-capital pressure, both firms are building huge capacity that’s steadily pushing Western memory out of the Chinese market.

• NAND/SSD margin split: YMTC has grabbed 13% of global NAND share as of early 2026. This scaling creates a split market: high-end finished products still have strong margins, while raw NAND components face constant price pressure from subsidized Chinese supply flooding the low end.

TAM Projections

SECTION 06

THE SETUP

Why now?

The obvious question is why now?

It’s a stock that significantly re-rated in the last twelve months, and there are still fears about cyclicality and Chinese competition.

1. Capacity & Demand - A fab takes three to four years to build. AI memory demand became a dominant driver in only eighteen months. Management expects tight conditions to persist beyond calendar 2027. Micron’s CEO said that while industry supply is expected to improve gradually in 2028, Micron currently does not have line of sight as to when memory supply will actually catch up with demand.

   

2. A larger share of demand is now contracted. Unlike prior cycles that relied heavily on short-cycle PC and handset forecasts, Micron now has multi-year SCAs and customers allocating memory years in advance, although hyperscaler capital plans can still change.

3. Roughly $100B of remaining performance obligations under 16 SCAs, with price floors and ceilings, backed by around $22B of upfront deposits. Micron’s SCA structure is unusually large and financially backed, but it is part of a broader industry shift in which Samsung, SK Hynix, SanDisk and Kioxia are also moving toward multi-year agreements. If it works as designed it does not make the cycle disappear but it does put some floor under it. That is the entire “this time is different” argument.

4. Execution matters too: Micron materially exceeded its fiscal Q3 revenue guide, while its HBM4 12-high ramp is progressing faster than HBM3E did at the same stage. Guided to $33.5 billion for FQ3, it delivered $41.46 billion. It’s 24% beat against its own number. Its HBM4 12-high ramp is running roughly twice as fast as HBM3E did at the same stage, meaning the yield curve that determines whether this segment is profitable is ahead of Micron’s own prior generation.
   And the CHIPS Act money has already been finalised against the Idaho and New York builds, with construction underway.

The points below are the issues I think matter most for testing whether this cycle is genuinely different.

SECTION 07

RISKS

What could go wrong?

Everyone is building into the same window

Micron, Samsung and SK Hynix are all expanding at the same time, and Micron’s own capex is guided from ~$27 billion to above the mid $40 billion range in one year. The company needs more fabs because supply is short, and building more fabs is precisely how supply stops being short. 2028 is the first credible window in which the supply response could begin materially changing the balance. As new capacity ramps across 2027–2029, ASP growth should eventually decelerate and could reverse, but the timing is uncertain and sell-side forecasts do not uniformly show a 2028 revenue decline across Micron, Samsung and SK Hynix.

The margin is price

An 85% gross margin on a commodity-adjacent product is not a durable structural achievement, it is all due to scarcity. Micron is allocating supply and taking price. This “advantage” can reverse faster than it was built, and it can reverse without any operational failure from Micron.

The SCAs are untested

The SCAs are binding take-or-pay contracts and the roughly $22 billion of deposits and financial commitments provide meaningful protection, but the structure has not yet been tested through a severe memory downturn; the remaining performance obligations are contractual commitments, not cash already collected, so enforcement and renegotiation risk still matter.

Chinese competition

Micron fell sharply around CXMT-related news, showing how sensitive investors are to the prospect of subsidized Chinese DRAM capacity, although I would avoid attributing the entire move to a single catalyst.

Micron is not first in the product that defines the cycle

SK Hynix holds HBM leadership and the deepest NVIDIA relationship. Micron’s HBM growth depends on the overall market expanding and on the company maintaining or increasing its allocation at major AI customers; it does not require customers to reduce SK Hynix volumes in absolute terms.

The volatility.

“Micron went from roughly $113 a year ago to an intraday high of $1,255, then fell to roughly $739 during the July sell-off before recovering toward $972, a peak-to-trough drawdown of about 41% in just over a month. Volatility is not making Micron a position you can easily size as if it was a stable large cap.

Bear case.

You are paying a trillion-dollar valuation for a historically cyclical manufacturer whose record margins are scarcity pricing rather than structural advantage. Micron’s competitive position in the defining product is third of three. Memory industry leaders are pouring unprecedented capital into the supply response that has ended every previous cycle.

SECTION 08

VALUATION

Here’s how Street values Micron

Micron’s recent financial performance broke all historical records, yet its valuation multiple has compressed to levels typically reserved for businesses in a structural decline. For the Q3 2026 (ended May 28, 2026), Micron posted GAAP revenue of $41.46 billion (up 346% YoY) and a record-breaking GAAP gross margin of 84.60%. This explosive growth was driven overwhelmingly by higher DRAM and NAND pricing, while COGS increased only modestly relative to revenue; pricing is the dominant driver of the current earnings step-up, although mix, yields and manufacturing efficiency also contribute.

Semiconductor Peer Benchmarking

The primary valuation argument for long-term investors is the massive multiple discount between Micron and other core AI hardware beneficiaries. While logic chip designers (NVIDIA, AMD) and integrated customized logic makers (Broadcom) command premium growth multiples, Micron and SK Hynix continue to be priced at double-digit discounts, despite posting matching or superior financial metrics:

Here’s how I value Micron

SECTION 09

MANAGEMENT & INSIDER OWNERSHIP

Who runs it, who owns it, and what they have been doing with their own shares

Micron is not a founder or insider controlled company. Directors and executive officers as a group owned about 2.69 million shares, or roughly 0.24% of the company based on 1.124 billion shares outstanding. CEO Sanjay Mehrotra was the largest individual insider listed, with about 1.08 million shares at the October 31, 2025 measurement date. By comparison, Vanguard, BlackRock and Capital World Investors owned 8.4%, 7.6% and 6.3%, respectively.

Insider activity in 2026 has been clearly skewed toward selling rather than buying, although this needs more context. Mehrotra has been a regular seller as Micron’s share price appreciated sharply. He sold shares on May 29 at prices largely in the $940–970 range, again on June 26 at roughly $1,128–1,181, and another block on July 24 around $943–966. The July sale was explicitly executed under a Rule 10b5-1 trading plan adopted on January 30, 2026, meaning the sale instructions had been established months earlier.

Other Micron executives and directors have also reported share sales during 2026, including Chief People Officer April Arnzen and Chief Legal Officer Michael Ray.

For the investment thesis, I view this as neutral to mildly negative rather than a red flag. Insider ownership is low, and management has been monetizing shares into a very large rally, so there is no strong insider-buying signal supporting the valuation.

SECTION 10

THE VERDICT

Read more

However, very few people post their performance from one or two years ago. This is likely because they either ended up underperforming or they simply were not invested at the time. On top of the YTD groups, MTD groups have emerged lately. In these spaces, not only is there an expectation of a 300% YTD return, but members are also expected to generate at least 100% to 200% MTD.

Then, there is the other side of the story.

Today’s online crowd not only ridicules legendary investors like Warren Buffett, but they also post comparisons of their own short-term performance against some of the greatest fund managers of all time.

When seeing someone achieve a 5% return, the initial impression should not be that they are bad investors or underperforming. There are several crucial points to consider:

• They could be managing millions of dollars and are uncomfortable chasing high-risk assets.

• They might be experts in specific industries currently going through a cyclical downturn.

• They could be long-term investors whose strategies underperform in volatile, short-term markets.

Furthermore, what does YTD return actually tell us about long-term performance? I believe an investor needs to perform consistently across a 5 to 10 year period before their track record can be accurately judged.

Someone just starting out could simply get lucky and jump 300% on a poor quality stock, while an unlucky investor could be down 20% on a fundamentally solid company due to short term market noise.

To summarize: never be discouraged by a 20% to 30% annual return.

Sustaining that level of performance over the long run is guaranteed to completely transform your financial life.

To put this into perspective, an initial investment of $10,000 compounding at 30% annually grows to $137,858.49 after 10 years. If you stretch that horizon to 20 years, that same $10,000 becomes $1,900,496.38.

If you think that’s too long, I have to remind you of my favorite quote from Lynch which is: “People don’t want to get rich slow, so they get poor quick.”

Stay patient. Don’t feel bad about having “only” 20, 30 or 50% annually.

You’re doing great.

Plus, I’m teaming up with this week to bring you the top headlines across macro, markets, FinX, AI, and tech.

Iran sanctions

Treasury Secretary Bessent unveiled the sanctions package Monday, targeting five channels - digital assets, technology, gold, aviation, and shipping - and blacklisting more than 60 Iran-linked individuals, vessels, and entities.

On Truth Social, Trump called it “the most crushing economic operation ever taken against any country” and warned countries that keep any economic ties to Iran would face severe consequences. Notably, at the press conference a reporter pointed out this wasn’t really a “D-Day” at all since no sanctions were imposed immediately and Bessent’s response was blunt: “Why would I want to blow up the global financial system?”

Iran escalated to formal diplomatic channels. Araghchi sent a letter to UN António Guterres calling the sanctions “economic terrorism” and asking the UN to reject and refuse to implement them. He also raised humanitarian concerns, arguing the measures could restrict access to food, medicine, and medical equipment for ordinary Iranians.

On Friday, Iran’s Foreign Ministry publicly urged its trading partners not to comply with the sanctions, calling them “illegal and unjustified.” But the same day, Araghchi struck a bit different tone in a separate comment, saying diplomacy with the US was “not impossible” if Washington understood that “pressure doesn’t work”.

Canada tariffs

Trump threatened 50% tariffs on Canadian cars, trucks, auto parts, and steel starting Jan. 1, 2027, writing on Truth Social “Canada has been ripping off the United States of America for years.” This came after Canada’s PM Mark Carney vowed to retaliate “dollar for dollar,” and after the US had already imposed 50% tariffs on about $20 billion of Canadian goods (wine, cement, hockey sticks) that Saturday when trade talks collapsed.

On Tuesday, Canada announced retaliatory tariffs on roughly $20 billion of US goods, hitting steel, dairy, and other categories, in response to Monday's US tariff moves.

The Conference Board's Consumer Confidence Index — a widely watched monthly gauge of how Americans feel about the economy — fell to 89.4 in August, down from 90.2 in July. It's the index's second straight monthly decline and its weakest reading since January.

PCE Inflation, July data

Headline PCE: +0.2% for the month, +3.7% YoY — both above the 0.1%/3.6% economists expected.

Core PCE: +0.2% monthly, holding at 3.3% annually — unchanged from June, meaning the disinflation trend has stalled rather than reversed.

Both readings sit well above the Fed’s 2% target, with little forward momentum toward it.

Venezuela oil deal

Trump posted on Truth Social: “The United States of America has just entered into an Agreement with the Country of Venezuela on, THE BIGGEST OIL DEAL IN WORLD HISTORY!” He said it gives the US majority control of more than 65 billion barrels of Venezuela’s proven reserves, negotiated by Secretary of State Marco Rubio, Defense Secretary Hegseth, and Venezuela’s acting President Delcy Rodríguez. Context: this comes as US gas prices hit about $4.09/gallon (up 27% year-over-year) and the Strategic Petroleum Reserve fell below 300 million barrels - its lowest level in decades - partly due to the Iran war disrupting the Strait of Hormuz.

Fed Chair Kevin Warsh’s Jackson Hole speech

He said directly: “While this summer’s PCE and CPI readings were better than expected, they do not tell me that underlying trends have meaningfully improved.”

Markets reacted immediately - odds of a September rate hike (priced via CME FedWatch-style futures) jumped to around 50–60%, up from roughly 35% before the speech.

The market’s reaction was immediate and connected across assets:

• Rate expectations: odds of a September hike jumped from about 35% (Thursday) to 50–58% (Friday) on CME FedWatch-style pricing.

• Treasury yields & dollar: yields rose and the dollar strengthened roughly 1% on the day.

• Gold: fell as much as 3.2–3.4%, breaking below $4,500 for the first time in about a week, as the stronger dollar and higher real yields made a non-yielding asset less attractive.

• Bitcoin: dropped roughly 3%, pulling back from its attempted breakout above $80,000–81,500.

• Equities: small caps (Russell 2000) were hit hardest, down ~1.4% Friday, more sensitive to rate expectations than mega-cap tech; the S&P and Nasdaq fell more modestly since Nvidia’s earnings glow was still providing some support.

Chinese bots

X's Safety team identified roughly 200,000 accounts tied to a suspected Chinese bot farm, with about 200 of them found actively posting content designed to sway American opinion against AI data centers - including material on grid strain, rising utility prices, and cartoons portraying data-center operators as profiting at the public's expense.

The concern: this looks like an attempt by a foreign adversary to slow down U.S. AI development amid the broader US-China AI race. Congressional Republicans have already been probing similar influence campaigns - a House Energy and Commerce Committee letter in June flagged China, Russia, and Iran as suspects, citing one estimate that opposition campaigns have delayed or blocked about $45.8 billion in data center projects in Virginia alone. Separately, OpenAI said in June it had disrupted two Chinese-linked covert networks using ChatGPT to generate fake content on the same topic, though those campaigns reportedly didn't gain much traction.

Now, let’s move to tech and some good news, covered by .

NVIDIA delivers

Where else to begin but from NVIDIA. Since July, AI build out has faced some uncertainty. Many were wondering whether window of opportunity was closing. Then came Jensen with a sledge hammer.

He basically delivered an historic quarter, guided the company to hit $100B ($108B to be more precise) quarter in the next 3 months and ended the call by saying he is expecting “just” 70% growth next FY because of supplier shortage. He also quashed all fears of Jalapeno, Trainium or TPUs by suggesting below.

I regret not investing more in the 2 frontier models OpenAI and Anthropic. We are confident of being the bet out there despite custom ASICs - Jensen Huang

There were three caveats though:

• NVDA guided their operating margin down to 73-74% instead of their usual 80s% level.

• Future commitment of $366B over next 5 years with $279B going towards supply & capacity.

• Inventory value and account receivables is heading only in one direction and that’s not appeasing any critics.

Moving Updates from Hot Chips:

Hot chips is usually a boring, geeky conference where electronics afficionados head to talk shop. This year though, it had much of fanfare. Many moving updates came from the conference on how future of Memory and Optical communication will look like in the next leg of AI development. Let’s see some on high level

Memory:

Memory is the defining constraint right now in AI build out. HBM bandwidth is scaling >2x every 2 years against GPU bandwidth scaling 3x in same 2 years creating “Memory wall”. Two main approaches were touched upon in Hot chips

• zHBM: Samsung’s approach towards solving the problem is to stack DRAM on top of xPU. Completely eliminates the need for 2D interposer targeting a ~70% I/O power reduction through wafer-on-wafer and hybrid bonding. Likely with HBM5/5e in 2030. This model eases power delivery but has thermal constraints.

• Raptor’s 3D Dram: D-matrix unveiled another solution for the memory wall problem - Login on DRAM. Single-die accelerators are put on top of vertically stacked DRAMs facing the cold plate that improves thermal performance. It offers SRAM-class bandwidth, HBM-class capacity at 10% of HBM energy. For given power level, they offer 10x bandwidth as current HBMs. This model needs engineering to find a workaround for delivering power through DRAM.

• NAND: Interestingly enough HBF again was pushed as one possible solution. Samsung also announced zNAND-O. 4-8 NAND wafers are bonded directly on top of logic die to shrink latency and improve bandwidth for edge AI use cases with real-time inference.

One other general theme in all this how hybrid bonding becomes an essential aspect to make that next jump in 3D memory. This can be another lever which can push investments resulting in tailwind for WFEs like Lam research, AMAT.

The benefits of it best illustrated from SK hynix presentation in the event below. Hybrid bonding not only provides thicker core die but also narrows the TSV pitch. This results in improved thermal efficiency (resistance reduction).

Optical Communication:

• Purpose-built Silicon: As developments progress, one thing is getting obvious: Wide range of use cases like agentic AI models, pre-training, sampling have unique demands requiring unique specifications. For example, Training and Inference have different compute, memory and latency requirements. These are showing up on various architecture emerging. Purpose-built accelerators seem to be the next trend to watch closely.

• Redefinition of “Networking”: AI factories used to be built around a homogenous networking principle. Moving away from this seem inevitable. Now, each AI factory seem to have a combination of approaches. For example NVIDIA announced that AI factories need purpose build networks with a combination of scale-up, scale-in, scale-across, context scale and scale-out networks.

• Expanding Scale-out and Scale-Across: Google shared its Virgo network in which 134000 TPUs at 48 petabit per second were in a single cluster spread across multiple data center sites. Buildings will require substantial switching and optical infrastructure.

What SaaSpocalypse?

Remember not so long ago when all the software names were left for dead. People were preparing their eulogy speeches and wanted to see how to reminisce the good ol days of “rule of 40”. Here’s a fun fact for you, this is the return of some software names and semi names. In last 6 months NVIDIA has been the worst performer amongst the names listed.

In the last month:

• Salesforce: Agentforce annualized revenue $1.5B, up 240% YoY; Agentforce plus Data 360 ~$3.9B ARR. Launched Claudeforce with Anthropic

• ServiceNow: AI ACV crossed the $1B milestone; full-year AI revenue guide raised to $1.5B. Agentic deployments up 9x in nine months; AI Control Tower live at 500+ customers.

• Strong results also from Atlassian, Workday, cybersecurity names.

This is not the SaaSpocalypse... this nonsense of this SaaSpocalypse, I think it’s time for it to stop. - Marc Benioff (Chair, CEO, and Co-Founder of Salesforce)

Stop being the herd and giving in to maniacal doomsday fear mongering. Strong structural businesses don’t go out of business unless they stop reacting.

World Humanoid Robot Games

China’s five-day World Humanoid Robot Games highlighted just how quickly the country is advancing in robotics. More than 2,000 humanoids from 660+ mostly Chinese companies participated, reinforcing China’s position as the current manufacturing leader; it produced roughly 12,800 humanoids last year, around 90% of global output.

The hardware progress was particularly striking. The fastest robot completed the 100m more than twice as fast as last year, while the best standing jump improved from 0.96m to 3.40m, helped by advances in motors, torque and liquid cooling.

The main limitation is still intelligence rather than physical capability. Current robots can become very good at individual tasks such as running or boxing, but struggle to generalize across unfamiliar real-world situations. Unitree founder Wang Xingxing said humanoids are “not yet capable enough for mass deployment”, with AI models for decision-making and interaction remaining the key bottleneck.

So the takeaway is that China is making very rapid progress in humanoid hardware and manufacturing scale, but autonomy, dexterity and genuinely useful general-purpose robots remain significantly behind. The industry’s eventual “ChatGPT moment” for humanoids may therefore still be several years away.

The last part is covering reactions from FinX where a lot of interesting events took place.

What FinTwit was actually saying this week

Data Centers!

Gavin Baker fired off a blunt declaration: any politician standing against data centers has no business holding office. He called the facilities a revival engine for small-town America, a rare win for blue-collar workers, a force that actually lowers household power bills, and a quiet helper for cleaner energy that barely sips water. Attached was a Wall Street Journal screenshot showing unions now entering the fray, turning jobs into the newest flashpoint after months of fights over electricity, water, and noise. Baker then spent the afternoon stacking receipts—utility filings from Georgia, California, Indiana-Michigan, and Louisiana where regulators were being asked to cut residential rates precisely because data-center customers were covering more of the system costs.

Anthropic ARR numbers look softer than what bulls had hoped

A once-bullish fund PM flew into Hot Chips looking rattled. Downstream demand, not supply, was the scare: circulating Anthropic ARR numbers looked softer than bulls had hoped, enough to question the whole AI CapEx cycle.Yet every supplier he spoke to still sounded extremely bullish. Orders were stacking. Expansions continued.

@labubu_trader called it pre-IPO fog and midterm politics, not a structural break. Conflicting data, seasonality, and campaign-season noise around data-center power and water were muddying the picture. Until those two dates pass, sentiment stays weak. After them, the real demand signal should appear. If it holds, this stretch becomes the buying window.

IREN earnings reactions

As usual, divided reactions. Bulls defending it. Bears attacking it.

The result:

Thursday after the close IREN printed $137.2M revenue (inside a wide $132–157M range, so “miss” depended on which estimate you used). AI Cloud did $70.5M, doubled QoQ, and passed Bitcoin mining for the first time. They also showed $4B contracted ARR for 2026 capacity and $1B operating. GAAP looked ugly: ~$450M Q4 / ~$639M FY non-cash mining impairments, $703M full-year loss.

Friday the stock dropped ~13% and lost about $1.85B of market cap.

X split into two camps: “dilution machine / where’s the revenue / $25–30B capex” vs “customers and lenders fund ~90% of each MW, 2026 is sold out, impairments are the old fleet dying.”

Lastly

No one knows what hypervisor means, but bulls will probably explain that it’s something bullish.

Cybersecurity next?

Michael made an excellent trend, and while I agree with the overall importance of this sector, I think we’re a bit late to the party. The time to buy these stocks was a few months earlier.

Paypal - can they do anything right?

Marvell Beats Q2 Targets and Raises Outlook

The company reported $2.739 billion in fiscal Q2 2027 revenue, up 37% year over year, with data center sales surging 46% to $2.2 billion which is now 79% of total revenue. CEO Matt Murphy highlighted robust AI bookings, raising full-year guidance to $12 billion (45% growth) and fiscal 2028 to $18 billion (50% growth), with data center revenue expected to grow over 60% both years.

Despite beating estimates and guiding Q3 at $3.15 billion, shares fell 7-10% in after-hours trading amid high expectations from a Google partnership and prior gains. Analysts see the long-term targets as credible but note execution risks ahead of the October 6 Investor Day.

SaaS jump

Wednesday night CRM and CRWD reported with NVDA. Thursday the “SaaS is dead” trade flipped. Salesforce had its best day since 2020 (~+21–23%). CrowdStrike ripped ~+18–21% and printed a high. OKTA was cited around +28%. $IGV leadership posts flooded the timeline: software was the rotation, not just one stock.

What the prints actually said: CRM beat on revenue ($11.35B vs $11.32B) and crushed EPS ($5.90 vs $3.27). AI & Data ARR hit $3.9B (+210% YoY). CRWD did a clean double beat: $1.47B revenue (+26% YoY), $333M net new ARR (+51% YoY, a record), raised full-year net-new ARR growth.

$CRM best day since 2020, up nearly 21%, back above $250.

$SIVE… at the end (of the article)

Let’s start with the recap by who has been the most vocal Sivers bull on the platform, or at least the most well known.

His “friend” also commented. I’m sure you know Moody by now if you followed Sivers on X.

The ER was a disaster, not going to spend time on numbers. Bulls are saying it’s about the future. Still, the wording that I disliked the most is constant use of “opportunity pipeline” like it means something.

ParadisLabs commented on it too:

This was my own recap of FinX reactions on Sivers ER.

One thing is certain - it’ll take a lot of time and effort to bring $SIVE even remotely close to its ATH.

Recommended Reads:

Damnang’s Substack

HBM Density Peak: Why AI Value Is Moving From Memory to Connectivity

Rubin Ultra’s HBM configuration is being reviewed downward, from HBM4E 12-Hi toward HBM4 8-Hi. Supply constraint is the direct cause, but supply alone does not make the choice work. It works because a wider scale-up domain and software tiering absorb the loss of per-GPU memory capacity at the system level…

Read more

12 days ago · 59 likes · Damnang

Nutty

An Investor's Handbook IV - AI Inference (1)

This is an analysis of industry and technology structure, not investment advice. The author may hold positions in securities mentioned…

Read more

5 days ago · 12 likes · 2 comments · Nutty

Know Your HBF: How Big a Deal Is It?

AI, the Fed and Market Opportunities

Space-Based Data Centers

Citrini Research

Macro Memo: Regime Change

One of the first reports Citrini ever circulated was a thesis arguing that Silicon Valley Bank was technically insolvent and at serious risk of failing. In March 2023, SVB collapsed and the report (along with a handful of…

Read more

5 days ago · 291 likes · 7 comments · Citrini

Glocal Investor

GE Vernova: The Company that sells electricity's Picks & Shovels

A single large AI data centre draws as much power as a small city. We are currently in the midst of building data centers requiring 15GW of power over the next year. SpaceX alone has announced plans to build 10GW data center by 2027. One company from one country. Simply put, energy demand is parabolic…

Read more

9 days ago · 2 likes · Arvind Srinivas

To get a chance to be mentioned in next week’s article, share an interesting market/company/investing story that happened in the week by Saturday evening.

If you have a good article to recommend, your or someone else’s, please share in the comments.

Do not forget to check and sub to ’ Substack (link below).

Glocal Investor (Arvind Srinivas)

Do let us know if you enjoy these recaps!

* The “FinX bubble” refers to a small corner of X where the community tracks artificial intelligence, semiconductors, and infrastructure bottlenecks.

Iran and Hormuz

The US-Iran conflict is still going on, now months in.

Trump added fuel this week, calling the Strait of Hormuz “American territory” / “new US territory.” Reports alongside that framing point to roughly 16 million barrels of oil moving through the strait overnight, with traffic up nearly 400% in recent weeks (allegedly).

Never a boring day following Trump’s posts. He announced sanctions, and called it “Economic D-Day”. All caps, of course.

Iran’s response has been defiant as usual. Their Foreign Minister mocked the latest round of sanctions, dismissing what’s been called the “most crushing economic operation ever” as another effort that will fail like the ones before it. Other Iranian officials are warning of wider strikes if neighboring states cooperate with US economic pressure.

Macro, Bonds, Debt

US national debt officially crossed $40 trillion - the headline number everyone’s anchoring to this week. Trump’s response leans on growth as the fix rather than tax hikes.

The Treasury has also stepped up its bond buybacks with repurchasing some of its own previously issued debt before maturity. Normally this is routine housekeeping, as they want to keep the bond market liquid by cleaning up older, thinly-traded (”off-the-run”) issues. But reports this week say the program was increased, possibly doubled, and some are reading that less as maintenance and more as a cushion. This could be a sign that the Treasury is worried about liquidity or demand heading into upcoming auctions.

The pressure behind all of this is simple: debt doesn’t just sit there, it has to be rolled over. When existing government debt matures, Treasury doesn’t pay it off in cash - it issues new debt to replace it. With yields elevated, refinancing a large slice of the roughly $40 trillion pile at today’s rates means interest costs eat a growing share of the federal budget.

Markets & Crypto

Bitcoin ran hard. FinX noticed and it was trending, even annoying a lot of investors there.

From roughly $64k toward the $79k range. Bulls are framing it as more than a normal bounce: institutional flows, ETF activity, and a short squeeze all getting credit.

Part of this week's rally ties back to renewed hope on the CLARITY Act which is the bill that would finally give crypto clear US regulatory rules. It's not a final vote yet, just a procedural one set for September 15, but the mere signal that it's moving again after being stalled is enough to fuel optimism.

What FinTwit was actually saying this week

@ParadisLabs set the tone Saturday, arguing that only the highest-quality AI names are worth holding from here as “the easy money has been made”.

There was a lot of responses from the wider community, where top voices largely agreed with the sentiment.

My response was the following:

Optics as the next shortage

@aleabitoreddit flagged the optical interconnect layer as the next bottleneck.

AAOI ATM, from FinTwit’s side

This was the most discussed topic of the week.

recapped the story excellently.

AOI indeed needs cash to fund its expansion, and for now this is the way they’ll do it. My main critique was that it was done right after the Rosenblatt event where they shared a lot of bullish updates. Event might have been used as a mean to increase the price leading to the ATM, but no evidence of that of course.

Here’s my Rosenblatt recap of AOI’s discussing. A must read for AOI investors.

Positive news for Space sector as Trump announces “Golden Age of Space Transportation”

This policy updates the National Space Transportation Policy to dramatically scale up the country's spaceflight frequency, commercial partnerships, and deep-space infrastructure.

Burry’s updated shorts

Our favorite bear is not just short but also

Without the receipts this could be just a good PR stunt, but also 95% of FinX operates without receipts, so we have to lower our expectations.

$3 trillion in hidden AI liabilities

The WSJ dropped a story this week that deserves more attention than it’s gotten: nine of the biggest US tech companies — Alphabet, Meta, Microsoft, Amazon, Oracle, Nvidia, Broadcom, SpaceX, and AMD — have accumulated roughly $3 trillion in AI-infrastructure liabilities that don’t show up on their balance sheets, per WSJ’s analysis of their latest filings (full recap here). That’s already about three times these companies’ official on-balance-sheet debt, and it’s growing faster than reported capex.

Data center bans keep spreading

Texas remains the headline case, but it’s part of a much bigger national wave this week. Recap:

• Texas: Governor Abbott’s early-August directive ordering PUCT and ERCOT to audit every data center in the interconnection queue before approving new ones is still in effect and was reinforced this week

• Wisconsin: data center backlash is now an active campaign issue — GOP gubernatorial candidate Tom Tiffany is running ads branding his Democratic opponent “Data Center David Crowley”

• The bigger picture: a Governing piece this week (Aug 20) notes Maryland alone now has 13 county-level moratoriums - more than half its counties — and that opposition is genuinely bipartisan: Chicago’s progressive mayor and Texas’s Republican governor both called for a pause within weeks of each other. New York became the first state to impose a moratorium via executive order in July; Pennsylvania’s Shapiro signed an order requiring local approval before state permits; Louisville and Fort Worth both passed local moratoriums in the past two weeks.

Six chip headlines from the week, and what they mean

• Broadcom is in talks to raise $60–80B in debt (some reports say the total package could hit $100B) to fund AI chip capacity for Anthropic and others, via a special-purpose vehicle (SPV) — a separate legal entity that buys the chips and leases them to Anthropic, keeping the debt off both Broadcom’s and Anthropic’s balance sheets.

• Samsung’s board approved returning up to ~110 trillion won ($79.5B) to shareholders in 2026 — the largest payout in Korean corporate history, about five times its prior record — funded by the AI-driven memory boom (Q2 chip profit surged more than 250x YoY).

• Nvidia pushed back hard on a report claiming it planned to ship a China-specific “LPU” (a chip type built on licensed Groq tech) by year-end, saying flatly it has no such product on its roadmap. This lands just before Nvidia’s Aug 26 earnings.

• Japan’s trade ministry (METI) is requesting another ¥150B (~$944M) for state-backed chipmaker Rapidus in the FY2027 budget, bringing total government backing toward ¥400B+ (with total planned support closer to ¥3T over time), as Rapidus works toward 2nm mass production by 2027. Japan continues doubling down on Rapidus as a strategic bet to reduce reliance on TSMC, even though Rapidus remains roughly two years behind TSMC and Samsung on the same node.

• Zhongji Innolight posted H1 2026 net profit up 241.7% YoY to 13.65 billion yuan, on revenue up 182.5% to 41.78 billion yuan — driven by 800G/1.6T transceiver demand from global hyperscalers, with overseas revenue up 210%.

Damodaran’s piece

Aswath Damodaran published a long framework piece this week worth summarizing properly rather than skimming through it (full piece here).

His core argument: every revolutionary technology goes through four phases

• hope/hype,

• investing build-up,

• business building, and

• recalibration

and AI is now entering the business-building phase, where investors stop pricing on story alone and start demanding real evidence: revenue, margins, and defensible moats.

To get a chance to be mentioned in next week’s article, share an interesting story that happened in the week by Saturday evening.

Do let me know if you enjoy these recaps!

Q2 showed us that the ramp is real and is happening. AOI exited Q1 with roughly 100,000 units of monthly 800G and 1.6T capacity, reached around 200,000 by Q2 and maintained its plan to scale toward 650,000 units per month by year-end. They also hit record revenue, 800G started accelerating, the first major 1.6T order moved closer to shipment, and management insisted that demand was running ahead of what the company could physically produce.

Check my detailed recap of Q2 earnings in the article below:

Rosenblatt event was a different type of discussion. No earnings or numbers to justify. We could listen about AOI story and important topics of the bull thesis: InP lasers, 800G and 1.6T capacity, U.S. manufacturing, CPO, long-term customer agreements, gross margins, supply constraints and how AOI plans to fund one of the most aggressive manufacturing expansions in the optics industry.

If you’re new to AOI’s story, check my deep dive in the article below:

AOI at Rosenblatt improved the setup and the thesis.

• Customers are willing to pay a premium for U.S.-made product.

• AOI is sold out through at least the second half of 2027.

• Several long-term agreements are already being negotiated.

• The company already has 300–400mW CPO lasers in customer evaluation.

• Five companies are discussing CPO with AOI, some are already talking about NPO, and management believes standalone CPO laser chips could eventually carry gross margins above 60%.

Very important to add that management again reiterated their target of $471 million monthly data-center transceiver for 2027. According to Murry (CFO), the company could do even more if it had the capacity.

The customers are there. The technology is there. Now AOI has to build enough product. That’s the key part of the thesis.

The Rosenblatt discussion gave us a lot of details on how management intends to achieve their plans.

InP lasers

Murry (AOI CFO) said: “Technology-wise, the foundation of the company is in the lasers,” he said, and added that their customers “definitely appreciate the fact that we have our own laser fab because it represents a differentiated supply chain for them.”

Murry was also confident about the quality of AAOI’s laser technology relative to larger competitors. He said:

“From a design standpoint, from a specification standpoint, I think we’re as good or better than anybody else, arguably better in some ways.”

And also acknowledged that:

“Our production capacity currently is just a lot smaller.”

Management is not arguing that AOI has to catch up technologically. They believe the laser design and specifications are already competitive. The gap it is trying to close is manufacturing scale.

U.S. pricing premium

One of the strongest comments from the entire event came when Mury was asked whether hyperscalers actually pay more for products manufactured in the United States.

Murry’s answer was direct:

“They’re willing to pay a premium for U.S. production.”

This adds a lot more to AOI’s thesis than just saying that customers prefer domestic suppliers. If domestically manufactured 800G and 1.6T can have higher ASPs, then Texas expansion could potentially help AOI on three fronts at the same time: capacity, strategic positioning and margins.

Murry said the current geopolitical uncertainty around Chinese optical vendors is already influencing procurement decisions, even without a formal ban in place:

“It’s the uncertainty that’s driving their decision-making at this point.”

and added:

“AOI’s U.S.-based production capacity, currently existing and planned growth, becomes very attractive to them.”

We don’t need to make our bull case based on the Chinese transceiver ban. It does not need to happen. The geopolitical risk is already improving AOI’s negotiating position.

300–400mW CPO laser

One of the more important CPO updates from Rosenblatt’s discussion was that AOI already has the high-power laser technology required for these architectures. Murry said the company has 300–400mW lasers and is already sending them to customers:

“We are shipping small quantities for evaluation.”

and he also added that the limitation isn’t connected to the product readiness:

“There’s no problem with our ability to produce these products other than the fact that we just don’t have enough capacity right now.”

This is important for the thesis because it reduces the technology risk around CPO. AOI is not waiting to develop it. The product already exists and is in customer evaluation, and as usual in AOI’s story the management needs to solve the manufacturing scale.

Murry also talked about AOI’s laser performance and mentioned that it is already competitive with its bigger competitors:

“From a design standpoint, from a specification standpoint, I think we’re as good or better than anybody else, arguably better in some ways.”

AOI’s weakness today is just that their “production capacity currently is just a lot smaller.”

The CPO opportunity is increasingly a capacity story rather than an R&D story, which means that if AOI can expand laser output, it should be in a position to convert current product evaluations into a meaningful revenue in the future.

That is also why the second half of 2027 looks like the more important CPO window.

“Into the latter half of next year when we see our capacity increasing meaningfully enough… then I think that’s where we really will start to see our participation in that market start to grow.”

CPO may help margin improvement

The CPO opportunity is not only about increasing revenue. Its importance is a lot bigger when it comes to AOI margin.

Murry reiterated that the company still sees roughly 40% gross margin as achievable by the end of 2027, but he also said that CPO products could take the business well above that level. When asked about the economics of CPO, he said:

“Margins should be meaningfully above 40%.”

And on the laser side:

“Certainly on the laser chip, they’ll be most likely above 60%.”

This is important because 800G and 1.6T ramp is mainly about increasing revenue. CPO could help improve the quality of that revenue by bringing in products with higher gross margins, especially if AAOI is able to sell more standalone laser content rather than only complete transceiver modules.

Murry said AOI may only have around 32% to 33% gross margin in 2026, below the earlier 35% target because of the lower 100G mix, but he still maintained the long term 40% target and then pointed to CPO as a potential path towards it.

Five CPO customers, with NPO discussions already happening

Murry said AOI is currently talking with five companies about CPO, and some of those same customers are already discussing NPO.

Read more

Is Sandisk back in play? Let’s dig in to see what they showed to shareholders during the investor day.

“You are going to make 10-year investments and then you are going to hold a quarterly auction to see who buys your stuff and at what price. It just makes no sense. I think we are rapidly transitioning to something that is very different.” - David Goeckeler, Chairman & CEO of Sandisk

That quote probably explains Sandisk’s 2026 Investor Day better than any slide in the presentation.

Sandisk’s Investor Day was much more ambitious than a conventional semiconductor product update. Management was asking investors to reconsider what kind of company Sandisk is today.

Historically, as you probably heard many times by now, NAND flash has been one of the more cyclical parts of the semiconductor industry. Manufacturers increase capacity when demand and pricing are strong, and then when the supply eventually catches up, prices fall, margins compress, capital spending gets cut, and the cycle begins again. This cyclicality has made memory/storage companies very difficult to value through normal earnings multiples because earnings at the top and the bottom of the cycle are dramatically different.

Sandisk Investor Day aimed to show several structural changes that their business is going through.

AI build out is creating a new source of large-scale NAND demand across the globe. Today, Sandisk is moving a large piece of its business into multi year agreements with customers. On top of that technological improvements are allowing the company to grow the number of bits produced without requiring a proportional increase in manufacturing capacity.

Management believes these changes can produce something the NAND industry has historically struggled to deliver: sustained growth, high profitability and stable FCF.

2026 has already transformed the earnings profile

The starting point for this Investor Day is management’s showcase of what happened during FY2026.

Sandisk reported approximately $20.2B of full year revenue, representing roughly 175% YoY growth. Full-year non-GAAP gross margin reached 71.6%, compared with 30.3% in the previous year, while adjusted FCF reached $8.7B. By the Q4, annualized revenue was running at roughly $36B and annualized adjusted FCF at approximately $20 billion.

Those are extraordinary numbers for a NAND manufacturer.

I’m sure a lot of you are wondering if 2026 represents an unusually strong memory cycle or the beginning of something more durable. The Investor Day was trying to make the latter case.

Sandisk’s long term framework relies on three main goals:

• increasing profitability

• reducing cyclicality

• generating consistent revenue growth

Instead of isolating these goals, management believes that technology, customer contracts and capital efficiency reinforce each other.

The business model

Sandisk calls its new contractual structure the New Business Model (NBM). The basic idea is this: instead of renegotiating most of its business every quarter, Sandisk is moving a large portion of its supply into multi-year agreements with their key customers.

CFO Luis Visoso described why the old model was problematic:

“We are moving from a quarterly price negotiation to large multi-year engagement... Over that time, when we were operating in that model, we practically generated no shareholder value. Made capital investments super difficult because they were very risky.”

That is arguably the core of the entire Investor Day. Sandisk is a company which makes manufacturing decisions years in advance, but historically it had very little visibility into the economics attached to it.

Management said the average NBM contract now lasts more than four years, with the longest agreement at five years. They expect about half of FY2027 bits and about 2/3 of FY2028 bits to be covered under NBMs.

Sandisk disclosed $93.9B of total contract value across eight NBM customers at floor pricing, with $91.1 billion of remaining performance obligations. Three of those eight customers are U.S. hyperscalers, according to management.

They has also secured $16.5 billion of financial guarantees related to these arrangements, primarily through third-party financial institutions. This is how it works: customers secure their long term supply of NAND, Sandisk gets committed volumes, and their pricing is protected by floors, and the guarantees provide another layer of financial protection. For customers, this model offers supply certainty and for Sandisk it creates far greater revenue visibility and downside protection than their old quarter-to-quarter NAND model.

The most important quote from Investor Day

The obvious bear case is that these contracts sound good today because NAND pricing is strong but what happens when pricing falls? Visoso addressed that:

“Even at the lowest price, at the floor pricing, at any point in time, we do not expect these NVMs to be below 80%. We expect that to be around 80%... every single year over the period, we should be able to be even at floor pricing around 80%.”

This may be the most important statement from the entire event. Management is basically saying that the contracted part of the business should still generate approximately 80% gross margins even if pricing falls all the way to contractual floors. That is a completely different economic setup from the historical NAND models, where the standard sequence of cyclicality is relatively straightforward.

Sandisk is working hard to break away from that cyclicality. The company is not claiming that NAND prices will never decline again. Goeckeler was actually careful about that detail:

“We are not saying cyclicality is going away. The whole world is cyclical. The whole business is cyclical. What we are trying to do is get this wild cyclicality of this business, just dampen that down, get more sustained value creation.”

To summarize: the bull case is not that NAND is no longer cyclical, but instead it’s based on the case that Sandisk’s earnings may become much less cyclical than NAND pricing itself.

Supply discipline

One of the reasons memory industries usually become oversupplied is that the technology creates massive productivity improvements. Every new NAND generation can produce a lot more bits from the same wafer footprint, and those improvements often turned into more output which eventually contributed to oversupply.

Today, Sandisk seems to be aware of that problem. Goeckeler explained it.

“Through the application of innovation, we can grow output per wafer at a rate of 27% a year. Which, by the way, tells you, you can’t just release nodes whenever they’re available. Otherwise you’re going to flood the market with supply and you’re going to have another ’23 situation.”

That is a significant change in the company’s philosophy. Technology enables roughly 27% annual productivity growth, while management is targeting only mid-to-high-teens bit growth over longer periods. Sandisk does not want or have to translate every available productivity improvement immediately into market supply.

Goeckeler made the point again during Q&A: “We just need to not overrun the market.” This is very important because new contracts alone will not protect industry economics if every NAND producer aggressively expands its capacity.

More bits without proportionally more capital

Sandisk’s technology roadmap explains one of the most important aspects of Sandisk’s financial targets: low capital intensity. Sandisk’s BiCS8 architecture already provides improvements in density, bandwidth and power efficiency, while BiCS10 pushes those economics further. The company’s 1Tb TLC BiCS10 product is based on 332 layers and management says it improves memory density by 59% versus BiCS8 while also improving interface bandwidth, read/write performance and power efficiency.

Sandisk CTO Alper Ilkbahar also discussed the 2Tb QLC version, saying:

“Each BiCS10 2 terabit wafer has 65% more bits than the BiCS8 2 terabit.”

What does this mean? It means that Sandisk can increase the amount of storage from its manufacturing footprint without increasing physical wafer capacity at the same rate.

Goeckeler described it this way:

“Growth is primarily driven in this business by the application of intellectual capital, not financial capital.”

If growth increasingly comes from engineering rather than building new fabs, the conversion of incremental revenue into free cash flow can become very attractive. This helps Sandisk’s long term model enormously.

Guidance

We were provided with targets for FY2028 through FY2030. Sandisk expects average revenue growth in the mid-to-high teens, non-GAAP gross margin around 80%, non-GAAP operating margin around 75%, adjusted free cash flow margin of approximately 50%, and capital intensity in the mid-single-digit range.

In relation to guidance, Visoso said:

“When you aggregate all of that for 2028 through 2030, we expect to grow revenue mid to high teens, consistent with [bit] growth... We expect non-GAAP gross margin to be around 80%... And then you get to 50% adjusted FCF after paying for taxes, working capital, capital spending.”

“Revenue growth consistent with bit growth” matters. If revenue and bit shipments grow at similar rates, Sandisk is not building its FY2028-FY2030 framework around continuous NAND price increases. They are arguing that the business can maintain improved economics even after the current pricing environment normalizes.

Why AI may make that possible

Sandisk story works with the demand, without it, not so much. This is where AI enters the thesis.

Historically, NAND market relied heavily on consumer electronics like smartphones, pcs, removable storage. Today, the focus is shifting to data centers.

Based on management’s estimates, the NAND industry revenue could exceed $300B in 2026 and approach $500B year later (for context, historical average was close to $60B). These numbers do not represent the industry consensus, but rather management’s own views.

HBF could move Sandisk closer to the GPU

High Bandwidth Flash (HBF) is Sandisk’s attempt to create NAND architecture specifically suited to AI inference. The idea is not to replace HBM everywhere. HBM provides a lot of bandwidth but relatively limited capacity and comes at a high cost, while HBF is designed to offer dramatically more capacity close to the accelerator while retaining significantly higher bandwidth than traditional storage.

Sandisk has described several potential architectures, including using HBF alongside HBM or creating disaggregated systems where HBM operates as the high-speed tier while HBF holds much larger model weights or KV cache datasets. The most interesting slide from the technology section compared HBF with HBM in an internal AI inference simulation.

Sandisk claims one HBF-equipped GPU could run a workload that required eight HBM-equipped GPUs simply to fit the model, providing an 8x advantage in minimum-system capital efficiency. At higher throughput, four HBF-equipped GPUs generated similar token output to eight HBM-equipped GPUs, implying approximately 2x GPU efficiency.

CTO Alper Ilkbahar summarized the result:

“With four HBF GPUs, we are able to match the performance of eight HBM GPUs.”

Then he added:

“Your capital will go twice as far... This is fundamentally going to change the economics of AI.”

While it definitely looks impressive, it’s important to add that these are Sandisk’s internal simulations. HBF is still a technology in development, but there has been tangible progress. Sandisk says the first HBF memory die has now taped out and first inference product samples are expected in 2027.

Sandisk is also working to create an industry ecosystem around this technology, including its collaboration with SK hynix and the HBF Technology Consortium.

For Sandisk investors, it’s important to mention that HBF does not need to succeed for Sandisk thesis to work. The near/medium-term argument is NAND demand, NBM contracts and free cash flow. HBF is optionality on top.

If HBF works out, Sandisk potentially moves from being a storage supplier into a more strategically important position within the AI memory hierarchy.

Capital allocation

Sandisk’s capital allocation plan is very simple. They say they wants to invest in technology and the business, maintain a strong balance sheet with no debt, and return 100% of excess cash to shareholders.

Yes 100% of excess cash to shareholders.

Visoso: “There is no trick. It is 100% of the cash. Excess cash will go back to shareholders.”

Thit commitment from Sandisk is very relevant if the company comes close to its free cash flow targets. A business generating 50% adjusted FCF margins while requiring only mid-single-digit capital intensity would produce massive amounts of distributable cash.

Is this killing the NAND cycle?

It’s too early to say. These NBM contracts have not been tested through an industry downturn. Historically, memory LTAs have not always been durable as investors expected when supply and demand conditions changed dramatically. Prior memory cycles have seen customers finding ways out of LTAs, while additional industry capacity could begin returning from competitors in 2027.

This risk should not be ignored.

But Sandisk does not need to eliminate cyclicality for their strategy to work. They only need to reduce the damage caused by downturns. Compared with previous cycles, there are now real structural mechanisms attempting to achieve that: multi-year committed demand, pricing floors, financial guarantees, supply discipline, technology-driven bit growth, low capital intensity and a commitment to return excess cash.

This is what makes the current Sandisk story more interesting than simply saying NAND prices are high.

Goeckeler addressed the skeptics during the Q&A:

“There are the die-hard cycle people... ‘This is the way it has been for 20 years.’ I happen to have a belief you can change the future.”

Maybe. The next downturn will tell us more.

I hope you enjoyed the recap!

Leave a comment

If you’d like to see deep dive next, let know in the comments!

Full webcast can be found here.

Investor day presentation.

Earnings Season

I would understand bears if AI infra companies had no revenue, and were talking about dreams in 2035. But they’re not.

Earnings after earnings, companies delivered. A lot of them sold off, not because of the disappointing earnings, but because they simply ran too much, or analysts set unrealistic expectations.

One of the hottest earnings on August 6, 2026 was definitely $AAOI. They delivered. The story got stronger, not weaker. They can’t produce fast enough. AOI makes optical transceivers, the components that let AI data centers move data at the speeds these clusters need, and right now demand is outrunning their ability to build them. It’s about the ramp and mid 2027 transceiver output, and all earnings from now till then are temperature checks on whether that ramp stays on track.

Nebius is another name that sent a message loud and clear: AI is not dead.

This is almost a 100% increase in ACV (Annual Contract Value, the revenue customers have committed to), and it tells you a lot about how strong current demand really is.

• Revenue $582.3M vs. $573.9M est.

• Adj. EBITDA $236.2M vs. $172.6M est.

• ARR $3B (+58% QoQ)

2026 outlook:

• Year-end ARR of ~$7-9B (reaffirmed)

• Raised contracted power from >4GW to 5GW

The bears

Fintwit’s favorite allegedly shorted both Micron and Nebius, and even after Nebius’s incredible run, there’s a message circling around how he doubled down. Now, since he isn’t obliged to disclose his positions, I’m not convinced he actually shorted anything. This could be a move to get more PR, but so far the PR has been mostly AI made clown faces of Burry.

I truly believe he is a smart guy, but he picked one of the worst companies to short. Their financials have never looked better, and sentiment just improved once again. Shorting in July, I’d understand, but in August, you have to be either crazy, brave, or both.

Memory is dead, it’s optics’ time to shine!

Here is what started this “trend”.

Now, there’s a lot to unpack, and a lot of it was said in the comments. Jukan seems to have changed his position on memory, and the community reaction was strong.

I made a satirical post on the same day.

A lot of memory bulls and Fintwit creators turned bearish on memory (short term, they say), and cited optics as the short-term winner instead.

My message was clear: no one knows what will fly next. We can guess, but we can’t claim to know. Memory recovered well last week. Optics (apart from AAOI) didn’t do so great.

While I disagree with Jukan, I very much condemn all the insults and name-calling sent his way. People should be allowed to share opinions without being lynched.

Then came this tweet, which got retweeted hundreds of times within an hour.

It was another sign of retail sentiment flip-flopping on memory, even though nothing about the fundamentals actually changed for the worse in July.

Citadel…

I ran into an excellent article recapping July, and giving important information, stats, and updates regarding overall market performance.

Some highlights.

”While volumes normalized from the extremes of the prior two months, retail investor behavior changed materially over the course of July. As many of retail’s highest-conviction AI positions continued to weaken, investors increasingly shifted from buying weakness to reducing exposure. That transition accelerated into the final week of the month, culminating in what is now on pace to be the largest week of retail equity selling since 2022.”

This is one of the reasons most retail investors lose money. They buy the top, and sell the bottom. A lot of companies, including our dear Citadel, “helped” retail sell more by leaking false stories about rate hikes.

Now onto Earnings Season performance (I spoke about it at the beginning)

What bubble? Earnings are performing excellently, yet stocks are still being sold off. Not saying it’s a bad sign, but it was showing increased volatility in July.

“Stronger fundamentals have not been accompanied by multiple expansion. Instead, stronger earnings have coincided with a meaningful compression in valuations.”

Sounds familiar? Most noticeable with memory names.

In any case, we went through the most important tests: the FOMC decision and hyperscaler capex updates.

I loved the last sentence from Citadel’s report, as it said what I was saying

“The technical reset we have been waiting for has largely occurred. July did not change the structural bull market. It reset it.”

July was a bigger breather than most expected but the one we needed. Stocks can’t endlessly go up.

U.S. Urges Apple Not to Buy Chinese Memory Chips

This ties directly into the memory supply crunch we’ve been talking about all month. Howard Lutnick said the Trump admin opposes Apple buying memory chips from China, telling Apple this "plainly" after touring an Apple facility in Houston.

Apple is reportedly testing chips from CXMT and Yangtze Memory Technologies (YMTC) as it faces a severe memory shortage driven by AI data center demand, which has driven up prices and forced consumer electronics makers to raise prices.

Apple can legally buy off-the-shelf parts from the Chinese suppliers without a license, though customized chip work would require U.S. government approval. Apple COO Sabih Khan wouldn't confirm the Chinese chip testing but said the company has to "look at all options," noting memory isn't heavily customized compared to other iPhone components.

Low VIX

One thing that doesn’t get talked about enough right now: the VIX is sitting low, and that’s not the all-clear signal people treat it as. Low volatility doesn’t mean the market is calm, it usually means it’s compressed, and compressed things tend to snap rather than drift. Historically, some of the sharpest volatility spikes have come right after extended stretches of a quiet VIX, precisely because everyone gets comfortable at the same time and positions accordingly: short vol, max leverage, light hedging.

This isn’t a call for a crash. It’s a reminder that a low VIX measures recent calm, not future calm. After a month like July, with earnings, capex updates, and Fed decisions all still in play, this is not the time to relax.

Do let me know if you enjoy these recaps!

Welcome back to the Retail Investor Interview Series. Today I’m sharing the story and interview with KaizenInvestor.

If you follow AI infrastructure, space, or semiconductor investing on X, you’ve likely come across KaizenInvestor. He launched the account in September 2025 and has grown a following of over 42,000 by publishing full theses, specific add-levels, and transaction updates rather than just picks. In three years, he says he grew a personal portfolio from €33K to €180K, and he reports beating the S&P 500 for four consecutive years.

His story is interesting not because of the growth numbers, but because of what sits behind them: he is a Commercial Director at a 50,000-person multinational who first learned to build a thesis writing sell-side company reports as a student, then stepped away from serious research once the corporate career took over, and only came back to it once AI tools made deep due diligence achievable again inside a two-hour daily commute. He describes his approach as buying into structural growth “waves” - space, photonics, batteries, drones, memory, compute, lidar - with published add-levels he waits for rather than chases, on names like $PL, $RKLB, $OUST, and $MRVL.

I talked to him to dig into the story behind the numbers: how his research process actually works, what the past year has cost him personally, and where he sees it going from here.

Intro

How it all started

He first encountered investing during university, through a student job at investment banks writing company reports used to update price targets. That role taught him institutional algorithms, decision-making processes, and sector analysis - by his own account, he spent more time learning about companies and industries there than in formal classes. His early personal investments were conservative large caps: Microsoft, Holcim, L’Oréal. His career then moved from Finance into Purchasing, then Marketing/Sales, culminating in a recent promotion to Commercial Director - one of the youngest to hold that role at his company. Serious personal research went on a break once full-time work took over, since pre-AI due diligence was too time-intensive to sustain alongside the job. He resumed it once AI tools made that depth achievable again, launching the public X account in September 2025. Since then he has grown a following of over 42,000 by publishing full theses, specific add-levels, and transaction updates rather than just picks.

The philosophy

His framework centers on structural growth “waves” rather than momentum or broad sector calls. Those are currently Space, Photonics, Batteries, Drones, Memory, Compute, Lidar, and GLP-1-related names. Within that, he emphasizes AI infrastructure, the space economy, HBM/memory dynamics, optical and photonics connectivity, and selective lidar exposure. He’s valuation-aware but growth-oriented, using reverse DCF, analyst upside screens, free-cash-flow trends, and competitive-moat analysis to build individual theses. He publishes specific “add” price levels tied to portfolio allocation and stacks cash deliberately, waiting for those levels rather than chasing strength once a name is already moving. He reports beating the S&P 500 for four consecutive years, while also being transparent about drawdowns - including a 26% pullback from highs at one point.

The process

He works roughly 12-hour days and does the bulk of his research during about two hours of daily commuting, turning it into personal research podcasts (voice notes) rather than written notes. He relied heavily on Gemini for deep dives before switching primarily to Perplexity AI once token limits capped him at roughly one deep dive per day, and also uses NotebookLM. His stated view is that the modern investing edge comes less from finding information and more from asking the right questions. He recommends deliberately reading opposing viewpoints and following high-effort researchers while still doing independent work, treating AI as a way to build conviction faster than previous generations of retail investors could.

Interview

— THE PERSON —

Q: Who is KaizenInvestor outside of investing?

Answer: A hardworking husband and father. Work and family take up almost all of my time. In my spare time I often go for a run or a walk with our dog. I used to train for a triathlon as well, but with the little one joining my life three months ago, I had to cancel these plans. I used to play football at a semi-professional level until I tore my ACL; now I just play on a low-level team with friends. We don’t even train, and I’m already happy if I make it to one game a month. I try to read one book a month as well, so that’s a goal of 12 books a year. But it’s not evenly spread out — I usually get through around seven books from January to November, and then have to read five more in December. We always go on holiday in December, so that’s when I catch up. I still hit my goal of 12 a year, just not month by month.

I’m always doing something, I can’t stand the thought of standing still. That’s one of the reasons I don’t watch TV any more. I always get nervous because I feel like I’m wasting my time. I do watch some football on my iPad while working in the evenings, but that’s basically it.

I might be a contrarian to all the financial-freedom guys, but I love corporate life. The pressure it gives me makes me feel alive. I always needed pressure to perform; without it I tend to get bored quickly. I think I’m doing a great job, as I was recently promoted onto the board of two legal entities of the company as Commercial Director/Manager, and I’m learning a lot of new things.

Learning has always been my number one priority — that’s also the main reason I started this account and named it Kaizen, to learn and interact with the community. Kaizen is a philosophy where you try to make yourself 1% better each day, and that’s exactly what I’m trying to accomplish.

The main thing I can do better on right now is organizing things for my wife. We had a challenge last year (before the kid) that I would organize 10 dates that year. I finished with three. I’m very lucky she loves me for who I am and makes sure I’m able to do all this while she takes care of me and the baby. She organizes our dates, weekends, holidays... I just have to show up. I’m a very lucky man to have her in my life.

Question: What does your life actually look like right now — between the Commercial Director role, a newborn, and a research habit that eats two hours of every commute, what does an actual day look like, start to finish?

Answer: Busy — I wake up between 5 and 6 AM, depending on the little one. If he wakes up at 5 AM, feeding him is the first thing on my morning agenda. Around 5:30 AM, I go for a walk with my dog. It’s my way to really start the day. We live near a couple of big parks, which are almost empty in the morning. During those walks, I prepare myself for the day: I check my emails — since I work for a multinational, emails from Asia and North and South America come in overnight — and I prepare my meetings and check what’s on my schedule for that day. I’m back between 6:15 and 6:30 AM, take a shower, and leave for work. It’s a one-hour drive for me, and I often listen to a self-made AI podcast about a stock I’m investigating, or have a discussion with an LLM on a work-related or investing-related topic.

I arrive a little before 8:00 AM at work and have an hour to do my emails. Since I’m in management, from 9:00 AM until 5:00 PM you can’t get anything done — you’re constantly in meetings or on the phone with one of your team members. After those meetings, I can start working again, so I head home around 7:00 PM.

I’m lucky my wife always makes dinner, so we eat around 8:00 PM. Depending on how much work I still have, I start some research after dinner. If I have too much work, I’ll just work a little. So from 8:30 to 10:30 PM I do research, work, or go for a run — it really depends on how much work I have, whether there are interesting earnings, etc. I never watch TV, so that helps save me some time. At 10:30 PM I feed the baby again, and at 11:00 PM I go for a small walk with the dog again. I try to be in bed around 11:30 PM.

Question: You’re a Commercial Director at a multinational with roughly 50,000 employees — one of the youngest to hold that title there. As an economist leading an engineering team, you’ve said your edge is “asking the right questions” rather than matching technical depth. How do you find the time and headspace to run an institutional-quality research process on top of that job?

Answer: Well, mainly AI, to be honest. I let AI do some research for me and turn that research into a podcast via NotebookLM. I listen to the podcast in the car and discuss it further with Gemini or Perplexity while driving. It’s also often in the car that I draft my posts for X. After a discussion with the LLM about upcoming earnings, for example, I dictate my post about it. I just have to check it when I’m home, tweak it a little, and post it.

On the headspace side, I get bored pretty quickly. I need to do a lot of different things in a day. When I leave the office, I often find myself with seven tasks still open and unfinished. Investing is just a natural interest for me — I like doing the research and learning about new companies. I also think it helps in my current role, knowing different company cultures and business models: what’s hot right now, and which sectors are struggling.

I love connecting the dots, and the more companies and business models you know, the more dots you can connect. I bought Planet Labs not because I knew every detail of the company, but because it connected the right dots with what I thought was the future for AI and space. I never go into too much depth — I just like to see the bigger picture.

Question: You became a father this year while carrying a demanding new promotion at the same time. You’ve written about it candidly on X — an emergency C-section after the baby’s heart rate dropped, your wife on strict three-month bed rest, and averaging just over three hours of sleep a night for the first two months while handling night shifts yourself. What made you decide to post about that publicly, on an account people mostly follow for stock research?

Answer: It’s still a social media account, and people can get my full story there. It honestly also helped me end a difficult period — I got a lot of messages from people who could relate and had a similar story. My main goal is to help people become better investors, and investing is part of life. You go through certain periods and emotions that all affect your investing journey. I think it’s part of what my account is about: letting investors know that everyone has difficult periods in life — it’s how you react to them that defines you.

Question: You took only five of your 25 paternity leave days because of the new promotion. Looking back now, would you make that same trade again?

Answer: I would now say I’d take them all. It was a very difficult period for me and my wife. I need energy and sleep to live the life I want to live. Without sleep I react differently to situations of pressure, which I normally live for. If my wife had been physically fine, I don’t think it would have been a major problem, but that wasn’t the case. I’m happy we got through it, and my wife is feeling better now. Let me phrase it differently: I’d recommend everyone take all their paternity days, to help your wife and rest a bit yourself. I tend to be a slow learner on these things myself.

Question: Outside of work, fatherhood, and the account — what do you actually do to switch off? You’ve mentioned Tomorrowland is basically on your doorstep — has becoming a father changed how you engage with that, or with anything else you used to do purely for yourself?

Answer: Like I said, I go for walks with my dogs or go for a run. I never run alone though — many of my friends run, so we try to schedule it together. Like I mentioned, before the little one I used to train for a triathlon with four of my friends. It was a month after he was born, and I underestimated the impact of that a little. So I wasn’t able to actually do the triathlon, another one of us got injured, and only two finished the race. I still play some football on a lower-league team with friends, and play some occasional padel and tennis.

I used to game a bit, but since I’m very competitive, I get addicted pretty quickly. I want to become the best version of myself, and with work that’s just not possible. I also don’t feel relaxed after gaming — I just want more and want to get better. That’s not something I have with sports; I want to win and will do everything to win, but my mind is empty after an hour of sport.

As for Tomorrowland, I used to go out every weekend, but that time is gone. I didn’t drink for three months while the little one was on the way, so I seriously changed on that front. I used to game a little when I was a student, but that changed once I started my career.

— THE ORIGIN —

Question: You got your start writing company reports for investment banks as a student, before you had any personal capital behind you. What did that job actually teach you — about how institutional analysts build and revise a thesis — that a retail investor coming up without it would never pick up on their own?

Answer: It obviously taught me a lot. I saw things from a completely different perspective. I saw how they built up reports and how the argumentation worked. I followed my first earnings calls and that’s how I got into investing. It’s like driving a car — you have to concentrate on a lot of different things, and that’s difficult in the beginning. But once you’ve practiced a lot, it moves to another part of your brain and you do it on autopilot.

The thing that struck me most is how fast-changing the environment is. My mentor is one of the most respected analysts in Belgium. He worked for funds in London and Frankfurt and started his own company because he had a fantastic prediction model. Yet he was short big tech. According to his model, it was all overvalued and a crash had to be coming. As we all know, it kept growing, and he was wrong on that call. The models of yesterday aren’t built for the models of tomorrow, and we have to keep innovating constantly to keep up.

That’s something I try to implement in everything I do. Standing still is moving backwards. The models that work today won’t work tomorrow. Everything is moving so fast, and we have to move with it. You can’t call every big wave a bubble just because it doesn’t fit your old model. I work in a fast-moving environment, and I mainly invest in fast-moving industries. That’s what I like most — keeping track of all the dots and making sure we can connect them.

Question: Your early personal investments were conservative large caps — Microsoft, Holcim, L’Oréal. What did those first years actually look like, and were there early mistakes before you found a process that worked?

Answer: I started with paper money, since I didn’t have any real money to invest. Times were also different — if I wanted to buy shares, it would have cost me around €35 in fees. So you couldn’t invest with €50, for example, since more than half would go to costs. That’s basically why the perception exists that investing in the stock market is only for the rich. I’m happy that’s changed.

But I made a lot of mistakes. I thought every company that dropped 60% was a good buy, because I assumed they’d reach a new all-time high eventually. I bought names I didn’t even understand, just because they’d dropped 60%. There was no AI to do deep research back then, so that made it a bit more difficult, and I made some awful decisions in those days — luckily most of it with paper money.

I kept all my trades in a spreadsheet I still have, and sometimes I force myself to look back at my old habits: what did I do back then, and why? Sometimes I don’t think I knew what I was doing, but that’s fine. I think it’s in the book Atomic Habits where they say you only know if you’re good at something once you’ve done it for 8,000 hours. You don’t just need to spend 8,000 hours on something — you also need to learn from your mistakes. Most of the time, learning from your mistakes isn’t that difficult; there’s a right way to do it. That’s different in investing: you never really know what the right decision was, which makes the learning process a lot more challenging.

Question: You’ve said serious research went dormant once the day job took over, because deep due diligence was too time-intensive without help. What did that gap actually look like day to day, and what was the specific moment that made September 2025 the time to start over, publicly, in higher-risk names?

Answer: I started my X account in September 2025, but I started taking investing seriously back in 2022. It was harder to do research back then, but I did some, mainly by reading other people’s research and trying to draw my own conclusions.

I started in 2022 with some quality companies: Holcim, Microsoft, Visa, Kinsale Capital, Adobe, Ulta Beauty, and Nu Holdings are all examples of companies I invested in. During my paper-money days I’d invested in even safer companies, but I felt ready to take it to the next level. I mainly used a reverse DCF to check how cheap these companies were relative to their historical FCF growth. I got some great results with it — I did beat the S&P 500 that way — so I was pretty happy with it, since I couldn’t do a lot of research otherwise. I automated my reverse DCF tool in Excel so I could run hundreds of reverse DCFs in about an hour, but that was basically it.

With just a reverse DCF, all high-growth companies look expensive. It was once I could do more research via AI that I could confidently move from my value companies into high-growth companies. In my opinion, I’d learned enough to make that step. I built a first framework for making investment decisions, which I perfected along the way. That’s what leaves me now with some higher-risk names.

I feel that most young investors get dragged into high-risk stocks too early. It’s like surfing the biggest waves without knowing how to stand on a board yet — they fall hard and may never recover from the experience. I started with very small waves to learn the ins and outs of investing.

Question: Was there a genuine blow-up along the way — a position or stretch that actually hurt and changed how you think about risk — or has the path been mostly steady compounding since you found your framework?

Answer: I started after the COVID crash, so the tariff sell-off is really the only one I’ve been through. During that period I shifted most of my value names into my high-growth names — a move that made me a lot of money. Crashes are always a great time to make changes to your portfolio. But my main way of handling risk is diversification. You’ll never see me holding only five companies, and I’m also against buying more than three names in a given industry.

Last year I earned as much from investing as from my regular income. That also means I always keep some cash on the side to invest when the market is down. I don’t DCA — I mainly wait for dips in the market. I’ve been in the game for quite a while now, and I’ve been through multiple dips where investors could have added.

Question: How did you actually learn the rest of it — was there a mentor, specific investors you followed, books, or was it mostly trial and error once you started applying the institutional habits from the student job to your own money?

Answer: I don’t really have a mentor, but I try to read at least one book a month — or let’s say 12 books a year. I read most of them on holiday. But I think I’ve learned the most on the job. Especially now, we have a lot of management training on decision-making under pressure, which is basically investing: you have to make decisions under pressure, figure out how to handle them, and learn from them.

I’m always eager to learn, but it’s not that easy to learn in investing. You have to make thousands of decisions a day, and there’s no single right decision. In gaming, for example, you can review your games and look at the decisions you made under pressure. If you review them with someone better than you at the game, they can show you the right decision. You can train that, so that under pressure the right decision becomes your reflex.

How do you do that in investing, where there is no right decision and personality is a big part of your decision-making process? You have to build a personal framework. It’s a fairly complex process, but once you’ve finalized it, you know exactly what to do under pressure and how to handle certain situations. Like in gaming, it’s not always easy to follow it under pressure, since you don’t have time to review the framework in the moment — but at least you can review your decisions afterward.

Question: You bought Planet Labs at $4, right as Trump’s tariff announcement hit the market — you’ve called that period “a nice candy shop” for investors who didn’t panic. Walk me through that specific buy: what made you pull the trigger while the stock was still falling, rather than waiting for it to bottom?

Answer: Haha, I don’t remember calling it a nice candy shop, but like I mentioned, a crash is a good time to review your portfolio. I felt ready for the next step, going from a value investor to a high-growth investor. The crash was my moment to go. My value stocks had fallen a lot less than the high-growth stocks, so the risk-reward was ideal for taking that step.

During the crash I was pretty patient as well. Planet dropped 80% and bottomed around $3, if I’m not mistaken. I believe you can’t time the market, so I don’t try to. When I bought Planet, it had just crossed back above the 21SMA, so I hoped the uptrend had started. But with a single Trump tweet, it dropped to $3. I was still very confident in my buy though — it was easy to hold because I’d done my research.

I’m always more aggressive with stocks and industries I don’t yet have in my portfolio, and more patient with stocks I already own. I really wanted an entry into the space industry, and Planet and Rocket Lab looked like ideal candidates — they connected all the dots for me. Looking back, choosing those two was the right decision. AST SpaceMobile was also a good one, but I didn’t have experience with companies that had no revenue yet. In the end it’s all about the good experiences you build, creating a framework around them, and sticking to that plan.

Question: Your original Planet Labs thesis was partly inspired by a small AI project you were working on at your day job — the idea that a model is only as good as the data behind it. How often does something from the day job actually leak into a stock thesis like that?

Answer: We knew AI was coming, and we were working hard on our data because we knew data would play a major part in AI. When you have bad data, AI gives you bad outcomes. Because I understood the importance of data, I looked for a company that owned a lot of well-structured data in a fast-growing industry. Planet Labs connected all the dots, and I pulled the trigger.

I see things across a lot of different industries, so it would be foolish not to use that advantage. Peter Lynch told us to “invest in what you know,” so I try to know about as many industries as possible. I won’t name companies, but I visit a lot of production sites and speak with a lot of people at some of the biggest names on Wall Street. It’s a big advantage.

— THE PROCESS —

Question: Walk me through what a research day actually looks like. You’ve said you turn roughly two hours of commuting into a personal research podcast — is that a monologue, a structured set of questions you run through, live modeling out loud? What does it actually sound like?

Answer: I have a small course on X about how to do it. It’s very easy to do via NotebookLM. I first run a deep dive on a company — it’s always the same prompt. Gemini or Perplexity runs it, I upload the output into NotebookLM, and it turns it into a podcast: two people discussing the company. It’s not enough to know everything about the company, but it’s a good start. After that, I continue the discussion with my LLM in the car to go a bit deeper.

Like I said, I don’t want to know everything about a company — I want to know the main things. Investing is about how you see the world in five years and which companies will play a role in that world. You don’t need to know every detail to know whether a company will play a big role in how the world looks in a couple of years; you just need to be able to connect the dots.

I’ve listened to the podcast for over 100 companies now. I really like the approach, and I think it gives investors an edge. You can always go deeper if you find the company interesting.

Question: You moved from Gemini to Perplexity once token limits capped you at around one deep dive a day, and you also use NotebookLM. What does that tool stack look like today, and where does each one’s job start and stop?

Answer: I honestly only just started working with Perplexity, so it’s too early to say. I used to do everything with Gemini.

Question: You’ve said the modern edge comes less from finding information and more from asking the right questions — of AI and of companies. What does a genuinely good question actually look like when you’re trying to build or break a thesis, rather than just confirm it?

Answer: I think it’s all about having an open mind and a genuine interest in the other side of the story. If you’re bullish, learn why other people are bearish. The problem most people have is that they only talk to people with the same views. You’re in a certain group around a stock you hold on X, for example, and you’re all bullish, so you miss certain points you should know.

You have to know the thesis to ask good questions. If you’re in a face-to-face discussion with someone, always ask open questions. But the real secret to asking good questions is listening. If you know your thesis and listen very carefully to a bearish one, you can ask very good questions to form your own opinion.

The problem people have now is that they don’t want to know the whole thesis — they only want the conclusion. That’s the whole reason TikTok is so popular, an app I don’t have. In five seconds you can’t defend a whole thesis; you can just shout a conclusion, buy or sell that stock. How can you form your own opinion, or ask great questions, when all you see are videos like that?

Question: What actually triggers you to add to a position versus waiting, in concrete terms — not just “conviction,” but the specific signal that tells you a name has hit your level?

Answer: Even with high-growth names, I use a reverse DCF. It tells me exactly how much free cash flow a company should generate in 10 years — is that realistic, do I think the company can actually achieve that much cash flow in 10 years?

To judge that, you have to study the company’s total addressable market, its market share, and its competitive edge. I make decisions based on that research.

I prefer to buy in an uptrend, so once a stock reaches a level I’m interested in, I wait until it’s back above the 21SMA before acting. You don’t have to predict the turn — you can just follow the trend. You don’t need to make it more complicated than it already is. There’s always a chance a stock above the 21SMA moves back below it, but I think over the long run it’s a good indicator of a stock’s trend.

Question: How do you separate “the market is wrong about this” from “I’m wrong about this” when a position moves against you?

Answer: With an open mindset, you don’t really think in terms of right and wrong — you think about what’s best for your portfolio. I own stocks that I believe will outperform the market over the next 10 years. A downturn of a couple of months doesn’t make that right or wrong. In the end the market is always right, but you need to give it time.

Take Planet Labs, for example. I bought at $4 and it dropped to $3 — according to the market, I was wrong. Then it jumped to $20, and by that same logic I was right. Then it dropped back to $12, so I was wrong again... you know how the story goes. If you think in terms of right and wrong, you’ll just end up following the market. I try not to do that.

You just have to make sure every decision you make is the best possible decision you could have made at the time. What I mean is: once you have your framework, you gather the information you need to confirm a stock is a good fit for your portfolio. Once you have that information and it all checks out, you pull the trigger. You followed the framework and got the necessary information — that’s the best possible decision you could make. If the market later turns against you because of new information, you weren’t wrong. You followed your framework, got the information, and at that time it was the best possible decision you could have made. No regrets.

Question: How much of your process is quantitative — reverse DCF, analyst upside, free cash flow — versus a qualitative read on management, moat, and market psychology?

Answer: I’m a big fan of market psychology and decision-making. It’s all about who you are and how your portfolio corresponds to that. It all starts with my personal framework — I know what my portfolio needs and what I’m looking for. I mostly think in terms of industries, like whether I need something in the space industry, for example. I don’t go too deep at that stage.

When I identify something my portfolio is lacking, I move deeper into the industry: who are the companies, what are the applications, which companies are the problem solvers, how is the TAM growing?

Only then do I look at specific companies. The qualitative aspects make me want to buy a company; the quantitative part tells me when. I treat trend as the one qualitative-technical aspect, using the 21SMA for that. I find companies interesting when they’re below the 200SMA and above the 21SMA — that shows they’re technically undervalued and in a positive trend. As I said, I only do this for companies that are already attractive from a fundamental and growth perspective.

Question: In “The AI Race to the Bottom,” you write about something you call the “Locksmith Paradox” — the idea that we treat time as equivalent to money, and agentic AI is only going to intensify that. Can you unpack that concept, and how it actually shapes which AI infrastructure names you own versus pass on?

Answer: The locksmith paradox came out of an interesting workshop we had at work. Ten years ago, if you locked yourself out, it took a locksmith two hours to open your door. Now it only takes 10 minutes. You get the same value, but the person puts in less time and effort — so is it worth as much money as it was 10 years ago?

I’m not too worried about that as an investor. I think a lot of companies are wrestling with this paradox as part of figuring out how to value their AI solutions — do you sell cloud, tokens, a software solution? You see many companies experimenting with their business models to extract the most value from AI. It’s interesting to follow.

— THE PORTFOLIO —

Question: Walk me through today’s actual allocation — main positions and weights.

Answer: I update my portfolio every month and I’m fully transparent about my buys and sells on X — people can follow exactly what I do, with no paywall. For those who don’t know: my biggest holding is Planet Labs, followed by Palantir, Ouster, Rocket Lab, Marvell, Google, ASM International, SK Hynix, Hims, Filtronic, Fluence Energy, Bloom Energy, Iren, Amprius Energy, and Wolfspeed.

Question: You publish specific add-levels tied to portfolio allocation — $RKLB at $60, $MRVL at $120, $PLTR at $90, $OUST at $20, SK Hynix at €850, and others. Walk me through how one of those numbers is actually derived, from thesis to specific price.

Answer: I’ve followed these stocks for a long time, so I know them pretty well. The levels mainly come from a reverse DCF analysis, showing where I think they offer a lot of value at certain price points. I also use asymmetric investing as a tool — I made a post on AST SpaceMobile explaining how I actually do that. For most of these holdings, calculating the downside is far more important than the upside. If everything goes right, a lot of companies have sky-high upside, but what if it goes wrong? How much cash do they have? What are the liabilities, etc.?

Question: How do you actually size a new position from scratch — what determines whether something starts small versus becomes a much larger allocation?

Answer: I almost always start with a 5% allocation. The upside of this is that when a company does well, its allocation grows and it becomes more important in your portfolio. When a company underperforms, it becomes less important. This is great for a long-term investor, because your top performers — the ones that keep performing well over the long term — are proven, while your worst performers matter less.

When I really believe the market is mispricing certain companies, I add to them so the allocation gets back to 5%. If a company drops to a 2.5% allocation, I might add to bring it back up to 5%.

Question: You’ve described stacking cash deliberately and waiting for those levels. What’s your actual cash management philosophy — do you hold cash as a deliberate tool, or does it always feel like a cost while you wait?

Answer: You make the best decisions when you have options. The last thing you want is to be forced into a decision. I don’t see cash as an opportunity cost — it’s my way of keeping things in my own hands and controlling how I engage with the market. I decide when I add, not the market.

During the tariff dip, I went from a 20% cash allocation to 0% cash. When I believe it’s time to add, I can go all in. When I believe it’s time to wait, I’m happy to build up to a 25% cash allocation. I never go above 25% cash though — I like to stay invested most of the time. You can never predict what the market will do next; you can only be prepared.

Question: $OUST has been framed as a multi-year, patient position rather than a quick trade. What’s the actual case there, and what would be the signal that the thesis had broken rather than just gone quiet?

Answer: I think Q3 will be a very important quarter for Ouster. I’ve held Ouster longer than most of the new Ouster bulls who’ve popped up recently, yet I see myself as more realistic than most of them. In my opinion, they don’t have a real competitive edge against Chinese LiDAR. The main thing here is that Chinese LiDAR is about to be banned in the US, and Ouster will be the main beneficiary of that ban. I value them around $40 at the moment, which I think is fair value right now. If they capture more market share than I expect in Q3, my price target might go up though.

The only thing that breaks this thesis is if American companies find a way around the import ban. I don’t really think that’s possible at the moment, but those big companies are smart about finding workarounds. It’s possible, for example, that they stockpile inventory just before the ban starts — that wouldn’t be positive for Ouster.

Question: Filtronic is a much smaller, more single-customer-dependent bet than something like Planet Labs — you’ve framed it explicitly as a chokepoint for SpaceX’s Stargate project. How do you size conviction differently when the thesis rests that heavily on one customer relationship?

Answer: I don’t really size positions by conviction — I don’t have more conviction in Planet than I have in Wolfspeed. I think Filtronic is a very interesting company with a bottleneck technology. At the moment only SpaceX depends on this technology, but other players are exploring new communication methods. Filtronic could become the go-to company for companies like AST SpaceMobile, Amazon (Project Kuiper), and Blue Origin, for example.

I bought Filtronic for the future technology, not the current customer base — they could just as easily have had no customers at all right now. It doesn’t really matter at this point in time; I think they have some great technology, and the space industry is waiting on new communications innovations. The Ka-band is essentially full at the moment, and we need a way to solve that. Filtronic is ideally positioned for it.

Question: Roughly 30% of your portfolio sits in space names — Planet Labs, Rocket Lab, Filtronic. Does a SpaceX IPO actually help that basket, or does it risk pulling capital and attention away from the smaller names you already own?

Answer: Well, that’s something we don’t know for sure. A SpaceX IPO would be the first of its kind, so we can only guess what happens. At the moment, space is going through a very rough period, so I’d guess it’s mainly pulling capital away right now. But that could change overnight.

I think it’s a good time to invest in space companies. They’re delivering strong earnings, and it’s one of the fastest-growing TAMs out there. Companies like Rocket Lab, ASTS, and Planet Labs are also floating between the 21SMA and the 200SMA, which is technically an ideal buying zone.

Question: Planet Labs hit an ATH of $51, then dropped 37% after management filed a $1.5B ATM despite a healthy balance sheet. How did you actually react in the moment, and what does that call say about how you read a management team’s style versus just their numbers?

Answer: They have a very healthy balance sheet, and I thought they’d be able to handle their upcoming projects with the cash they already had. They have some interesting AI projects at the moment, but we haven’t had any comments yet on where that capital allocation is actually going. I trust Planet’s management — they obviously know more than we do, so in cases like this it’s more about trust than anything else.

I think the timing was a bit off, since it came right after an earnings call that was actually great. We now have to wait for the next earnings call to get more information on where the money is going. I still think Planet Labs’ future is big, since they’re combining AI with space.

Question: You’ve been transparent about a 26% drawdown from highs at one point. What’s your internal test for “this is a normal drawdown in something I still believe in” versus “I was wrong and need to exit”?

Answer: When the whole market is going down, I’m not worried. I don’t look too much at the price of companies I own — more at their allocation within my portfolio. I bought heavily into the tariff drawdown, so I wasn’t affected by it too much.

I’m also lucky to have a great salary at the moment — according to Gemini, I’m in the top 1% in Belgium, though I don’t know how representative that is. We bought a house, had our first child, and got married last year, so a lot of big expenses are behind us now — I’ve only been able to invest around €35K so far. I’m sure I’ll be able to invest more in the coming years, since I won’t have those big expenses any more.

I’m very relaxed about my financial situation, which makes it easier not to worry about dips. Ninety-nine percent of the time, a dip creates a buying opportunity if you wait long enough, and being relaxed about your finances helps with that. I’m sure I’ll have 1 million invested one day, since I’m able to invest a good chunk of my income and I’ve beaten the S&P 500 for four consecutive years now. I hope to get there around age 35.

Question: How do you think about correlation across your top names — space, memory, compute, and lidar exposure can all move on the same macro or rate narrative. Does that worry you, or do you see them as genuinely differentiated bets?

Answer: Someone who thinks these are similar bets is being too short-term minded. Yes, it’s all technology- or AI-related, but as you’ve read, my first mentor was against tech, and look where that got him. He saw too little difference between Alphabet, Amazon, and Microsoft — in his view they were all competing for the same thing. I believe I’m well differentiated across the industries that will grow the most over the next 10 years.

Question: Do you set stop-losses or hard downside limits, or is every position held purely on thesis until the thesis breaks?

Answer: I never use stop-losses or hard limits. I want to feel in control of the situation. When I use a stop-loss, I feel like I’m not in control any more — I know that’s not entirely true, but it’s a feeling I have. I’ve also had some bad experiences with them, or the reverse, where I wanted to sell but the stock barely missed my sell price. Instead of just selling at $199, I set a limit order to sell at $200, but the stock never reached it and fell to $130 instead. Now I just sell when I want to. Like I said, investing is all about experience — make sure you repeat the good ones.

Question: How often do you actually rebalance or reassess the whole portfolio top-down, versus just reacting position by position?

Answer: I always have my whole portfolio in mind — I never look at a single position in isolation. You should always start with the portfolio, like a football coach who buys players to fit his system. I don’t buy a top player and then change my system around him. I might miss big opportunities because a company doesn’t fit my framework or portfolio, but I believe you need to stick to your strategy.

I don’t think you can do a bit of everything. If you’re good at something, make sure you perfect it. Don’t try too many different things — you’ll only make mistakes, and your goals and strategy will stop being clear. All great investors have had a framework and stuck to it. The biggest mistake an investor can make is trying too many different things.

Question: If the portfolio doubles again from here, does the strategy change at all — different sizing, more diversification, less concentration — or does the same playbook just scale?

Answer: Always the same framework. I worked hard on my framework and my decision-making process in investing. As I said, I don’t think changing your way of working is a great idea.

— INVESTING STYLE & PHILOSOPHY —

Question: How would you describe your investing style in one or two sentences, to someone who’s never seen your feed?

Answer: I run a high-risk but well-diversified portfolio, built around several themes that I believe will determine the future of our world. I hold my stocks long-term and let them run to maximize my gains.

Question: Your framework is built around structural growth “waves” — space, photonics, batteries, drones, memory, compute, lidar — rather than broad sector calls. Where does that framework actually come from, and how do you tell a wave that’s genuinely structural from one that’s just a well-marketed narrative for a year or two?

Answer: I don’t think we’ve ever seen a wave like AI before, but I have seen several smaller ones. The world is moving faster than ever, so I think waves build up faster and also crash faster now. I like to think in terms of 10 years, but that might be an outdated way of thinking too — maybe the best investors now only think in two or three years, I don’t know. The fact is a wave never comes alone; it brings several other waves with it. The key is identifying them and seeing how far they can go. You can’t predict everything correctly, which is why I diversify. When you diversify across different fast-growing waves, I’m confident one or two of them will go very high.

Question: Is there a name for your process, or a specific framework you run every new idea through — the way some investors have a checklist or a set of filters?

Answer: I have my framework, but it’s not easy to explain in a single interview. My goal is to teach people how to build their own framework, because I think investing is something personal. You can’t just copy someone who has a different risk appetite, a different personality type, a different timeframe — you’ll always lose that way. When you build your own framework, you’re able to learn from your own decisions and make your own picks.

It’s not easy to build a framework like that, since it takes a lot of time and effort. People want to start buying high-risk stocks without even knowing their own personality — what their strengths are and what their weaknesses are as investors.

I really want to teach people this, but it takes a lot of time and effort. I’m working on some articles — I’m already at 10 long articles just to get people started. I don’t know if I’ll ever finish it, or in what form I’ll put it online, but I want it to be one of the best investing courses on the internet. We’ll see if I can make that happen.

Question: You’re heavily concentrated in AI infrastructure, space, and semis specifically. Was that a deliberate specialization, or did it happen organically because that’s where the returns were?

Answer: It happened organically. Like I said, I mainly buy a 5% allocation in each stock and let everything develop organically from there.

Question: Do you have an explicit sell discipline, or is every exit a judgment call? What actually gets you out of a position?

Answer: It’s a judgment call. When I want to exit a position, it’s because I doubt the thesis — I’m a fundamental investor, not a technical one. The only technical thing I do is check the 21SMA. I want to buy in a bullish environment and sell in a bearish one; that’s how I maximize my profits. People who claim to sell at the top and buy at the bottom aren’t trustworthy, in my opinion. I’ve read many books about strategies and market-timers, and no one has ever beaten the market by timing it. I don’t believe an anonymous account on X is going to be the first to do that.

Question: You’ve said you’ve beaten the S&P 500 for four straight years. Is there a year in that stretch that was closer than the headline number suggests, and what did that one teach you that the good years didn’t?

Answer: It was actually pretty close in most years. Only last year was a truly fantastic year for me — I moved from value stocks into high-risk growth stocks at the right moment. A bit of luck, but it certainly helped. In the other years I beat the index by around 5% to 10%, which is still great.

If I learned one thing from gaming, it’s that you build up advantages by protecting your small gains. I used to play League of Legends (Diamond 4 was my highest rank, so not great, but okay), and I was mechanically awful, which matters a lot as a mid-laner. But I respected the small advantages I built up over the course of a game — in the beginning you don’t even see them. In year one, I beat the S&P 500 by 5%. I didn’t get rich from that; you need patience to start the snowball rolling.

It’s the same with League — if you have an early advantage worth 5%, it feels like nothing. People don’t respect that small 5% edge and end up taking extra risks. In League you see it when someone with a small lead who’s dominating their lane suddenly starts making reckless tower dives — they take on more risk and lose all the advantage they’ve built.

It’s exactly the same in investing: building your advantage takes time. I see people buying a single company like Iren, or something similar — they’re young and looking for risk, but how will you ever start snowballing an advantage that way? I can take more risk now because I’m ahead. I’ve put around €35K into my portfolio so far, so it’s fine if my portfolio drops from €180K to €130K. I’ve built up the comfort to be patient in this market and buy the dips again to gain even more of an edge. You can’t snowball when you’re just starting out.

Question: What’s a mistake in your process — not a single bad trade, but a flaw in how you think about investing — that you’ve had to actively correct?

Answer: I already explained this one, and it’s a bit controversial — I used to think that every company would make new all-time highs eventually. If you believe that, then every company that drops 60–70% looks like a good investment.

— THE ACCOUNT —

Question: You’d been researching privately for years before the account existed. What was actually missing from doing it privately that made going public, at scale, the right next step in September 2025?

Answer: I’m always looking for opportunities to keep learning. I was on holiday with my wife last September, reading the book The Diary of a CEO. He talked about how much he learned from his own Twitter account — just putting his thoughts on the internet and interacting with people he didn’t know.

That’s the main reason I started the account: to keep learning and share what I learn with others. I don’t have specific goals on the platform, and I don’t do it for the money — if you compare my following to how much money I earn on X, I’m probably in the lowest percentile. As long as I keep learning on the platform, I’m fine with that.

My usual philosophy of tracking what works and writing it down doesn’t really apply to my X profile. I don’t look at my stats, and I don’t check which posts did well and which didn’t. I just want to give people useful information, and if 10 people read a post and really got something out of it, I’m a happy man.

Question: You grew to over 42,000 followers in under a year. What do you think actually drove that — the timing, the transparency, the specific names — versus what you expected going in?

Answer: I don’t know exactly what triggered it. Maybe it’s the fact that I don’t care too much — I’m here to learn, so that’s mainly what I’m doing. I’m fine with the full transparency, since people comment on my portfolio and I can learn from that.

I’m very process-minded. I don’t believe in goals like “I want to reach 50K followers by the end of the year.” I set goals like wanting to become a better investor, and learning on X is part of that. It’s the process of learning that matters to me, not the follower count.

Question: Your Planet Labs deep dive pulled 165K views but only 35 replies and 319 likes — a pretty wide reach-vs-engagement gap. What does that tell you about who’s actually reading you, and does it change what you write next?

Answer: I honestly have no idea if that gap is big or not. I put information out there that I think is valuable about a company. I don’t do any engagement farming, like asking questions at the end of a thesis just to get replies — I really don’t care about that. If people have questions, they can always ask. If not, they can just read the information and move on.

Question: Do you have any plans to monetize the research directly, the way some accounts move to a paid newsletter or community, or is staying fully public part of the model for you?

Answer: My picks will always stay fully transparent. I’d also never subscribe to someone who just shares picks. If I ever finish my “course,” it’s possible I’d monetize that — I don’t know in what format yet though. I’m not racking my brain over it yet; I’m fully focused on finishing it first, and then we’ll see.

Question: What does the account look like in a year — same format at larger scale, or something more structured?

Answer: I’m not planning to change much. We’ll see where I am next year.

— WHAT THIS IS REALLY ABOUT —

Question: If you had to distill your approach into one thing that separates investors who compound well over time from those who don’t, what is it?

Answer: Patience. As I mentioned before, compounding takes time and a well-structured approach. First build your strategy and decision framework, then make your first buys and build your portfolio — don’t start by buying a single stock you think might rocket. If you treat the stock market like a casino, the house always wins.

Question: How has your actual relationship with risk changed since you started, if at all — are you more or less willing to concentrate hard into a single idea today than you were a few years ago?

Answer: I explained this one already in a previous question. I feel like I can take more risk now, since I’ve built up an advantage. Time is on my side now.

Question: You’ve framed the account as being about teaching, not just picks. Does that mission feel in any tension with also managing your own capital and reputation publicly, or do you see them as the same thing?

Answer: No — I don’t care too much about my reputation on the platform, and only my wife knows I even have an X account, so there’s no conflict at all. If I make a pick, it’s public now. If that pick goes down and it damages my reputation on the platform, so be it. I think I’m qualified enough to teach people about building their own framework and decision-making. If they only want to use my picks, I’m fine with that too.

A big thank you to KaizenInvestor for being so open — both about the process and about the personal side of the past year. Interviews like this only work when people are willing to go beyond the picks, and he did.

For anyone new here: this is the Retail Investor Interview Series on Substack, where I sit down with retail investors who’ve built something worth learning from — not just returns, but the process, the mistakes, and the story behind how they actually think. The goal is to pull back the curtain on how different investors build conviction, manage risk, and stay disciplined over time.

If you enjoyed this one, here are a few other interviews from the series worth checking out:

And as always, if you have an interesting suggestion on who to interview next, let me know in the comments!

The report came in and the story is intact.

If you read my $AAOI Earnings Report article, you knew what my expectations were. I was waiting for updates on:

• Q2 revenue

• 800G shipments

• Texas and Taiwan facility progress

• 1.6T qualification and first shipments

• Q3 guidance

• Gross Margin

Let’s unpack everything.

Q2 2026 Earnings Release Highlights:

• Revenue: $191.9M (Est. $190.0M); +86% YoY

• Adj. EPS: $0.06 (Est. $0.01)

• Non-GAAP Net Income: $5.5M (Est. $1.7M)

• GAAP Gross Margin: 27.7%; -260 bps YoY

• Non-GAAP Gross Margin: 29.8%; -60 bps YoY

Q3 2026 Management Guidance

• Revenue: $255.0M–$290.0M (Midpoint: $272.5M vs. Est. $278.0M)

• Adj. EPS: $0.11–$0.26 (Midpoint: $0.185 vs. Est. $0.28)

• Non-GAAP Net Income: $10.1M–$24.0M (Midpoint: $17.05M vs. Est. $25.2M)

• Non-GAAP Gross Margin: 29.0% to 30.5%

AOI delivered a solid performance in Q2, and marked its fifth consecutive quarter of record revenue and returned to non-GAAP profitability. Demand for its products is surging, far outpacing what AOI can currently deliver. Management confirmed that customer demand exceeds the company’s current build capacity by 20% to 40%. Although guidance came in slightly weaker than analysts expected, AOI’s position as a critical supplier for AI infrastructure remains intact.

There’s so much to unpack from the recent AOI earnings report and the call. Let’s start with these:

1. Q2 Revenue

2. 800G Transceiver Shipments

3. Texas and Taiwan Facility Progress

4. 1.6T Qualification and First Shipments

5. Q3 Guidance

6. Gross Margin

7. Earnings Call Recap

8. Q&A Recap

9. Verdict

And at the end, I’ll cover everything else you might have missed that is important for the AOI story.

1. Q2 revenue

Leading-up to Q2 ER, management guided Q2 revenue between $180-$198 million, with consensus among analysts at $190.5 million. AOI delivered the figure of $191.9 million hitting slightly above the midpoint of their guidance.

Here’s the revenue breakdown per segment:

• Data Center: This was historically AOI’s secondary business. In Q2 it reached $107.7 million, up 140.4% YoY and representing 56.1% of total revenue. Today data center segment is AOI’s main growth driver.

• CATV: Despite fears of a cyclical dip, the CATV set a record with $80.6 million in revenue, up 43.8% YoY.

• Telecom: Telecom contributed $3.4 million, up 75.8% YoY, while FTTH and other segments contributed a minor $0.27 million.

One of the main highlights of this ER was AOI’s return to non-GAAP profitability, posting $5.5 million in adjusted net income ($0.06 per diluted share) which is a significant turnaround that reverses a long streak of sequential and YoY losses.

Customer and Credit Concentration

A key structural risk for AOI remains its extreme customer concentration. For the six months ended June 30, 2026, the company’s top ten customers accounted for 99.0% of total net revenues, with the top three accounting for 43.0%, 26.0%, and 24.0% respectively. Management said that they "continue to prioritize new customer acquisition and growth of diverse revenue streams." which is good news, as any loss of top customers would significantly impact AOI’s revenue.

This concentration has also impacted the balance sheet. Net accounts receivable grew to $314.0 million. This is mainly coming from a single customer Digicomm International, and represents $211.0 million or 67.2% of all outstanding trade receivables.

2. 800G Transceiver Shipments

While the base-case thesis expected 800G shipments to jump to $18.4 million sequentially from Q1’s $4.6 million, the actual delivered figure was $12.8 million. This is still +178.3% QoQ, accounting for 11.9% of total datacenter sales.

Data Center performance was marked by a massive surge in 400G transceivers, which brought in $48.4 million (45.0% of total Data Center revenue). This suggests that hyperscale customers utilized 400G modules to bridge their immediate optical interconnect needs while they wait for the massive 800G capacity ramp.

As usual in AOI’s case, the Q2 limit was strictly capacity based. During the Q&A, CFO Stefan Murry said “Our ability to deliver revenue in general, and specifically when it comes to 800G, is limited by our production capacity right now. If we could produce more, we could ship more.”

Management expects 800G shipments to go vertical in Q3 2026, guiding for a nearly 5x QoQ increase to $62.0 million to $65.0 million.

3. Texas & Taiwan Facility Progress

Read more

It’s been a while.

Let’s see what happened since the last update.

Read more

This time we’re going back to microcap land where we’re talking about Optical Cable Corp.

Why $OCC?

• The company posted its best quarter in years. - In Q2 FY2026, OCC’s net sales jumped 26.6% YoY to $22.21 million, while their gross profit surged by 42.4%. The company swung from a net loss in the prior year to a $1.10 million net profit.

• AI servers need a lot of wiring. - Management mentioned that the accelerating demand for cloud computing and AI applications is a primary driver for expanding their data center product offerings.

• It’s a microcap which could be at an inflection point. - OCC has a market cap of $131 million and it was recently added to the Russell Microcap Index. OCC currently has excess manufacturing capacity and management believes its existing manufacturing and staffing levels is able to support further revenue growth. This gives the company room to increase production without needing to do a major expansion of its manufacturing.

The Stock

OCC flew high during October 2025, but then got sold off in December and dropped to $4. I added it to my watchlist in March where it traded around $9, but at the time there were way too many networking names which were popping off quickly.

I started paying attention after OCC reported its Q2 FY2026 as the stock jumped to its ATH $27.98.

ER came in very strong:

• Net sales up 26.6% to $22.2 million

• gross margin at 34.2%

• net income of $1.1 million

• Backlog grew 82% in six months.

• The improvement in profitability required relatively little incremental capital spending which allowed higher sales to translate into higher gross profit.

In July, AI related stocks went through a massive drawdown, and OCC joined the party. The stock is today trading roughly about the same level as it did BEFORE the Q2 earnings report.

Who said that Mr. Market isn’t generous?

Table of Contents

• How it all started?

• What is OCC’s business about?

• How they make money?

• OCC’s product portfolio?

• Why Now?

• What’s Standing in the Way?

• Lightera Wildcard

• Fiber Market Analysis

• Financials

• The Last Earnings Call

• Insider Ownership

• What’s it Worth?

• Bottom Line

How it all started?

Optical Cable Corporation ($OCC) was founded in 1983 in Roanoke, Virginia. It was founded by Robert Kopstein after working on U.S. military fiber programs at ITT. He took his learnings with him and used the technology originally engineered to survive hard conditions on the battlefield and adapted it for commercial markets.

The company went public on Nasdaq in 1996. What happened after is quite a turn of events. Kopstein controlled 96% of the stock and pledged those shares as margin collateral across seven brokerage firms to invest on dot-com stocks at the top of the bubble. When the market crashed and his margin hit, and brokers seized and dumped millions of OCC shares in the open market to cover his personal trading losses. The stock dropped, shareholders sued him and Kopstein had to leave the company.

CFO Neil Wilkin took over and settled the class action. In January 2003 he made a move to save the company. He used $2.95 million in cash and credit to buy back over 21% of OCC’s own shares directly from A.G. Edwards (these shares were held as collateral against Kopstein’s defaulted margin loans) at just $1.00 per share.

Wilkin is still the CEO of the company and he added acquisitions in 2008 (SMP) and 2009 (AOS) to form OCC’s current three-site manufacturing footprint across Roanoke, Asheville, and Dallas.

OCC is a survivor and a turnaround story. Its product was designed in the 1970s to survive harsh military conditions, and the company itself proved tough enough to survive a founder margin liquidation, the dot-com crash, 2008, and decades of micro-cap obscurity.

The main question for OCC today is: Will AI allow the company to utilize its potential to the fullest?

What is OCC’s business about?

OCC manufactures fiber optic and copper data communication cabling and connectivity. This is the physical layer of a network.

OCC does not produce chips, switches, or software.

What OCC does is the wires, the connectors, the patch panels, the racks and enclosures that everything else plugs into.

How to simply summarize its products - data travels through data centers one of two ways:

• through copper wires as electrical signals

• through glass fibers as pulses of light

Fiber can carry a lot more data over longer distances which is the main reason why every modern network, data center, and cell tower runs on it. However, the glass fiber is fragile. OCC’s engineering value add is the cabling: how you buffer, armor and jacket those fibers so they survive being pulled through conduit, buried in a mine, or run across a military staging area. OCC’s specialty is exactly that ruggedization, plus the connectivity hardware on each end.

The company organizes its offering into three product families which correspond to its three distinct manufacturing facilities, all sold through one integrated sales team and reported as a single segment:

• The Roanoke, Virginia facility produces OCC’s core Fiber Optic Cable products which includes tight buffered, hybrid and heavily armored specialty cables customized for military ground tactical operations, mining, and broadcast environments etc.

• The Asheville, North Carolina facility manufactures the Enterprise Connectivity Products which provide a suite of copper and passive fiber hardware such as high-density data cabinets, relay racks, patch panels, and residential wiring enclosures.

• The Dallas, Texas facility is dedicated to the Applied Interconnect Systems Product products which includes tactical fiber optic connectors, hermaphroditic deployable systems, and the patented Modular Advanced Reel System (MARS) designed for harsh military and industrial environments.

End Markets

OCC's end markets spans across commercial enterprise networks, AI data centers, residential installations, and Passive Optical LANs, alongside a long range of severe duty niches that include the military, industrial facilities, mining, petrochemical operations, renewable energy.

How they make money?

OCC makes money the traditional way which means they do the manufacturing and sell physical networking and cabling hardware. They do not sell any software, subscriptions, or digital services. OCC’s revenue comes purely from selling the passive equipment that physically connects communications networks.

OCC has three core product categories:

• Fiber Optic & Hybrid Cables: Tight-buffered, armored, and hybrid/composite cables. These are designed for commercial enterprises, data centers, corporate campuses, universities, and heavy industrial sites.

• Enterprise Connectivity: Category 5e through Category 8 copper cabling, fiber patch panels, wall jacks, patch cords, data cabinets, and server enclosures for commercial office LANs and enterprise IT rooms.

• Harsh-environment Connectors: Ultra-ruggedized fiber optic connectors and tactical cable assemblies made for military field operations, oil and gas rigs, and outdoor mobile broadcast setups.

OCC sells their products primarily through major and regional distributors, original equipment manufacturers (OEMs), and resellers. While they do occasionally sell directly to end users, their products are mainly purchased through these intermediary distribution channels by contractors and system integrators who ultimately do the final installations.

Current OCC’s top-line growth and its rapidly expanding sales order backlog are being driven by strong demand from multi-tenant AI data centers (often referred to as Tier 2 data centers) and enterprise data centers, alongside robust continuing demand from military and severe-duty end markets.

OCC’s product portfolio?

In the previous section, we talked about OCC’s primary product offerings, which represent how the company divided its internal manufacturing. However, when OCC presents its product catalog to their buyers, distributors, and contractors, it organizes these products into five commercial categories.

OCC provides a comprehensive suite of passive infrastructure and connectivity solutions. OCC focuses on highly engineered components where reliability is of critical importance.

• Fiber Optic Cables & Connectivity: OCC’s signature product. They manufacture rugged, tight buffered fiber optic cables that offer extreme physical protection against crushing and moisture for both enterprise and harsh environments, together with the hardware needed to connect them.

• Copper Cables & Connectivity: An end-to-end portfolio of copper datacom cables which are supported by OCC’s patented, high-performance connectivity hardware like jacks and patch panels.

• Hybrid Cables & Enclosures: These highly specialized cables combine optical fiber for high bandwidth data and copper wires for electrical power within a single jacket. They are used for powering security cameras, cell towers, and industrial plants.

• Cabinets, Racks, Enclosures: High density data cabinets, relay racks, and wall mount enclosures which are needed to securely house and organize these complex network systems.

• Deployable Solutions: Built for the military and extreme duty sectors, this line includes ultra rugged tactical connectors and OCC’s patented MARS reel system designed for rapid cable deployment and retrieval in the field.

Why Now

The Demand

The AI revolution which is creating demand across many sectors is heavily impacting networking sector. It is pulling structured cabling demand up across the entire supply chain, and OCC could have a massive upside, bigger than any of the giant producers, if it manages to land enough orders.

In June 2026, management said that the demand is coming from the “enterprise, data center and severe duty market sectors,” and that sales into data center markets positively impacted Q2 revenue and are expected to keep contributing through the second half of fiscal 2026.

Previously, management mentioned opportunities in Tier 2 and Tier 3 data centers specifically for a company of OCC’s size, since hyperscaler primary builds are locked up by industry giants.

Growing Backlog

OCC’s forward visibility is currently growing fast. By the end of Q2 FY2026 (April 30), OCC’s sales order backlog had reached $13.30 million. To put that into perspective, this represents a massive 82.2% increase in just six months (up from $7.30 million at the end of fiscal 2025).

Operating Leverage

OCC currently has excess manufacturing capacity and management believes its existing manufacturing and staffing levels is able to support further revenue growth. This gives the company room to increase production without needing to do a major expansion of its manufacturing.

Because OCC still has some excess capacity this means that the near term revenue growth may not require an increases in capital spending. This creates operating leverage as higher production volumes absorb the company’s fixed manufacturing costs. As revenue grows, OCC can generate gross profit faster than sales because a portion of its manufacturing cost base is fixed.

In Q2 FY2026, revenue grew $4.66 million YoY, while gross profit rose $2.26 million. That represents an incremental gross margin of roughly 48.5%, far exceeding the reported average margin of 34.2%. This sort of operating leverage is the foundational engine of the bull case.

Expanding TAM and the Lightera Catalyst

OCC’s TAM is growing fast mainly due to data center infrastructure and defense spending. To capture the accelerating demand for cloud computing and generative AI applications, OCC formed a major strategic collaboration with Lightera in July 2025.

This partnership is a major growth driver for OCC. It guarantees a stable supply of premium optical fiber, which protects the company from industry-wide material shortages. It allows OCC to expand its catalog with next-generation products aimed particularly at modern data center and enterprise applications. Finally, Lightera showed a strong confidence in the deal and invested $2.0 million to buy a 7.24% ownership stake in OCC.

Institutional Visibility

It’s not a secret that microcaps suffer from a lack of liquidity and visibility, but for OCC things changed. On June 26, 2026, the company was added to the Russell Microcap Index. Russell Microcap inclusion increases OCC’s visibility to institutional investors and can create incremental demand from funds that track or benchmark against the index. For a relatively illiquid microcap, that exposure can improve trading liquidity over time.

What’s Standing in the Way

These are some of the risks worth paying attention to.

Liquidity: $145,600 of cash

Yes, this is all the cash they have against $7.33 million drawn on its revolver.

OCC operates with very little cash because its credit arrangements use a cash-management structure in which customer receipts are used to reduce their borrowings. That means that the reported cash balance can be quite low even while the company continues funding normal operations through its revolving credit facility.

The debt is secured by company assets and includes customary covenants and events of default. A breach of those provisions could allow the lender to accelerate repayment, making continued compliance with the credit agreement important given OCC’s limited cash balance.

Fighting against giants

OCC is facing a strong competition in the networking sector. A lot of companies they compete with have a larger R&D budget than OCC’s annual revenue.

When it comes to fiber for data centers, the supply is contracted for years in advance by companies that can guarantee big volumes.

Since OCC is too small for such big orders, their most realistic opportunities are Tier 2/3 data centers. That opportunity is real, but it also means that OCC gets AI capex as a secondary wave, and not through a direct allocation. If AI capex slows down, secondary waves are first to get hit.

The valuation itself

The stock went through a violent repricing. The current price assumes that the growth continues. A stock that repriced this violently on one strong quarter can reprice just as violently on one soft one. With the stock being a microcap, its liquidity amplifies swings in both directions.

The Lightera “Put/Call” Overhang

This partnership is a catalyst, there’s no doubt about it, but the equity structure behind it creates a drag on the balance sheet which I have to mention.

Lightera bought OCC stock but didn’t get standard common stock. They got 642,199 redeemable restricted shares with a put/call feature. Under GAAP, this deal keeps the shares outside permanent equity, marked each quarter to the stock price.

The key date here is July 7, 2027 when Lightera’s put becomes exercisable. OCC’s cash exposure is capped at the original $2.0M but for anything above that OCC must do one of the two: register the shares so Lightera can sell them on the open market, or issue a promissory note. Neither option is great for OCC. It either means sudden dilution hitting the float, or there’s added debt to a balance sheet that’s already very low on cash.

The Lightera Wildcard

On July 7, 2025, OCC announced a strategic collaboration with Lightera.

Who is Lightera?

It is the U.S. optical fiber manufacturer that goes back to the old AT&T/Lucent fiber operations and today it is part of Japan’s Furukawa Electric Group which is one of the world’s major fiber producers.

We mentioned earlier that Lightera bought OCC stock but didn’t get standard common stock. They got 642,199 redeemable restricted shares with a put/call feature.

The deal mechanics

The structure of this deal is very unusual.

• Lightera bought 642,199 newly issued redeemable restricted shares for $2.0 million in cash. That is roughly $3.11 per share, granting them a 7.24% equity stake in the company.

• Under a Joint Collaboration Agreement, the two companies integrated their product portfolios. What this means for OCC? It means they can sell Lightera’s products alongside their own, aimed at the enterprise sector, multi-tenant data centers, and Passive Optical LANs.

• OCC holds a call option to buy their shares back, while Lightera has a put option to force OCC to buy them back. None of these options is exercisable for two years (until July 7, 2027), and they expire 90 days after that unless their collab is extended. OCC’s cash repurchase obligation is capped at the $2.0 million invested, but if Lightera exercises the put when the shares are worth more than $2.0 million, OCC must either register the excess shares for public sale, issue Lightera a promissory note, or help them dispose of the stock.

Why it matters for the thesis

The collab expanded OCC’s data center offering as now they can combine selected Lightera products with its existing cabling and connectivity portfolio. That gives OCC a broader solution set for multi-tenant and enterprise data center customers and could increase the size of the opportunities for them.

Additionally, this deal is a supply-chain hedge for OCC. It locks in a reliable, premium fiber source in today’s industry where there are shortages of optical fiber with extended lead times.

While this is just an option, there could be a possibility that Lightera acquires more of OCC due to the setup of their ownership, and two year evaluation window.

Remaining Sections:

• Fiber Market Analysis

• Financials

• The Last Earnings Call

• Insider Ownership

• What’s it Worth?

• Bottom Line

Fiber Market Analysis

Fiber Supply Layers

The optical fiber supply has three layers:

• Bare fiber production (melting silica into glass and drawing it into strand)

• Cabling (buffering, bundling and jacketing that strand into a usable cable)

• Connectivity (connectors, patch panels and terminations)

OCC is in the second and third layer. Not in the first one.

Read more

If you’ve followed this name for even a bit, you already know that AOI is probably the most volatile stock on the planet. To put it briefly, it’s an absolute rollercoaster: after jumping to an all-time high of $233.67 on May 13, the stock got caught in a brutal AI sell-off. A combination of the broader AI hardware sell-off, the forced liquidation of the Leopold Aschenbrenner “Situational Awareness” public stock portfolio, and geopolitical tensions pushed the stock down over 65% to a low of $76 in late July.

Over the last few sessions, AOI moved from $76 to $134 jumping 76% and seems that there’s no slowdown to its price action.

If you haven’t followed AOI before, please check the article below to get all the details about the company and my investment thesis.

Quick Intro

AOI is a vertically integrated manufacturer of fiber optic transceivers. It makes their own laser chips, builds their own components and also assembles all of it in a finished product, all under a single roof. While others buy critical parts from someone else, AOI makes everything themselves, and the production starts with the laser chip grown in Sugar Land, Texas.

The laser chip is the most critical component inside of every optical transceiver. That’s also the core of AOI’s bull case. They control the supply chain that most competitors depend on third parties for.

AOI’s business has four revenue streams:

• data center

• CATV broadband

• telecom

• FTTH

The revenue is concentrated and you can guess very lumpy.

In 2025, CATV was AOI’s largest market with $245.1 million (53.8% of revenue), while data centers came in with $195.7 million (42.9% of revenue). In 2026, the data center is the main growth driver. In Q1 2026 data centers were at $81.4 million (53.9%), where CATV came at $66.8 million (44.2%).

AOI is going through their most aggressive manufacturing buildout in their history. They are expanding their Houston-area footprint toward 900,000 square feet where they will be growing laser fabrication capacity by approximately 350% by 2027, and targeting over 650,000 transceiver units per month by year-end.

Management said that revenue is limited by production capacity and supply chain, not market demand.

AOI’s only issue is how fast they can build.

The Setup

Management guided Q2 at $180M–$198M.

Analyst consensus is between $190.48 million to $194.29 million.

The July selloff hit AAOI alongside the entire AI sector. It was a combination of macro fears, together with the forced liquidation of the Situational Awareness public portfolio, and wider sector selling pressure pushed the stock down over 65% from its ATH. Going into ER, the stock has recovered fast fueled in part by news that the Trump administration is drafting a plan to ban Chinese transceiver imports.

The thesis was never touched. The stock was sold with the sector.

2026 Revenue

Before we get to more details, here’s AOI’s revenue growth before 2026. The table shows 2026 per quarter where we can see Q1 actual, Q2 guided and Q3-Q4 implied.

Read more

Welcome back to the Retail Investor Interview Series. Today I’m sharing the story and interview with .

If you follow AI infrastructure, semiconductor, or data center investing on Substack or X, you’ve likely come across . He has been posting portfolio updates and trading commentary on X since the end of August 2025, but in mid-July 2026 he co-launched with two other investors, and . In just twelve days, the new publication and its Discord community had already attracted over 1,000 paying members.

His story is interesting not because of the growth numbers, but because of what sits behind them: a software engineer who started in 2017 with roughly $10K while working at a gas station and has spent years building toward it through active, concentrated positioning rather than passive holding. He describes his approach as trading around core positions (buying weakness, trimming strength) rather than simple long term holding, and says it has meaningfully outperformed a static approach on names like , , and .

I sat down with him to dig into the story behind the numbers: how his investing style actually works, what led to the Edelbridge Alpha launch, and where he sees it going from here.

Intro

How it all started

He has been investing since 2017. He started with roughly $10K while working at a gas station, then moved into software engineering toward the end of 2018 while adding consistent monthly contributions and building his process through both steady compounding and a few extremely high-risk decisions. For most of that time, he did it almost entirely in private while working his day job. His professional background is in software engineering, not finance, and what he has is a process he refined himself:

• deep focus on AI infrastructure and semiconductors

• active trading around core positions

• willingness to share both the wins and the painful lessons publicly

He had already been posting portfolio updates, theses, and journey reflections on X for some time. The bigger shift came in 2026 as his performance became more visible. In mid-July 2026 he co-launched with two other investors. Within days the Substack and Discord had attracted over 1,000 paying members. He has described the combination of transparent performance updates and the new collaborative platform as the moment that significantly expanded his reach, because it showed there was real demand for detailed, honest research and real-time portfolio transparency in the AI infrastructure space.

The philosophy

His framework centers on high-conviction ideas within AI infrastructure and semiconductors, combined with active management. He looks for opportunities where he can build a strong thesis and then improve returns by trading around the core position. This means buying weakness, trimming strength, and adding back instead of just simply holding long term.

His portfolio is heavily concentrated. He started 2026 essentially all-in on NBIS and has traded around it and a handful of other names. Today he emphasizes sticking to process and managing risk while remaining highly concentrated in his highest-conviction ideas. He holds through meaningful drawdowns when the long-term thesis remains intact. He doesn’t pretend conviction feels comfortable, he simply refuses to let short term volatility override his thesis if it hasn’t broken.

The model

He has consistently said he will keep sharing the journey and research publicly, just as he always has. In May 2026 he announced plans for optional paid access for those who wanted more (earlier ideas, live trade details, direct discussion). This evolved into the July 2026 launch of Edelbridge Alpha Substack and Discord which is a collaborative, paid platform offering deeper dives, portfolio updates, executive interviews, trade alerts, and community discussion.

His reasoning is straightforward: he is doing the research and management anyway. Sharing the core work publicly costs nothing extra and creates a valuable feedback loop that sharpens his thinking. The paid layer provides additional depth and community for those who want it, while the public X presence remains the foundation. Non-paying followers continue to receive the same core updates and transparency they always have.

Interview

— THE PERSON —

Who is Babyfolio outside of investing?

First and foremost, I’m a husband and a father. I feel incredibly grateful to be in this position. My wife and our two kids are the driving force behind everything I do and the reason I work so hard to build a better future for our family.

I’m also a gamer at heart. Before becoming a father, I spent countless hours playing games, especially strategy games. After my two kids were born, the time and commitment required to game disappeared almost overnight. Even so, that mindset never left me.

Strategy games taught me to think several steps ahead, weigh risk versus reward, and make decisions with the long game in mind. They shaped the way I approach investing and life in general. To me, life is a game you can win with patience, discipline, and consistently making better decisions than the average person.

If you’re not working or investing, what other hobbies do you spend time on? Also, is it true you regularly go afk while playing ARAM (League of Legends game mode)?

On the very rare occasions when I’m not investing or working, I usually go back to one of my old passions: gaming.

I’m still a gamer at heart. These days, with two young kids, I don’t have the time I used to have, but I still try to squeeze in a few games at night when I can, mostly League of Legends (usually 1–2 ARAM games) or some TFT.

And yes, it’s true, I regularly go AFK during ARAM.

The reason is simple: my kids still wake up at night, and unfortunately, “my team needs me” loses every time to “my kid needs me.”

So if you ever see me randomly disappearing mid-fight, blame the toddlers, not me.

What does your life actually look like?

At the moment, I still work a full-time job (although I plan to leave it soon). As you can imagine, balancing a full-time career, two very young kids, investing, and working on high-quality research makes life pretty chaotic.

Most of my days revolve around my family. My kids always come first, and that’s where the majority of my time and energy goes.

The nights are when I get to focus on my own work. That’s when I dive into research, write for Substack, engage with the community, and share my thoughts on X. Occasionally I’ll sneak in a post during the workday, but the real work happens after everyone else is asleep.

It’s a demanding schedule, but I genuinely enjoy it. Every late night feels like an investment in the future I’m trying to build for my family.

You’re a software engineer by day — how do you find the time and headspace to actively trade concentrated AI infrastructure positions alongside that?

One advantage of the way I invest is that I don’t need to make decisions every single day. My strategy is centered around building high-conviction core positions and swing trading around them rather than constantly chasing new opportunities.

Like most investors, I watch the market every day, but I rarely make changes unless something significant happens or the thesis changes quickly. That gives me the luxury of spending more time thinking than reacting.

Compared to a day trader who’s glued to the screen from market open to close, my schedule is much more flexible. It allows me to focus on researching businesses, understanding industries, and making deliberate decisions instead of constantly feeling the need to trade.

What does a normal Tuesday actually look like, hour by hour?

A typical Tuesday looks something like this.

My wife and I wake up around 6:00 a.m. to get the kids ready for the day. By 6:45, we’re out the door. We drive both kids to kindergarten, and then my wife drops me off at the train station so I can head to work.

The next nine hours are spent at my full-time job. It’s demanding and can be pretty exhausting, although I usually find a few moments throughout the day to check the markets or post something on X.

After work, I head straight home to my wife and the kids. We spend about an hour to an hour and a half together playing before it’s time to start the bedtime routine.

If you’re a parent, you already know what’s coming.

Getting two young kids to sleep can easily take up to two hours. Some nights everything goes smoothly. Other nights... not so much.

Once they’re finally asleep, my second shift begins.

I sit down at my computer and spend the rest of the evening doing what I love: catching up on the day’s market action, reading news, researching companies, writing for Substack, interacting with the community on X, and talking with friends. Despite what people might think, we don’t spend every conversation discussing stocks. Most of the time we’re just talking about life, joking around, and occasionally arguing over whose investment thesis is better.

By the time I finally shut everything down, it’s usually somewhere between 1:00 and 2:00 a.m.

Then a few hours later, the alarm goes off, and we do it all over again.

It’s a tough schedule, and sleep is definitely the first thing I sacrifice. But I know it’s temporary. Every long day and late night is another step toward the life I’m trying to build for my family.

You’ve said this is about generational wealth for your two kids. Do you plan on making your kids rich, or do you prefer to teach them how to become rich themselves?

I think it’s a mixture of both.

Of course, I want to create opportunities for my kids and give them a strong foundation, but I don’t want to hand them an easy life where they never have to struggle or earn anything themselves.

If possible, I would actually like to keep the extent of what I’ve built private from them for as long as I can. I want them to understand the value of money, appreciate hard work, and learn that most people have to work hard to earn what they have, and that they should too.

There’s a common saying that the first generation builds the wealth, the second generation grows it, and the third generation loses it. While it’s not always true, I think there is an important lesson behind it: wealth without the right values can disappear quickly.

For me, the most important thing isn’t just leaving my kids money. It’s leaving them the mindset, discipline, and humility required to handle it responsibly.

I want them to respect hard work, treat everyone as an equal regardless of their financial situation, and never let wealth create a sense of entitlement or superiority. Money is a tool that can provide freedom, but character is what determines how someone uses that freedom.

Now that many people know about the baby’s hat on your avatar, tell us more about it.

That’s actually a funny story.

Before I changed my profile picture to an animated version of my real kid, I used the “Baby Boss” profile picture. Whenever my Nebius allocation went above 80%, people started noticing that I would put a Nebius hat on the Baby Boss.

It slowly became a little tradition in the NBIS community. Every time people saw the hat appear, they knew I was extremely confident in the position.

Unfortunately, I haven’t kept up with it as much recently, but it became a fun little symbol that people associated with my highest-conviction bets.

— THE ORIGIN —

You started in 2017 with around $10K and monthly contributions. What did the first couple of years actually look like, before any of this was concentrated or high-conviction — what were you buying, and how bad were the early mistakes?

I started with around $10K in 2017 while working at a gas station. I only became a software engineer toward the end of 2018, so back then I was just a 21-year-old trying to figure out how to make money online.

My first mistake was falling into the day trading rabbit hole. I found a day trading course, completed it, and thought I had found the answer. I spent 1-2 months day trading, made terrible emotional decisions, and lost around 50% of my portfolio.

That was my first wake-up call. I realized day trading wasn’t for me, so I started consuming as much investing content as possible, learning about different strategies, investors, and styles. Growth investing immediately clicked with me.

My first real investments were Apple and Nvidia. I put 50% of my portfolio into each and held them until late 2018. Then Trump announced China tariffs, the market sold off, and I lost another 50% of my portfolio. This time, the damage was amplified because I was using margin to double down.

At that point, I realized I was doing something wrong. I didn’t need more conviction, I needed a system.

I started researching what characteristics historically increased the odds of a company succeeding. Things like founder-led companies, management with skin in the game, strong revenue growth, margin expansion, and large market opportunities and fairly low market cap (which allows for more upside).

That research eventually led me to Tesla (it was a $30B company at the time).

Elon Musk, at the time, represented almost the perfect setup. A founder-led company, massive insider ownership, a CEO with extreme conviction, and a company attacking a huge market. I started with 50% of my portfolio in Tesla in early 2019, and by mid-2019 I was essentially all-in when the stock was trading around the split-adjusted equivalent of $14 and the majority of people were convinced Tesla was going bankrupt.

During this period, I still had other positions and used margin occasionally to try to increase my Tesla exposure. Looking back, it added unnecessary stress, but at the time I was extremely confident in the thesis.

My wife, who was my girlfriend at the time, trusted me completely and gave me her savings to invest as well. We started putting every salary we earned into Tesla.

By February 2020, I was feeling unstoppable. I was up around 300% from the lows, my conviction was at an all-time high, and everything seemed to be working.

Then March 2020 happened.

My portfolio dropped around 60%. It was a brutal reality check. We had almost no cash, our salaries were relatively small at the time, and everywhere you looked there were predictions that the world economy was collapsing.

I sat down with my girlfriend and told her: “Either everyone is right and this is the end of the world, or this is the opportunity of a lifetime.”

She trusted me, and I made what was probably the riskiest decision of my life. I took loans from multiple places that approved me, around 4-5 loans in total, and went heavily on margin through IBKR.

When people hear this part of my story, they usually think it was a mistake. Honestly, I agree. It was an enormous risk. I could have lost everything and still been paying those debts today.

I would never recommend anyone do what I did.

But it also happened to be the decision that completely changed my life.

March 2020 ended up being the bottom, and over the following period I grew my portfolio massively, eventually around 20x from those lows. I paid back the loans and bought two properties in 2021.

During that period, I continued trading other names like Cloudflare (NET), Sea Limited (SE), and Trade Desk (TTD), and a few losers like SKLZ, to try to increase my Tesla position. It was the SaaS and e-commerce boom, where almost every growth company seemed unstoppable, very similar to what the AI trade has felt like recently.

My biggest mistake was not following my own system. I became emotionally attached to Tesla and started ignoring warning signs. By 2021-2022, revenue growth was slowing, margins were compressing, and Elon was increasingly focused on Twitter.

The result? I became a Tesla bagholder for years.

Eventually, in 2025, I decided I needed to move on. Around that time, I started looking for companies that reminded me of Tesla in its early days, companies with a massive opportunity, strong leadership, and asymmetric upside.

That company was Nebius.

I originally found it through , then did my own research. The more I studied it, the more it reminded me of Tesla’s early setup: a founder with a strong track record, a huge market opportunity, and the potential for a massive growth curve.

I started slowly rotating out of Tesla into Nebius around the $50-60 range. The Microsoft contract was the moment everything changed for me. It significantly reduced the execution risk and validated the thesis.

At that point, I had enough conviction to go all-in around $104.

And the rest is history.

How did you initially learn? Did you have a mentor? Did you watch YouTube videos, or read books or something else?

I would say the majority of what I learned came from experience.

Early on, I consumed a lot of content to understand the basics. I watched YouTube videos, read about different investing styles, and even took a day trading course.

Ironically, that course was probably one of my most valuable lessons because it cost me 50% of my portfolio. It quickly taught me that day trading wasn’t the path for me and pushed me to look deeper into investing, business analysis, and finding a strategy that actually fit my personality.

Over time, the biggest teacher has been making mistakes, analyzing what went wrong, building a system, and continuously improving it.

Was there a genuine blow-up — a position or stretch where things went badly wrong and it actually hurt — that changed how you think about risk, separate from $TSLA?

Yes. While Tesla was the biggest turning point in my journey, I had another major realization after experiencing a brutal 50% drawdown in October 2025.

That moment forced me to really rethink my approach to risk.

I realized there was no reason I should simply accept massive drawdowns as part of investing. Yes, volatility is unavoidable, especially when investing in high-growth companies, but there should be ways to manage risk better, minimize losses, and potentially even take advantage of those periods.

That was the point where I started becoming much more focused on trading around my core positions instead of just holding through everything.

I’ll also give credit where it’s due. I learned a lot about trading from Daniel Koss. I took many of the concepts he uses and refined them to better fit my own personality and risk tolerance. My style is a bit more conservative, but the principles helped me become a much better investor and trader.

Since then, I’ve incorporated more active management into my strategy, especially around my highest-conviction positions. The goal isn’t to constantly trade, but to protect capital during downturns and maximize opportunities when volatility creates them.

A good example is Nebius. As of writing this, NBIS is up around 149% YTD, while my portfolio is up around 270% YTD. The difference comes from not just picking the right companies, but also managing positions effectively along the way.

— INVESTING STYLE & PHILOSOPHY —

How would you describe your investing style in one or two sentences, to someone who’s never seen your feed?

I combine deep research with active position management, holding high-conviction names while trading around them when the risk-reward changes.

You call yourself an active trader around core positions rather than a buy-and-hold investor. What’s the actual edge you believe active management gives you that holding statically doesn’t?

The edge is flexibility.

By actively managing around core positions, I can take advantage of volatility, reduce risk when the upside becomes limited, and increase exposure when the risk-reward becomes attractive again.

A great company can still be a bad stock at the wrong price, and active management helps me adapt.

Is there a name for your process, or a specific framework you run every new idea through — the way some investors have a checklist or a set of filters?

I don’t have a specific name for it, but I do have a framework I use for every new company.

I focus on finding factors that increase the probability of success, such as strong revenue growth, expanding margins, founder-led management, insider ownership, large TAM, and strong industry tailwinds.

The goal is simple: identify companies where the odds of winning are higher than average and where the risk-reward is asymmetric.

You’re heavily concentrated in AI infrastructure and semis specifically. Was that a deliberate specialization, or did it happen organically because that’s where the returns were?

It was a bit of both, but the main factor was following where the strongest tailwinds were.

I believe the best opportunities usually come from being positioned in industries experiencing massive growth, strong demand, and significant capital investment.

AI infrastructure and semiconductors became the areas where I saw the biggest opportunities, with huge secular tailwinds and companies benefiting from the AI buildout. That’s where I believe the highest risk-reward opportunities are right now.

Do you have an explicit sell discipline, or is every exit a judgment call? What actually gets you out of a position?

I don’t have a strict sell rule or a specific percentage that forces me out of a position.

Every exit is a judgment call based on a combination of fundamentals, new information, valuation, and the stock price.

If the business changes, the thesis weakens, the valuation becomes unattractive, or I find a better risk-reward opportunity elsewhere, I will adjust or exit the position.

How much of your process is quantitative — valuation, growth, margins — versus a qualitative read on management, narrative, and market psychology?

It’s a combination of both, but I would say the qualitative side plays a bigger role.

The numbers matter, revenue growth, margins, valuation, and financial metrics help me understand the business and the opportunity.

But investing is not just about numbers. Management quality, founder mindset, market opportunity, industry tailwinds, and understanding investor psychology are equally important in determining whether a company can become a massive winner.

Has your philosophy changed meaningfully since 2017, or is it mostly the same instincts applied at larger scale now?

My philosophy has changed significantly since 2017.

Early on, I was mostly investing based on excitement and stories. Over time, I learned that I needed a real framework and a process to increase my odds of success.

The core idea stayed the same, find great companies with massive opportunities, but today I combine that with better risk management, deeper research, and more active position management.

What’s a mistake in your process — not a single bad trade, but a flaw in how you think about investing — that you’ve had to actively correct?

The biggest flaw I had to correct was becoming too attached to the story instead of staying focused on the business.

The Tesla experience taught me that having a great narrative and a great company doesn’t mean you should ignore changing fundamentals. I was too emotionally attached to the investment and started defending the story instead of objectively reassessing the thesis.

Since then, I’ve focused much more on staying objective, constantly challenging my assumptions, and making decisions based on data, execution, valuation, and risk-reward rather than just conviction.

— THE PROCESS —

You describe trading around core positions — buying dips, trimming rallies — rather than holding statically. What actually triggers a trim versus a hold, in concrete terms, not just “conviction”?

A lot of it comes down to what price my model gives me for the stock.

For example, let’s say I believe Nebius can reach around $320 fair value by the end of 2026, and the stock is already trading around $290 in June 2026. At that point, I have to ask myself, “How much upside is actually left compared to the risk I’m taking?”

If the remaining upside is limited, I become more comfortable trimming and reallocating capital to better opportunities, rather than staying fully exposed just because I still like the company.

Another factor that helps me make those decisions is technical analysis. I know many people dismiss it as “chart doodles,” but in my experience, it can be a very useful tool when combined with strong fundamental analysis.

I look at things like how far the stock has moved away from key moving averages, momentum, and whether the stock has become too extended in the short term. These signals don’t determine my decisions by themselves, but they can increase my confidence when deciding whether to trim, hold, or add.

The goal is not to perfectly time the top or bottom. It’s about managing risk, maximizing upside, and making sure my capital is always positioned where I believe it has the highest potential return.

Another thing that could cause me to trim or sell is simply seeing better risk/reward on another stock.

How do you separate “the market is wrong about this” from “I’m wrong about this” when a high-conviction position moves against you?

That’s a tough question, and honestly, it’s something I think about a lot.

I’ve written about this publicly before, but the sunk cost fallacy is a very real thing, and it’s one of the hardest psychological traps for investors to avoid. The more time, research, and emotional energy you put into a position, the harder it becomes to admit you might be wrong.

The only way I know to fight it is to be as objective as possible. Ignore the story you fell in love with and go back to the data. Look at the fundamentals, the execution, the original thesis, and ask yourself: “If I didn’t already own this stock, would I buy it today at this price?”

You have to constantly separate the company from the stock and avoid becoming emotionally attached to a narrative.

That being said, it’s much easier said than done. I think every investor, including myself, struggles with this. The goal isn’t to completely remove emotions, because that’s impossible. The goal is to recognize when they are influencing your decisions and actively fight against them.

How concentrated is too concentrated for you? You’ve gone close to 100% in a single name more than once — what has to be true for you to do that again, and what would make you never do it again?

For me, there isn’t really a specific percentage where I say, “this is too concentrated.”

I believe the right level of concentration depends on where I see the best risk-reward opportunity. If I find a company where my conviction is extremely high, the fundamentals are strong, the opportunity is massive, and the downside is manageable, then I am comfortable allocating heavily.

At the end of the day, I don’t optimize for diversification. I optimize for maximizing returns while managing risk.

That being said, concentration only works if you truly understand what you own. Going heavily into a single name without doing the work is just gambling. The more concentrated the position, the higher the responsibility to constantly challenge your thesis and make sure you’re not just holding because of emotional attachment.

So for me, there is no universal “too concentrated.” The concentration level should match the quality of the opportunity and the strength of your understanding.

Definitely not for everyone though.

You’ve used a little margin at times, separate from the 2020 $TSLA leverage. What’s your actual rule for when leverage is acceptable versus when it isn’t?

I use a small amount of margin (up to 15-20%) occasionally when I see a specific opportunity and believe there are near-term catalysts that could drive the stock higher.

For me, margin is mainly a tool for short-term swing trades, not something I want to use as a permanent part of my portfolio. I don’t want to be exposed to margin for months at a time because, in my opinion, that is a recipe for disaster.

The biggest lesson I learned from the past is that leverage can amplify both the upside and the downside. Used carefully, it can be a tool, but when you rely on it for long-term investing, you create unnecessary pressure and risk.

My preference is to have my core investments funded with my own capital and only use limited leverage when I have a clear setup, a defined reason, and a shorter time horizon.

You sold covered calls on $NBIS and it cost you real upside. Walk me through that mistake specifically — what did you think you were protecting against, and what did you actually give up?

I wasn’t really using covered calls as a hedge. My original goal was to generate additional income and use the premiums to accumulate more NBIS shares.

The problem was that it worked too well at first.

The market was very choppy, which is usually an ideal environment for selling calls, and after a few successful trades I started feeling like I had figured it out. I became overconfident and underestimated the risk of a sudden move higher.

Then the Iran war happened in March 2026, the market reacted, and NBIS had a massive run-up. Because I had sold calls at much lower strike prices, I ended up giving away a significant amount of upside.

I estimate that I missed out on over $1 million in potential gains. Looking back, I was essentially making pennies compared to the dollars I was giving up.

The biggest lesson was that covered calls can be a useful strategy in the right situations, but when you have a high-conviction position with a potentially asymmetric upside, you have to be extremely careful. Sometimes the best trade is simply owning the shares and letting the thesis play out.

Trading actively around a core position adds a lot more decisions than buy-and-hold. How many of those decisions do you think are actually adding value versus just adding noise or stress?

I think the value creation is real, at least based on my own results. As of writing this, NBIS is up around 149% YTD, while my portfolio is up around 270% YTD. The difference comes from actively managing the position rather than simply holding it.

That being said, this approach is not for everyone.

The biggest challenge is not necessarily the strategy itself, but the psychological pressure that comes with it. There are two main sources of stress.

The first is seeing your average cost basis change. When you trade around a position, your cost basis can look much higher than someone who simply held from the beginning. Some people struggle with seeing a higher average cost and feel like they made a mistake.

Personally, I had to learn to detach emotionally from cost basis. The only thing that matters is where you are today and what the future risk-reward looks like.

For example, if I bought NBIS at $150 and sold at $290, then later bought back at $220 and the stock dropped to $200, I’m not looking at it as a 10% loss. I’m still sitting on a significant gain, and potentially a better overall result than someone who simply held the entire way.

The second challenge is being in cash. Having cash on the sidelines can create its own psychological battle. You start feeling impatient, you get FOMO when a stock runs without you, and you constantly question whether you made the right decision.

Active management can add value, but it requires a lot of emotional control and discipline. If someone can’t handle the stress and uncertainty that comes with making more decisions, then a simple buy-and-hold strategy might be the better approach.

— THE PORTFOLIO MANAGEMENT —

Walk me through today’s actual allocation — main positions and weights.

— EDELBRIDGE ALPHA & THE SUBSTACK —

You’d been sharing research and updates on X for years before Edelbridge Alpha existed. What was actually missing from doing it there that made a paid Substack and Discord the right next step?

The main reason was that putting dozens of hours into research and sharing everything for free simply wasn’t sustainable.

But beyond monetization, I wanted to build a community of high-quality investors.

Me, Daniel Koss, and Alasdair Mann wanted to create a place where we could combine our knowledge, challenge each other’s ideas, and make better investment decisions together.

A lot of people think we are all doing the same thing, but behind the scenes we often have very different opinions and approaches. We constantly debate, verify, and challenge each other’s assumptions, and when we do agree on something, it usually means the thesis has been thoroughly tested.

The paid Substack and Discord allow us to keep producing high-quality research while building a community around it.

You announced plans for paid access back in May 2026, well before the July launch. What happened in that gap — what were you and your co-founders actually building or debating?

I started out by sharing ideas on X, mostly short trade ideas and thoughts. It was a great platform for reaching people, but it has limitations. The format is mostly one-way, where I post an idea and people react in the replies, but it doesn’t really create deep discussions or meaningful investor conversations.

During that gap, we were thinking about how to build something different.

Substack and Discord felt like the right combination because it allowed us to go deeper with research, have higher-quality discussions, and create a community where investors can challenge ideas and learn from each other.

The goal was never just to post more ideas. It was to create a place for better conversations and a more enriching investment process for everyone involved.

Why co-found with @daniel_koss and @itsalasdairmann specifically, rather than launch solo? What does each of you bring that the others don’t?

First of all, I genuinely love working with both of them. They are exceptionally smart, have achieved amazing things, and each brings a unique perspective.

The main reason I wanted to do this together is because investing is better when ideas are challenged and verified. We don’t always think the same way, and that’s exactly the point.

Daniel, Alasdair, and I each have different experiences, expertise, and ways of looking at opportunities. That mix creates better debates, helps us challenge our own assumptions, and ultimately leads to better investment decisions than any of us would make alone.

What does a paying member get that a free X follower doesn’t? Walk me through it concretely.

On X, I share ideas and thoughts, but it’s mostly short-form content. Edelbridge Alpha is where we go deeper.

Members get high-conviction ideas, actionable trade setups, full portfolio updates, deep dives, financial models, price targets, industry analysis, management interviews and access to discussions where ideas are challenged and peer-reviewed.

The goal is not just to give people more stock picks, but to show the full process behind the decisions, including the risks, assumptions, and what could make us wrong.

You grew to 1,000+ paying members within days. What do you think actually drove that — the timing, the names, the free track record — versus what you expected going in?

I think it was a combination of everything.

Our track records definitely played a big role, but also the fact that we already had existing communities and subscribers who trusted our work. We had been sharing research and ideas independently, and many people were already following our journeys.

The launch was also driven by the fact that we wanted to build something we were more passionate about. We moved away from our previous formats and created a community built around deeper research, better discussions, and collaboration between investors.

The response was much bigger than we expected, and we are extremely grateful for the trust people showed us.

Free X updates continue as the “public foundation” — how do you decide what stays free versus what moves behind the paywall, and has that line moved since launch?

I don’t see it as a hard line between free and paid, but more about depth and actionability.

X will always remain the public foundation where I share thoughts, ideas, and parts of my investment process. It’s important for people to see how we think.

The paid community is where we go deeper: full research, detailed theses, price targets, portfolio updates, trade ideas, and discussions where ideas are challenged and analyzed.

The line hasn’t really changed. The goal is to keep providing value publicly while making sure the deeper work, which requires many hours of research, has a sustainable model behind it.

What does Edelbridge Alpha look like in a year? Is the plan to stay three individual voices under one roof, or does it become something more structured — more contributors, more tooling, something else?

In a year, I hope Edelbridge Alpha becomes the best investor community on the internet.

The core will always remain the same: independent voices, original research, and investors challenging each other to make better decisions.

I do think it can grow beyond the three of us, with more high-quality contributors, experts, better tools, and a stronger community, but the bar will always stay extremely high. We don’t want to become a place with endless content. We want to stay focused on quality, actionable insights, and real alpha.

— WHAT THIS IS REALLY ABOUT —

If you had to distill your approach into one thing that separates investors who compound well over time from those who don’t — what is it?

If I had to pick one thing, it would be having the ability to adapt.

A lot of investors fail because they become attached to their opinions, their stocks, or their stories. The best investors constantly challenge themselves, learn from mistakes, and adjust when new information appears.

Conviction is important, but the ability to stay objective and change your mind when you’re wrong is what allows you to compound over time.

Looking back at the 2020 $TSLA trade next to where you are now — all-in on high-conviction AI names in 2026 — how has your actual relationship with high-risk, high-leverage bets changed, if at all?

The biggest difference is that today I have a much better understanding of risk.

The 2020 Tesla trade was a life-changing opportunity, but looking back, the amount of leverage I used was unnecessary and could have ended very differently.

Today, I still believe in taking concentrated positions when I see exceptional risk-reward, but it’s based on deeper research, a stronger framework, and better risk management. I am still aggressive, but I am much more aware of what can go wrong and constantly reassess whether the thesis is still valid.

The goal is not to avoid risk. The goal is to take calculated risks where the upside justifies them.

You’ve said you want to help create more millionaires through this. Does running a paid product sit in tension with that, or do you genuinely see them as the same mission?

I don’t see them as conflicting at all. In fact, I think they go hand in hand.

The reason I started sharing my journey publicly was because I wanted to help people learn how to think about investing and potentially improve their financial future.

At the same time, putting hundreds of hours into research, analysis, and building a community has to be sustainable. The paid product allows us to keep producing high-quality work while creating a place where serious investors can go deeper.

The mission is still the same: help people become better investors. The subscription is simply the vehicle that allows us to dedicate more time and resources toward that mission.

That’s it for this version of retail investors interviews.

I’d like to thank for taking time to tell us his story. I hope you were able to learn more about him and his journey.

Thank you!

I’m going to tell you why HDDs matter and why WDC is a company you should definitely keep track of.

In my previous article I talked about Kioxia, a Japanese pure NAND play. This time we’re traveling to the US to visit one of the oldest tech companies. We’re still in storage but we’re covering the slow tier. It provides massive space, sits far from the GPUs and CPUs, and never gets the spotlight. But it’s a must-have nonetheless.

Today’s AI buildout isn’t just a GPU/CPU story. Every training run, checkpoint, and inference log has to live somewhere, and at exabyte scale, hard drives are still the only economical answer. WDC and Seagate are the only two publicly listed suppliers left standing, both locking in multi-year contracts that didn't exist in prior cycles.

WDC enters this moment net cash positive for the first time in a decade, with record margins and a business that's finally being valued for what it actually does well.

The company posted the best quarter in its pure play HDD history: 50.5% gross margins, a net-cash balance sheet, and 45% revenue growth, and the story is far from over.

CEO Irving Tan, on the Q2FY26 call:

“We’re pretty much sold out for calendar ‘26. We have firm POs with our top seven customers. And we’ve also established LTAs with two of them for calendar year ‘27 and one of them for calendar year ‘28. Obviously, these LTAs have a combination of volume of exabytes and price.”

This level of forward visibility this industry has never had before.

How I met your HDD maker?

While I’ve known WDC almost my whole adult life, I didn’t invest before January 2026. One of my first gaming PCs in the early 2000s had Nvidia’s RIVA TNT2 GPU, and I remember running a WD HDD alongside it. It was an 80GB drive, one of the biggest on the market at the time.

I followed the IT sector for quite some time through my IT-related work, but I never imagined there would be such insatiable demand for these products. Back then, the PC industry was the main growth driver.

I bought both WD and Seagate in Q1. After a couple of months, I sold them at a decent profit to concentrate on memory manufacturers.

After the recent sell off, and after getting information about HDD demand going forward, I returned to WD. I believe the management is doing right steps to create shareholder value, and I believe WD has a lot of upside going forward.

Table of Contents

• Where it all began

• What do they actually do

• What is a hard disk drive (HDD)?

  • Memory vs. Storage

  • How does a HDD work?

  • Technology Stack

  • Vertical Integration

• Storage in AI Data Centers

  • Types of storage in AI data centers

  • Where are HDDs located in AI data centers?

• How do they make money?

• Why now?

• What just changed?

• What’s standing in the way

• Sandisk separation

• What’s it worth?

• Bottom line

  • Bear/Base/Bull Cases

  • Final Thoughts

Check my previous article about Kioxia. It is a pure-play NAND producer from Japan.

Where it all began?

Western Digital did not start as a HDD company. It was born as a chip company, but it didn’t work out.

WD was founded in 1970 in Southern California as General Digital Corporation, the company made calculator chips. By 1975 it was the largest calculator chip maker in the world. Fast forward one year later, it was bankrupt. Its biggest customer collapsed, the oil crisis hit, and WD filed Chapter 11 in 1976.

The pivot that saved the company happened the same year, 1976. WD launched its first storage product - a floppy disk controller.

Before CDs and DVDs took over, people used floppy disks. WD made controllers for these disks. I wanted to show what these are or rather, were. Many PC users today have probably never seen one.

From left to right:
8-inch: The earliest commercial format introduced by IBM in 1971.
5.25-inch: The standard PC format of the late 1970s and 1980s.
3.5-inch: The rigid plastic format that dominated the 1990s.

3.5-inch was probably the most popular one and had the size of… wait for it… 1.44MB. A single MP3 song from could not fit in.

In 1983 it won the contract to supply IBM’s PC/AT with disk controllers.

In 1988 WD bought Tandon’s drive manufacturing and became an actual hard drive maker. Komag in 2007 and Hoya in 2010 made WD vertically integrated - WD started manufacturing its own platters.

HGST, bought in 2011 for $4.8 billion, gave WD IBM's original enterprise drive engineering (via Hitachi) and helium-sealed drive tech - strengthening the high-capacity segment WD relies on today.

WD then acquired a string of flash-storage and SSD companies (sTec, Virident, Skyera, Amplidata) between 2013 and 2015. It was clear where management thought the future was.

SanDisk in 2016 was the $19 billion acquisition that marked the exclamation point on the flash thesis.

In 2025, WD admitted the flash acquisitions were not the right step. The flash business was spun off as an independent public company, resurrecting the SanDisk name (NASDAQ: SNDK). What stayed behind is today’s WDC: a pure-play hard disk drive company, the thing it had been building toward since 1988.

What do they actually do?

WDC makes hard disk drives. That’s the whole company. After the spin-off of its SanDisk flash memory business in February 2025, WDC operates as a pure-play tech company focused entirely on developing, manufacturing, and providing data storage devices based on hard disk drive (HDD) technology.

TL;DR: They develop, manufacture, and sell HDDs.

What is a hard disk drive (HDD)?

At its most basic mechanical level, an HDD provides non-volatile storage by recording magnetic information onto rapidly rotating disks, known as platters.

Today, HDDs are not just data containers, but highly engineered tech designed to solve massive-scale data challenges for cloud providers, hyperscalers, and enterprises.

Memory vs. Storage

Before we move further, let’s get the terminology right.

I’ve seen people blur the line between memory and storage. They’re not the same thing. Both relate to preserving data, but they serve different functions.

Memory (DRAM, HBM) sits next to the processor and holds data being actively worked on right now. Storage holds everything else. It holds the data not in active use but that still needs to live somewhere. Storage itself splits into two tiers:

• NAND flash (SSDs) fast, expensive.

• Hard drives - slow, cheap, enormous.

Here’s a complementary table to help you understand the difference between different types of memories and storages.

How does a HDD work?

The basic operating principle didn’t change since IBM shipped the first commercial HDD in 1956. Everything HDD makers have done since (HelioSeal, ePMR, HAMR, UltraSMR - we’ll talk more about this later) is about doing that same basic trick with smaller spots, with more platters, packed tighter, without the data degrading or getting corrupted.

An HDD stores data using millions of microscopic magnetic spots on a spinning platter. The read/write head flies just above the surface and it’s setting each spot as it writes information, or checking each spot as it reads information back.

That is all there is. It’s an enormous grid of these magnetic spots, arranged in a specific pattern. A document, a video, an AI training data set - they’re all just different arrangements of these magnetic spots on a platter.

Because HDD’s head has to physically travel to reach those spots is the reason why HDD is slower than flash storage (SSDs, NAND). Its head has to move across the platter and wait for the right spot to spin underneath it before it can read or write anything, whereas flash memory has no moving parts and can access any spot almost instantly.

This same spinning-disk mechanism is also why HDDs are so much cheaper per TB (terabyte) to manufacture at massive capacities. Due to that lower cost, HDDs still dominate bulk, cold-and-warm data storage in data centers, even in the AI era.

Technology Stack

The main challenge that WDC’s industry has is simple: they need to fit as much data as possible onto the same physical disk.

Let’s briefly talk about its tech and how it works.

Every technology that WDC is talking about in its product roadmap is really an answer to one of two separate questions:

• how do you make each bit of data smaller and more stable?

• how do you arrange those bits most efficiently across the disk?

Confusing these terms is easy and they get thrown around constantly in HDD lingo. But they each do a different job. Here’s an explainer.

Problem 1: How do you make each bit smaller and more stable?

• HelioSeal - swap air for helium inside the drive. Less drag on the spinning platters, so you can fit more platters into the same case.

• ePMR - apply a small jolt of energy while writing. Lets WD use more stable magnetic material without losing the ability to write to it. Today’s standard.

• HAMR - the next step up. Briefly heat the spot with a laser to write on even more stable material. Replaces ePMR, doesn’t stack with it.

Problem 2: How do you arrange the tracks once they’re written?

• UltraSMR - overlap tracks like roof shingles instead of laying them side by side. Squeezes out more capacity from the same platter, regardless of which recording technology wrote the data underneath.

Why they combine so easily:

• HelioSeal = the drive’s environment (container-level choice)

• ePMR/HAMR = how a bit gets written (physics-level choice, pick one)

• UltraSMR = how tracks get arranged (layout-level choice, works with either)

The tech shift which needs to be separately mentioned is called HAMR (Heat-Assisted Magnetic Recording). It briefly heats the disk with a laser while writing, which allows for much denser, more stable data storage and unlocks the next generation of huge drives (50TB and beyond). WDC is testing early versions of this technology, while Seagate is already shipping it. We’ll talk more about this later on.

Vertical Integration

When it comes to making HDDs, the hardest part isn’t assembling it. It’s the components inside.

There are three key parts that define a good HDD maker:

• the platters - magnetic discs where data lives.

• the recording heads - writers which add the data on magnetic discs.

• the mechanics that position everything with nanometer precision.

WDC makes nearly all of it itself. They make the platters, including the magnetic media coating them. They also produce the recording heads which are designed and built in-house. That’s rarer than it sounds. In most hardware businesses, the manufacturer is really an integrator, which means that they design the product and then assemble parts other companies make.

Why does this matter?

• Cost. Every component bought from a supplier includes that supplier’s profit margin. By making parts in-house, WDC keeps those margins for itself. In a business where winning means saving dollars off every single drive, this compounds across hundreds of millions of units. It is also part of how an HDD maker prints a 50.5% gross margin.

• Speed. When the people building the recording heads, the media, and the drive itself all work for the same company, a new technology doesn’t have to wait on three separate suppliers to agree on specs and timelines. That is the reason why WDC can push out a new drive capacity whenever its own engineering is ready, instead of waiting on someone else’s factory.

• Control. WDC doesn’t need to wait for suppliers to hit a quality target, fix a defect, or allocate capacity. When nearline demand exploded with AI data centers, WD didn’t have to fight anyone for components.

However, there are downsides to this too. Owning the full supply chain also means owning its fixed costs when the demand drops.

Storage in AI Data Centers

AI data centers sort their storage into layers, with each technology doing a different job. Data gets tiered based on how fast it needs to be accessed and that access speed determines where it lives and what it runs on.

In this section two questions matter: what kind of data lives where, and physically, where does that data actually sit in the building. We’ll go through both.

Types of storage in AI data centers

To balance performance, capacity, and cost, modern AI data centers divide their storage architecture into three tiers.

The highest performing tier is Online, or “hot” storage. It utilizes HBM and enterprise SSDs. Because flash memory is a lot faster than HDDS but also significantly more expensive, this tier sits directly in the compute racks and is reserved strictly for latency-sensitive workloads like real-time AI processing and active model training.

The massive middle tier is Nearline, or “warm” storage. It relies on high-capacity HDDs to act as the true workhorse of the cloud. Today, roughly 80% of all data stored in the cloud resides on HDDs. This is the most economical way to store vast and also rapidly growing pools of unstructured data like AI training datasets and inference histories.

The foundational base is Offline and Archive storage for “cold” data. Why cold? It’s dedicated to long term retention and regulatory backups. This tier relies on incredibly cheap but extremely slow magnetic tape. When we say slow, it means really slow.

Where are HDDs located in AI data centers?

From what we've seen in the previous section, data is tiered based on how fast it needs to be accessed. There's a similar split when it comes to where that memory and storage physically sit inside the data center.

On the image above, on the left side, there’s the compute rack. It holds the GPUs, with HBM sitting directly on the package for real-time processing, and local NAND/SSDs right alongside handling the actively-used “hot” data. This is reserved for fast storage, and there’s no space inside for HDDs.

On the right side, we have HDDs. They live in the storage rack, connected back to compute over a high-speed network. This is where the bulk of AI’s data actually sits (training datasets, inference histories, audit trails) because storing that volume on flash isn’t economical at scale. HDD racks are typically built as “JBOD” (just a bunch of disks) configurations, letting capacity scale horizontally as demand grows.

The storage rack is where WDC operates. Its entire product line lives in that rack down the aisle.

How do they make money

WDC operates and reports its financial results under a single reportable segment: HDD. All of the company’s revenue is generated from the sale of tangible products, with only an immaterial amount coming from services (such as post-contract support and maintenance) and intellectual property licensing.

WDC structures and analyzes its revenue across three primary end markets:

• Cloud: WDC’s largest addressable market and its primary growth driver, which represents approximately 89% of the company’s total revenue as of the third quarter of fiscal 2026. This segment represents HDDs sold to public and private cloud environments, hyperscalers, and large data center operators.

• Consumer: is about 6% of total revenue. This represents the retail market.

• Client: is the remaining 5% of total revenue. This market provides OEM customers with high-performance HDD solutions for desktop and notebook computers.

Cloud now represents roughly 89% of WDC revenue, up from the mid-80s a year earlier, and has stayed around that level recently. The growth itself accelerated from +25% YoY in Q2FY26 to +45% in Q3FY26. WDC isn't just getting bigger but it's also getting more concentrated at the same time.

WDC relies heavily on a small number of big buyers. Recently, the company’s top 10 customers accounted for roughly 71% to 77% of its total net revenue, driven largely by hyperscale cloud providers making multi-year purchase commitments.

To stabilize revenue and protect its margins, WDC has moved away from volatile spot markets, and opted to secure LTAs with its largest cloud customers that extend out to calendar years 2028 and 2029.

The management said that its capacity is fully committed for calendar 2026 and essentially sold out through 2027, with LTAs extending into 2028–2029.

Why now?

This is the thesis: AI creates data faster than the industry can afford to store it on flash.

The active working set lives in memory and on SSDs, close to the GPUs. Everything else (80% of data) needs some large and cheap storage. At exabyte scale, this means cheap hard drives.

The cost gap between SSDs and HDDs is massive, and it’s widening at high capacities.

In 2024, nearline HDDs achieved a cost of roughly $15 per terabyte, preserving an estimated 2.5x cost advantage over enterprise SSDs.

As data centers deploy massive capacities, the economics skew even further toward hard drives. At the 30TB capacity point, the price delta between HDDs and SSDs widens to roughly 3x to 4x. The flash memory continues to have consistent overall cost premiums of 6x to 10x compared to HDDs across various workloads.

The technology that was supposed to kill the HDD is unable to compete with its cost-at-scale. Cloud providers today are allocating 1/5 of their massive capex specifically to storage infrastructure to house expanding AI training sets and cold-data tiers. They are forced to rely on the cheaper HDDs.

In the Q3FY2026, WDC shipped 222 exabytes of data, up 34% YoY. The previous forecast was 23% annual exabyte growth through 2028, and the management explicitly stated on the Q3 earnings call that they now have conviction in a “greater than 25% CAGR”.

The company claims this acceleration is being driven by three distinct AI waves:

1. Training Data Retention: As models continuously retrain and reinforce learning, the underlying data generated must be kept permanently to improve model quality.

2. Agentic AI and Inferencing: Autonomous AI agents are shifting the paradigm from simple Q&A to continuous workflow execution. Every automated action and inference generates new data that must be stored.

3. Physical AI and Synthetic Data: Robotics, autonomous vehicles, and industrial systems generate massive streams of video and sensor data, which is then used to create even larger synthetic datasets for further training loops.

According to Mordor Intelligence, the global HDD market size is roughly $51.8 billion in 2026 and is projected to reach nearly $69.7 billion by 2031. The resulting ~6% revenue CAGR looks modest, but that is because cost-per-terabyte declines naturally eat into revenue growth even as exabyte volumes explode.

While TAM is growing 6%, this is more about two companies capturing nearly all of it at margins this industry has never seen before. WDC recently crossed a record 50.5% non-GAAP gross margin. This tells us that the profitability of the HDD sector has fundamentally changed.

To summarize - why now:

• Demand is compounding at 25%+ in volume terms.

• Nearline supply is sold out and allocated through calendar 2027.

• Flash memory got more expensive relative to magnetic disks.

• Only WDC and STX are capable of supplying the hyperscalers at scale.

What’s standing in the way

Losing HAMR race - HAMR is the new technology that will make HDDs cheaper and allow them to hold more data. WDC’s competitor Seagate is already selling these new drives to two big cloud companies. WDC is still in the stage of testing its HAMR drives. In order to buy time, they are adding more and more capacity out of their proven technology. For now, this is working perfectly as WDC is actually making a higher profit margin than Seagate today. However, this is a race against time. If Seagate figures out how to make its new HAMR drives cheaper than Western Digital's older drives before WDC gets its own HAMR drives to market, Seagate will steal the profitability lead. This race is the most important competitive dynamic to watch.

Read more

Welcome back to the Retail Investor Interview Series. Today, I’m sharing my interview with HatedMoats.

If you follow fundamental investing on Substack or X, you’ve likely come across his work. He launched his publication in late July 2025 with just 50 followers and no name recognition. Less than a year later, he has over 3,000 subscribers and 12,000 followers on X.

His story is compelling not because of headline numbers, but because of what it represents: someone who spent over two decades investing quietly, built a serious track record in total anonymity, and then decided to share his research - completely free and without a personal brand attached.

What I genuinely appreciate about him is his discipline. He spends three or more months analyzing a single company before publishing, holds through 20% drawdowns on positions he believes in. He also refuses to chase narratives and remains honest about when he is wrong.

Investor Snapshot

From: Central Europe

Investing since: 2002 (index investing), 2010 (individual stocks)

Substack: hatedmoats.substack.com

Website: www.hatedmoats.com

Twitter: @HatedMoats

Subscribers: 3,000+ (free)

Portfolio YTD (2026): +14.77% vs S&P 500 +5.68%

Intro

How it all started?

He has been investing since 2002 - which means he entered markets in the aftermath of the dot-com collapse, learned through 2008, and has been actively building and refining a process for over twenty years. For most of that time, he did it entirely in private.

His professional background is in medicine, not finance. He has no formal training in equity research, no Bloomberg terminal, no institutional framework. What he has is a process he built himself over two decades which consists of deep fundamental analysis, DCF modelling, and a specific lens he calls the “Hated Moat” which is high-quality businesses with durable competitive advantages that the market is temporarily mispricing due to negative sentiment.

He launched his Substack on July 28, 2025. Within days, his Novo Nordisk deep dive went viral on X with over 300,000 views, 2,000 new followers in a matter of days. He describes it as the moment that changed everything. Not because it made him feel validated, but because it showed him there was an audience for serious, long-form fundamental research that didn’t dumb things down.

Since then, he has published full deep dives on UnitedHealth, ASML, Adobe, dLocal, Salesforce, Masimo, Duolingo, PayPal, Pfizer, ServiceNow, and others. Each one represents months of work. Each one comes with a real position attached.

The philosophy

His framework is built around a single question: is this a great business that the market currently hates for the wrong reasons? If yes, he researches it until he has enough conviction to buy and then he holds, sometimes through significant pain.

His private portfolio, investing since 2010, is heavily concentrated. Alphabet alone represents over 32% of his holdings and is a position that he initiated in 2011 at a split-adjusted average of roughly $12.92, now a confirmed 20-bagger. His second-largest position, Nebius (NBIS), got there almost by accident after a strong run in early 2026.

He is currently down over 20% on Adobe and adding. He holds Novo Nordisk through a frustrating stretch of underperformance in a bull market. He doesn’t pretend conviction means comfort and he just doesn’t let discomfort override a thesis that hasn’t broken.

The model

Everything is free. No paywall, no subscription tier, at least for now. He has announced a paid “IPO” for the one-year anniversary in July 2026, but the core research has always been available to anyone. His reasoning is simple: he’s doing the research anyway. Sharing it costs him nothing extra, and the feedback loop makes him sharper.

Interview

— THE PERSON —

Q: Who is HatedMoats outside of the investing world? What does your life actually look like?

A: Outside of the investing world, I look for ways to get back inside. I’m a board certified psychiatrist and a researcher with PhD, in a senior position at university hospital. This could change soon, as I’m thinking about starting a company and going for private practice. My life can thus be a challenging mess quite soon enough.

Q: What does a normal Tuesday look like for you? How does someone who works in medicine find time to spend 300+ hours per quarter on investment research?

A: I wake up at 6am, I hate myself for going to sleep at 1am, I go to work, I get home after 3pm, and I spend time with my family and with many of my hobbies, investing research being one of them, and I go to sleep at 1am again. As for how I can spend so much time researching while working in medicine - as I said, I’m in a senior and researcher position after many, many years, and the workload and work-life balance is not as brutal as it used to be. 5 years ago, I couldn’t spend even 30 hours per quarter, though.

Q: You’ve been investing since 2002. That’s over two decades of a parallel life running alongside your medical career. At what point did you realize investing wasn’t just a hobby anymore, that it had become something serious?

A: During the dot-com crash period, I’ve read an interesting take on the very few true ways of getting rich - inheritance, lottery, owning a business or having a high paying job and investing heavily. I inherited virtually nothing and still haven’t won the lottery. While owning the business might be on the table, I went for the route of investing heavily but since my job wasn’t really a high paying one at the time, I started trading as well. Investing itself, I truly and fully realised how important and serious it is during GFC, when my mum panic-sold despite the warnings, and seeing the consequences in real time of what impact this had on her finances from around 2010 till today. At the same time, I started researching individual stocks as well, and it got a bit more serious.

Q: Most people in medicine are trained to think in probabilities, evidence hierarchies, and long time horizons. Do you think that background actually makes you a better investor and if so, where does it show up most in your process?

A: As a matter of fact, doctors, lawyers, engineers and other highly educated people can be one of the worst investors and especially traders, in my opinion. They love to be the smartest ones in the room and they take being wrong very badly. In investing (and trading), the best loser wins. What helped me become a better investor was not medicine per se, but my field - psychiatry - and it helped me immensely in understanding the psychology of the markets and human behaviour in them. I’d consider it being one of my main edges in investing.

Q: You’ve been investing since 2002. That means you’ve lived through the dot-com aftermath, 2008, COVID, and the current cycle as an active participant. Which period shaped your thinking the most and which one did you handle worst?

A: I was quite young and inexperienced during the dot-com and lost decade era so I think what shaped me the most was the GFC, and as I mentioned above, first-hand experience of panic selling and its consequences. GFC trumatised us all and I think it’s a reason why I’m generally too cautious at times when I should be much braver. That’s why I think I handled COVID crash the worst - I did the right things but at a much, much smaller scale than I should have.

Q: Outside of markets and medicine - what do you do? Is there a life that isn’t consumed by research?

A: Oh yes, and it’s mostly at sleep’s expense. Not recommended. I like reading, playing chess, watching series and films and truly love travelling!

— THE ORIGIN —

Q: You spent over twenty years investing in complete anonymity before starting the Substack. What finally made you want to share the work publicly?

A: I didn’t even know a place like Substack existed, I wasn’t (and still am not) much of an internet person generally but once I discovered it, I really like it and I saw that I’ve been doing just that for years, only without posting about it, so why not try to share and help some people.

Q: The Novo Nordisk deep dive in August 2025 got 300,000+ views and drove almost all of your early growth. You went from 50 to 2,000 followers overnight. What was that actually like and did it change anything about how you approach the research?

A: It was quite unbelievable because everywhere, I’ve been reading about how you must persevere posting to nothingness, so to speak, oftentimes for a year or more, before getting to your first 1,000 followers and that it’s the hardest part. I got there within two weeks after only posting two theses - United Health and Novo Nordisk. I just hit a good thesis at the really right time (Novo was really a hot name back then), got reposted by a bigger account and the rest was one big chunk of luck. I believe it didn’t change anything about my research approach, though.

Q: Your professional background is in medicine, not finance. How did someone with that background end up building DCF models in their spare time and spending three months researching ServiceNow’s ARR growth?

A: During my PhD, I was fascinated with statistics and I always liked maths. Our research team collaborated with many people, including psychologists but also guys from finance. I made some friends, they taught me a thing or two, and the rest was just me studying financial models and filings for years. Nothing sexy at all, just a boring grind that I enjoy.

Q: You’ve mentioned blowing up is a formative experience for most serious investors. Did you have a moment like that - a position or a period where things went badly wrong and what did it change?

Not really with investing as I already knew better and had the painful experience from trading futures. I had some legendary, life-changing trades in futures but also many painful ones. It taught me a lot, especially the “best loser wins” philosophy, that risk management is essential and that position sizing is everything.

— THE PROCESS —

Q: You spend three or more months on a single company before publishing. Most investors think they’ve done deep research after reading the earnings call transcript. Walk me through what your actual process looks like from the beginning - from hearing a name to deciding to buy.

A: Well, that would be an article-length type of answer, so long story short: I briefly read about the company, I need to understand what it does and what kind of moat it has within 10 minutes. I look at its basic financials. I look at its growth, margins, FCF, all that jazz. Then I do a valuation model. If I like what I see, I dig deep from 10-K onwards.

Q: What do you open in the first 60 seconds when a new name crosses your radar? What immediately makes you close the tab?

A: I open the latest filings and check revenue growth, margins, free cash flow, ROIC, debt and share-count dilution. I close the tab quickly when I see a structurally poor business disguised as a cheap stock, i.e. persistent cash burn, serial dilution, excessive leverage, weak pricing power, deteriorating unit economics or a thesis that depends entirely on TAM, adjusted EBITDA and management promises.

Q: What documents do you actually read before buying - 10-K, earnings transcripts, investor presentations? What’s mandatory and what’s optional?

A: Mandatory are the latest 10-K, recent 10-Qs, several earnings transcripts and enough historical financials to understand the full cycle, not just the latest good year. I also read the risk factors, segment disclosures, debt structure, SBC and capital-allocation history. Investor presentations are useful, but optional, as they are basically marketing documents. I believe that filings show what happened, transcripts explain why it happened, and valuation decides whether I buy.

Q: The “hated moat” framing requires both a genuine moat and genuine market hatred. In practice, how do you tell the difference between a moat that’s being temporarily mispriced and one that’s actually deteriorating? What does the bear case have to look like for you to abandon a thesis rather than add to it?

A: I try to separate price damage from business damage. If customer retention, pricing power, margins, market share and returns on capital remain intact, the hatred may be temporary. If those fundamentals deteriorate structurally, the moat may be eroding. The bear case must be well structured and defined in the first place. For example, I abandon the thesis when the bear case stops being cyclical and becomes structural and when the original competitive advantage is weakening or the new facts invalidate the assumptions behind my valuation.

Q: You run a DCF for almost every name you cover, but you’ve also acknowledged how sensitive the outputs are to small assumption changes. How much of your actual buy decision is the model versus qualitative judgment the model can’t capture?

A: The DCF is a discipline tool, not a thesis. It forces me to make assumptions explicit and shows what the market is pricing in. But the buy decision still depends heavily on qualitative judgment which includes moat durability, management, industry structure and whether the business can outperform the model’s assumptions. My assumptions are always rather conservative and the results serve as guide rails for my options strategies - cash secured puts and covered calls.

Q: Do you have a target price in mind before you buy, or is it purely thesis-driven?

A: I have a valuation range before buying, not really a single rigid target. The thesis tells me what I want to own and the valuation tells me what I’m willing to pay. I believe that a great business can still be a bad investment at the wrong price.

Q: How do you think about position sizing? Do you have rules for how large a single position can get and would you ever go all-in on one name?

A: I think it’s the most crucial part of investing. To me, position size should reflect both conviction and downside risk. Usually I start smaller when uncertainty is high and add as the thesis is validated or the margin of safety improves, or there is a panic nonsense selloff.

Generally, I would never go all-in on one name. Maybe Google at the right time in the past. The fact is that even the best thesis can be wrong, and survival matters more than maximising the outcome. That’s something I learnt the hard way while trading futures.

Q: Are you an average up or average down investor or do you try to size the full position upfront?

A: Usually I’m both. Adding to a position should be a matter of conviction and thesis strength, not current price alone. Also, I prefer building positions in stages rather than sizing fully upfront. I average down only when the thesis is intact and the margin of safety improves and I average up when execution strengthens the thesis.

Q: You’ve described rotating from index gains into individual names at market highs. What does that decision actually look like in practice and what triggers the rotation, and how do you decide where the capital goes?

A: I don’t rotate simply because the index is at a high. Usually, the trigger is a widening gap between index valuation and the expected return available in specific companies. I also start to dislike broad indices because soon we’ll be forced to buy ridiculously overvalued companies like SpaceX with them. So I think I’ll rotate even more. Capital then goes toward names where the moat remains intact, the bear case is understood and the margin of safety is meaningfully better than the index.

Q: You’ve been down over 20% on Adobe and you’re still adding. You held Novo Nordisk through months of underperformance in a bull market, fully aware of the opportunity cost. At what point does patient conviction become stubborn denial and how do you know which one you’re in?

A: I think that patience becomes denial when I start defending the stock price instead of reassessing the business. I stay patient while the moat, earnings power and long-term thesis remain intact. For example, I already noted that the thesis for Novo has changed, though not broken. It resulted in fair value reassessment and change from $90s to $70s. I stop adding when the position size is appropriate for me and sell when the evidence shows structural deterioration, not just weak sentiment or temporary underperformance.

Q: You use options - covered calls on NBIS, short puts on MU but you’re clearly not an options-first investor. How does that layer fit into a long-term thesis-driven framework without undermining the core position?

A: Options are a secondary tool, but an important one for cash flow and total return. I sell puts only at prices where I genuinely don’t mind owning the shares, and covered calls only when I’m comfortable reducing/closing the position at the strike. They should improve entry prices or generate income, not turn a long-term investment into a short-term trade. Of course, it can be risky and I made some mistakes, for example this year with NBIS and RKLB due to greed and their massive run-up. It cost me adequately, yet still in my opinion, selling options should be a strategy of every investor.

Q: What are your actual sell triggers? When does the thesis break for you?

A: I sell when the moat clearly deteriorates, management destroys capital, the balance sheet turns to pile of poo or the original earnings and valuation assumptions no longer hold. Price weakness alone is not a sell trigger to me, a broken thesis is. The latest example can be PayPal - chronically incompetent management that won’t learn.

Q: Do you hold cash strategically or are you almost always fully invested?

A: I have periods of raising cash strategically when broad indices are near all time highs and I don’t see any good opportunities in individual names. I have some cash for the sake of having available cash, and usually I use it for cash secured puts.

— THE PORTFOLIO —

Q: Your private portfolio is over 32% in Alphabet - a position you’ve held since 2011. Most people would have taken at least some of a 20-bagger off the table by now. Why haven’t you?

A: Because the thesis has strengthened over time, not weakened. Alphabet still has exceptional economics, a fortress balance sheet and multiple reinvestment opportunities. I genuinely can’t think of a way to justify selling the best company in the world to buy… not as good company.

Q: Nebius became your second-largest position almost by accident - it ran hard enough to displace ASML in your top five before you’d formally sized it as a core holding. Walk me through the original thesis. What did you see before the institutional validation arrived?

A: I initiated the position in November 2025, when Nebius was briefly a Hated Moat and the market was questioning its valuation, capital intensity and ability to compete in AI infrastructure. I saw a technically strong platform, competent management, substantial funding and a credible and believable path to becoming a major AI cloud provider.

Q: Compass Pathways is a very different type of bet from the rest of your portfolio- early-stage biotech, binary clinical risk, nothing resembling a traditional moat. Why does this belong in a hated moats framework at all?

A: It doesn’t fit the traditional moat framework and I don’t try to pretend that it does. The brand is “Hated Moats” but it doesn’t mean I personally invest exclusively in such companies, and when I do invest in something else, I’ll just share the thesis as well. Compass is a small, asymmetric bet in a field I understand the most.

Q: You cut PayPal after a long drawdown, calling management “ever-incompetent.” But Adobe is in a structurally similar situation - market ignoring it, AI disruption narrative lingering - and you’re adding to Adobe. What’s the actual difference?

A: The difference is that PayPal’s problem became operational and was present long enough to become structural as well. Weak execution, deteriorating competitive positioning and management repeatedly failing to rebuild confidence. Adobe still has strong pricing power, recurring revenue, high margins and deep workflow lock-in. The market fears AI disruption, but so far the moat is being tested, not broken. If there were no AI fears and Adobe were producing the earnings it does today, the stock could easily be trading in the $600s.

Q: What does the full portfolio look like today - walk me through the main positions and weights.

A: Top 5 are Google (over 28%), Nebius (almost 12%), Novo Nordisk (7.6%), Cameco (7.35%) and ASML (6.31%).

It’s also a good place for advertising and you all can check the whole portfolio out on www.hatedmoats.com. Portfolio is the main part of the website right now and everyone can check out all positions and weights!

Q: Any names you’re currently watching that aren’t in the portfolio yet?

A: Yes, many. SAP, Wolters Kluwer, Karman, Cellebrite, TransMedics, SoFi, Kraken Robotics and many more.

— THE PLATFORM —

Q: You’ve structured Hated Moats as a public company with quarterly earnings calls, guidance, and shareholder votes. It’s a clever device but it also creates accountability you could have avoided. Why go that route?

A: Because I don’t want to avoid accountability. Too many people online do, and I want to take the opposite approach by making transparency and accountability core pillars of the platform.

Q: You’ve written that Twitter is now “dead” for your type of content. You built your initial audience almost entirely on X. What’s the realistic ceiling for a long-form fundamental research publication on Substack and do you think the community there can replace what Twitter once offered?

A: Twitter provided the initial push, but overall only around 10% of my subscribers come from there. The vast majority have discovered Hated Moats directly through Substack. I genuinely don’t know where the ceiling is but I hope it is much higher. :) The Substack community is clearly better suited for long-form fundamental research in my opinion. Unfortunately, X has increasingly become a low-effort, attention-grabbing, engagement-baiting wasteland, especially since monetisation began rewarding that behaviour.

Q: You’re planning a paid tier for July 2026, the one-year anniversary. What are you actually selling? The deep dives are already free. What does a subscriber get that a free reader doesn’t?

A: Yes, “initial paid offering” should be coming soon though it needs to be postponed as I don’t want to place it rigidly by date but rather by the amount of free subscribers - not yet sure if it’s gonna be 4,000 or 5,000; as I want the launch to be properly prepared. Paid subscribers are not buying just one extra article, they are buying the full Hated Moats research ecosystem.

That will include real-time portfolio buys and sells, the complete monthly Watchlist, full access to the “Vault” (detailed guides on how the research is conducted) and research archive, a customisable Valuation Map through the Valuation Editor, valuation and Margin-of-Safety alerts, custom portfolio Heatmaps, and access to future “Moat Labs” tools as they launch. There are many ideas. The main one is to get from passive reading to active research.

Free readers get the analysis and paid members get the models, signals, tools and ongoing decision-making framework behind it.

Q: At 3,000+ subscribers and 12,000 followers, people are likely moving real money based on what you publish. What do you think about that responsibility?

A: I take it seriously, though I don’t think my work on its own should move anyone’s money. I never want readers to treat a post as an instruction to buy. My responsibility is to provide honest, transparent research and their responsibility is to make an independent decision based on their own risk tolerance, circumstances and additional sources. I should be just one piece of their research, never the “source of truth”.

— WHAT THIS IS REALLY ABOUT —

Q: If you had to distill everything since 2002 into the single most important thing that separates investors who compound well over decades from those who don’t - what is it? Not a framework, not a valuation tool. The actual thing.

A: Buy and hold. It’s that simple. Simple but definitely not easy. Best performing investors are either dead, or those who have lost access to their account.

Q: You’ve said no one cares about 10 years of 18% CAGR, and that boring tends to be best in finance. You’re trying to make rigorous, boring, long-term investing compelling enough that people actually subscribe and stay. Is that a contradiction you’ve resolved, or one you’re still living with?

A: The boring part is passive index investing, where I still believe most retail investors should have most of their money in. Our research is about the “exciting part” - individual stocks. I believe one should have a mix of them both, the “boring” and the “exciting”, and the ratio heavily depends on how much time and dedication one is willing to put in. The odds are still brutal and against retail investors outperforming indices. I’ve been outperforming the market with individual stocks for 15 years, that’s why I’m here sharing my work.

That’s it for this version of retail investors interviews.

I hope you were able to learn more about HatedMoats and his journey.

This was the third article in the series, and I hope to find more interesting investors and share their stories very soon.

Thank you!

Q3 net sales hit a record $479 million, Integrated Memory more than doubled year-over-year, and AI-driven businesses now represent nearly three-quarters of company revenue. More importantly, the company raised FY2026 guidance again and gave an early FY2027 view that points to roughly 30% growth in both revenue and non-GAAP EPS.

This article talks about what happened in Q3. It looks at what changed versus the original thesis I published recently and what needs to happen next for the bull case to stay intact.

Penguin has gained a lot of attention since its last earnings report in April. Following the release, the stock re-rated quickly, with its price surging from $20 to $60. Even so, the company’s market cap remains under $4B.

Since April 1, Penguin shares are up more than 300%.

Q2 Earnings Release Recap

Here’s a quick recap of Q2 ER to give you context for some of the topics in Q3 ER:

Penguin is going through a transition, while growing different segments of its business.

Advanced Computing: Penguin made a deliberate strategic move to move away from hyperscaler hardware sales. It’s a low margin segment which relied on a few big customers. The company is now diversifying into enterprise/Neocloud/sovereign AI, while the Penguin Edge business is winding down entirely.

Integrated Memory is the main growth driver of the company. It’s showing strong demand and favorable pricing.

The company’s new positioning is all about its “AI factory platform” spanning infrastructure, memory, software, and services, riding the shift from AI training to inference deployment.

Q3 Earnings Release Intro

The article covers the full earnings presentation, call, financial analysis, and a review of the "What to Monitor Going Forward" section from the Investment Thesis.

This review checks if the investment thesis is developing as planned, identifying any potential red flags.

If you’re new to Penguin Solutions, feel free to check the Investment Thesis article below.

Earnings Call & Earnings Presentation Recap

Let’s start.

First thing to notice is that 74% of Q3 net sales came from what Penguin calls “AI-driven business.” What does AI-driven business consist of? It’s Integrated Memory plus the non-hyperscale part of Advanced Computing, and that part grew 104% YoY. That’s the major piece of information sitting right there: they’re diversifying away from hyperscalers and growing triple digits.

Penguin’s CEO used this number to back up the quarter’s record net sales and the raised EPS outlook, calling it proof the company is turning into an “AI factory platform company.” That raise wasn’t small. Full-year guidance moved from 12% net sales growth and $2.15 non-GAAP EPS to 22% growth and $2.60 EPS at the midpoint, driven mainly by the memory business, whose full-year outlook jumped to 90-95% growth.

It's important to add that this wasn't the first guidance bump this year. Penguin raised its full-year outlook in April from 6% to 12% net sales growth, and now, this quarter, from 12% to 22%.

On the AI-driven growth, the CEO added: “This gives us confidence that the AI opportunity is expanding as enterprises increasingly adopt agentic AI workloads powered by inference at scale. Our results this quarter demonstrate that Penguin Solutions is evolving into a leading AI factory platform company. We believe we are still in the early stages of a significant long-term profitable growth opportunity.”

Let me repeat the last bit, just in case you missed it.

Here, the CEO is covering four key topics for Penguin Solutions.

AI moves to production, Agentic AI

CEO added that since Penguin’s April earnings call, the demand environment has strengthened as AI continues to transition from early prompt experimentation to production scale inference and agentic AI. Additionally, this transition is increasing the demand across the whole data center stack, not just GPUs/HBM, but also compute, memory, storage, and networking. This is increasing demand for Penguin’s memory solutions. Agentic AI/context heavy workloads need more capacity, faster context access and better orchestration.

The industry is seeing memory as the key bottleneck for large context inference.

Penguin’s MemoryAI appliances target this bottleneck, improving inference performance and token economics.

This reinforces Penguin’s core thesis: production AI needs a full-stack “AI factory” platform, connecting infrastructure, memory, software, and operations to help customers deploy and run at scale.

Our AI Factory Platform

The CEO explained how Penguin is becoming a leading AI factory platform company at the intersection of memory and AI infrastructure. Penguin’s platform combines five core elements with their partner ecosystem. We covered this in detail in the investment thesis article.

CEO talked about Penguin’s proprietary products and services and how customers increasingly need more than just AI hardware procurement. He also added that the customers need architecture that performs in production at scale, an accelerated deployment path, and an operating model that helps them generate profitable revenue or deliver operational efficiencies.

Customer and partner momentum

Penguin added four new AI infrastructure customer logos in Q3 2026.

Penguin is continuing to add new customer logos and deepen engagement with existing customers.

CEO explained how their new customer engagement often lead to longer sales cycles, but also can lead to deeper customer relationships. Then he mentioned a few examples.

• Deepgram - Optimized inference environment (Dell + NVIDIA) for voice AI; expanded in Q3 for more production capacity, plus a new ClusterWareAI + services purchase.

• Tier 1 Financial Institution - Expanded engagement adding CXL-powered MemoryAI KV cache server, ClusterWareAI software, and Dell compute.

• Haein (South Korea) - Local Neocloud CPU-as-a-service deployment, showing Penguin’s reach across varied customer environments.

• Quantitative Trading Firm - New Q3 customer win - no further detail disclosed.

• Spectra - Sovereign supercomputer built with Sandia National Laboratories and NextSilicon, proving Penguin can deliver in national-security-grade environments.

From Q3 2025 to Q2 2026, Penguin added 16 new logos, and five of those customers subsequently increased their business with Penguin.

Also, CEO mentioned two major wins for Penguin:

• Named an NVIDIA AI Factory Specialized Partner, a recognition of Penguin’s expertise in designing, building, deploying and managing full stack NVIDIA-based AI factory infrastructure.

• Recognized as the 2026 Dell Technologies Global Alliances Americas AI Partner of the Year.

Go-to-market execution

Penguin's CXL-based MemoryAI KV cache server claims 2x inference performance, 8x lower latency, and 4-5x lower cost than GPU HBM and the tier-1 financial customer added in Q2 already came back and purchased more of these in Q3.

Business Review

Penguin delivered a strong quarter across the key metrics that matter for the thesis.

Let’s look at the details:

• Record quarterly net sales of $479 million (48% YoY and 40% QoQ)

• Record Non-GAAP operating income of $64 million (67% YoY)

Key information on the slide (below) is:

Integrated Memory + non-hyperscale AI infrastructure represented 74% of total company net sales and grew 104% YoY.

This means the company is successfully pivoting away from low margin hyperscale sales and growing their AI factory business with the focus on enterprise, neocloud and sovereign AI factories.

Advanced Computing

• Net sales reached $138M (29% of total company), growing 4% YoY.

• Demand remains strong across enterprise, sovereign AI, and Neocloud sectors.

• Non-Hyperscale AI is scaling fast, with Q3 net sales up 81% YoY, now accounting for 58% of the segment’s total sales (33% YoY).

• The shift to inference and agentic AI aligns with Penguin’s AI factory platform. The MemoryAI appliance gained traction with new revenue, bookings, and a strengthening pipeline.

Integrated Memory

• Exceptional net sales of $275M, jumping over 111% YoY due to higher volume and pricing.

• Demand is structural and durable rather than cyclical, similar to what we see in memory/storage makers businesses. It’s heavily driven by agentic AI data center needs.

• CXL memory expansion cards generated new revenue, bookings, and pipeline growth.

Optimized LED

• Q3 net sales totaled $66M, 7% YoY.

• While the company prioritizes memory and AI infrastructure, the LED segment continues to be managed in a disciplined manner. It is generating positive cash flow and hitting its innovation roadmap.

Ending CEO remarks

CEO mentioned that on June 1st, the company announced an improved financial outlook. At that time, they expected full year fiscal 2026 net sales and diluted EPS to be at the high end of their previously issued ranges.

This quarter, based on Q3 results and due to strong AI driven demand, they are further increasing fiscal 2026 outlook for both net sales and diluted EPS.

This is what CEO said when he presented Penguin priorities going forward:

1. Invest in differentiated software, AI memory, and compute infrastructure solutions.

2. Execute with discipline and speed.

3. Add new logos and deepen existing customer relationships.

4. Grow the partner ecosystem and diversify the customer base.

Takeaways from CEO’s ending speech is that the AI is entering a production phase and agentic AI makes inference more persistent and demanding, several capabilities are becoming more essential.

Penguin is strategically positioned at the intersection of memory and AI infrastructure to lead the next phase of AI.

CEO ended with “We are still in the early innings of a very large growth opportunity.”

Financial Review

This review was focused on non-GAAP results.

For Q3 FY 2026, total Penguin Solutions net sales were a record $479 million, up 48% year-over-year.

Read more

When I first started monitoring memory companies, it was in May 2025. When I seriously started looking at them, it was in Q4 2025–January 2026. During this time, it was mostly Micron and Sandisk making the headlines. There wasn't a lot of talk about SK Hynix or Kioxia. While I invested in SK Hynix at that time, I didn't pull the trigger on Kioxia until May.

In May, Kioxia had the real inflection point.

May 15, 2026 earnings release.

They blew past guidance.

They had guided FY2025 revenue growth of 27.7%–33.0% and operating profit growth of 57.1%–77.0% and delivered 37% revenue growth and nearly doubled operating profit.

The forward guidance brought even more surprises. Shares jumped after the company said it expects operating profit of ¥1.3 trillion for Q1 FY2026 alone. This is more than the entire prior fiscal year’s record profit.

Let’s dive deeper to understand what’s driving Kioxia’s growth, and, most importantly, whether there’s more upside left in this stock.

Table of Contents

• How it all started?

• What is Kioxia’s business about?

• What is NVM/SSD memory?

• Why NVM and SSDs are critical today

• TAM growth and future demand

• Investor’s Day Recap

• Challenges faced by growing inference systems

• Sandisk relationship

• Financial picture

• Capital allocation and shareholder returns strategy

• Valuation

• Final thoughts

How it started?

Kioxia's roots go back to Toshiba Corporation which invented NAND flash memory in 1987. Following heavy losses at their US nuclear subsidiary Westinghouse, the memory division was spun off in 2017 as Toshiba Memory and officially rebranded as Kioxia in October 2019. In 2018, a Bain Capital-led consortium including Korean, US, and Japanese investors acquired a controlling stake in the company for approximately $18 billion. After several delays, Kioxia completed its highly anticipated initial public offering on the Tokyo Stock Exchange Prime Market in December 2024 under the ticker 285A.

What is Kioxia’s business about?

Kioxia is often referred to as the Sandisk of Japan, which already gives a hint what their business is about.

Kioxia is a pure play memory manufacturer and one of the top three NAND flash producers in the world.

The company develops, produces, and sells non-volatile semiconductor memory (NVM) and solid-state drives (SSDs). Its storage components are critical infrastructure found across multiple end-markets, feeding data into smartphones, personal computers, automotive systems, and most importantly (and focus of our story) massive enterprise AI data centers.

Kioxia operates a massive and capital heavy manufacturing primarily in Japan. As their fabrication requires billions of dollars in equipment, Kioxia shares the costs of factory construction and R&D through a 25 years joint venture partnership with SanDisk.

At the moment Kioxia's business is undergoing a major structural shift due to the generative AI boom. AI data centers and their inference servers require a lot more storage to hold large language models and AI generated outputs. To get the best out of this new business reality, Kioxia is transitioning away from relying on cyclical consumer electronics (like smartphones) and is targeting AI market with intent to increase sales for data centers and enterprises.

What is NVM / SSD memory?

Non-Volatile Memory (NVM / NAND Flash)

NVM, specifically NAND flash memory is a storage medium whose contents are permanently retained even when a device’s power is turned off.

The main value of 3D NAND flash memory is its ability to deliver a huge data storage capacity and efficient data transfer. Compared to older 2D planar NAND, modern 3D NAND significantly reduces the energy consumption required for data processing.

Solid-State Drives (SSDs)

While raw NAND chips physically store the data, an SSD represents the completely finished storage device built around them. An SSD functions by packaging these NAND flash memory chips together with a controller (System on Chip) that manages all data flow, a DRAM cache for temporary data storage and mapping tables, and a high-speed interface (like PCIe) to connect to the motherboard.

Unlike HDDs that rely on spinning magnetic platters, SSDs have no moving parts which makes them a lot more durable, quieter and energy efficient. Due to advanced controllers and high speed PCIe interfaces, SSD is able to deliver a lot faster input/output operations per second (IOPS), massive bandwidth throughput and a very low latency which eliminates the data loading wait times that usually come with mechanical drives.

Why NVM and SSDs are critical today

Both were historically used in PCs and smartphones for fast, reliable storage, but SSDs have evolved into a core component which dictates the performance of entire AI systems alongside GPUs.

As the AI market is moving from training to inference (Agentic/Physical AI) the volume of data processing is increasing exponentially. This creates a massive system bottleneck. GPUs and their built-in High Bandwidth Memory (HBM) lack the capacity to hold large datasets needed for these continuous computations. To help with this problem, data centers today are redefining SSD and are using it not only as a storage, but also as an extended memory layer for GPUs.

How they are used in AI Data Centers:

• Extended GPU Memory: Frameworks like NVIDIA’s Context Memory Storage (CMX) utilize high-bandwidth SSDs to temporarily expand the GPU’s cache. This allows the system to efficiently store past computational results on the SSD, bypassing strict GPU memory limits and keeping expensive GPU clusters fully utilized without waiting for data reloads.

• Retrieval-Augmented Generation (RAG) Servers: RAG servers must instantly search massive external databases to improve AI response accuracy. Because DRAM capacity is too limited and HDDs are far too slow, data centers are utilizing super-high IOPS SSDs to act as extended memory and achieve fast search speeds.

• Massive Output Storage: AI models generate a large amount of output data. High-capacity SSDs are being deployed as the foundational infrastructure to store this rapid surge of AI-generated content efficiently in data centers.

TAM growth and future demand

The insatiable hardware requirements of hyperscalers are currently driving AI data center build-outs faster than the NAND supply chain can keep up.

Two categories of data exist for NAND demand growth: unit/bit demand forecasts and dollar-value market size forecasts. They diverge significantly and should be reported separately.

Bit demand

The company guides for flash memory bit demand to grow from 997 exabytes in 2025 to 1,807 exabytes in 2028, a 22% CAGR. Within that total, Kioxia forecasts data center demand rising from 295 exabytes (30% of total demand) in 2025 to 909 exabytes (50% of total demand) in 2028.

Kioxia’s Integrated Report 2025 separately states an expected ~20% CAGR in flash memory demand from 2025 to 2029, citing TechInsights’ NAND Market Report Q2 2025 as the source. The two figures are consistent with each.

Dollar-value market size

These four estimates span a $19B range for essentially the same base year, and CAGR estimates range from 5.3% to 11.0%.

For comparison, Counterpoint Research reported global NAND market revenue of $46B for Q1 2026 alone, up 246% year-over-year. Annualized, that quarter alone approaches or exceeds several of the full-year 2026 estimates above.

Management’s estimation

1. Overall NAND Market Growth 22% CAGR

2. Data Center 46% CAGR

3. Inference as the Primary Catalyst 86% CAGR

We covered what is Kioxia about. We explained their products. We talked about the importance of NAND today, and estimates for its TAM expansion.

Now, let’s talk about the Investor Day. The company shared a lot about its strategy, capex, growth plans, shareholder value creation, new products, current issues and so on.

I believe there’s a great value in learning more about Investor Day topics to understand more about Kioxia’s story.

On Kioxia’s investor relations page you can find the full presentation, the transcript and investor day Q&A. I focused on what I thought was the most relevant for this investment thesis, and I did not include all the slides.

Investor’s Day Recap

Transition from a cyclical hardware manufacturer to a structural backbone of global AI infrastructure

On June 2, 2026, Kioxia hosted its Investor Day under the theme "Flash Memory Scales AI Inference". This event marked a definitive "phase change" for the company, where it outlined its transition from a cyclical consumer hardware manufacturer into a structurally stable and highly profitable backbone of global AI infrastructure.

What did they talk about during the presentation?

Kioxia is doing strategic shift in the revenue mix away from commodity smartphone and PC memory and toward Data Center and Enterprise.

The target is to have over 60% of revenue from Data Center & Enterprise, medium to long term which is quite high from where the business is today.

Why they’re focusing Data Centers and Enterprises? See on the next slide.

Total flash memory demand: 997 exabytes (2025) → 1,807 exabytes (2028), a 22% CAGR.

The growth isn’t evenly spread. Data center demand 295 EB in 2025, growing to 909 EB by 2028 is compounding at 46%, more than double the market-wide rate.

Smartphone and PC demand is described in the deck as flat or slightly declining near-term.

Data centers are roughly 30% of total flash demand in 2025 to exactly half by 2028.

Within data center demand, inference workloads are growing at an 86% CAGR, against just 16% for training. That’s the most accurate way to confirm the training-to-inference shift. Kioxia is saying that the inference alone is growing more than 5x faster than training within their own demand model.

Supply demand (left)

Kioxia’s own supply-to-demand ratio starts above 100% in Q1 2025, then falls around 93-95% through late 2025 into early 2026, and only gradually recovers back toward the 100% line by Q4 2027, without fully closing the gap in the window shown. A ratio below 100% means demand is running ahead of supply.

Management’s own label on the slide is direct: “Tight supply-demand to persist throughout FY27.”

Revenue base (right):

The bars showing 2024 and 2025 sit flat and very low. 2026E jumps sharply, and 2027E jumps even more. Chief Strategy Officer Junichiro Yaguchi said several research firms estimate NAND market revenue will be roughly four times larger in 2026 than 2025, with that trend continuing into 2027.

The two charts tell a single story from different angles. We have the shortage (left chart) as the cause, and the revenue explosion (right chart) is the effect.

Here are two slides about Kioxia’s strategy.

• Kioxia plans to spend ~¥470B ($2.89B) annually on capex over the next three years, a 60% jump from FY2025, driven by AI/datacenter NAND demand.

• Growth is being locked in via customer LTAs for stable demand, while management commits to keeping capital efficiency above the industry average.

“We are targeting an average annual reduction in front-end cost per gigabyte (GB) in the 10% range. Through the migration of BiCS FLASH™, we have factored in bit density improvements of 50% or more per each new generation.

During transitions between product generations, we don’t base our key performance indicators solely on the layer count; rather, we plan to achieve competitive cost reductions by pursuing the optimal mix of 2D shrink (lateral scaling) and layer count.”

Key info here is roughly 10% annual reduction in front-end cost per gigabyte. That’s the number that drives margin expansion regardless of NAND pricing, meaning if NAND price stays flat, it would push their profitability higher.

Each BiCS FLASH transition (as mentioned on the slide: Gen.6 to Gen.8 to Gen.10) comes with a density jump of 1.5x to 1.6x per step. More bits per die with the same manufacturing footprint and lower cost per gigabyte.

The real target is the optimal combination of layer count and 2D shrink (shrinking the cell horizontally, the older 2D-NAND lever, still in play alongside vertical stacking). That’s a deliberate move by Kioxia as they’re trying everything they can to optimize cost per bit.

Challenges faced by growing inference systems

We’ve established that data center demand is what’s driving this NAND cycle and we also covered that inference is growing faster than training. Now we’re going one level deeper. What does inference actually do to a storage system, and why does it need SSDs specifically to solve it?

Kioxia’s SSD department talked about this question during the investor presentation. Here is their answer in three parts:

• the problem inference creates

• the new SSD use cases that problem unlocks

• and the specific products Kioxia built to solve these problems

A basic AI query used to be simple. GPU ran one inference, then looks something up if needed and returns an answer. The storage didn’t stress a lot, even old HDDs were fine for the data sitting there.

Agentic AI changed things. Today, it’s not one single inference. There’s a chain of them, thinking, checking, repeating. Due to this, two problems show up.

• KV cache problem - every time a model processes information, it builds a memory of everything in the current session - this is KV cache. In a single query it’s usually small enough to sit inside GPU’s own memory. In the Agentic AI chain of events the cache keeps growing and at some point it can’t fit the GPU memory anymore. When this happens, the system has to “throw” the old cache and reload it later, and that reloading becomes a bottleneck slowing everything down.

• RAG problem - DRAM memory used for this simply doesn’t have the necessary capacity.

Kioxia has an SSD based fix for inference problems we mentioned earlier.

KV cache overflow - instead of losing that cached data or being limited by GPU memory size, move the data on an SSDs located close to the GPU. There are two technologies enable this:

• Context Memory Storage (CMX), which treats SSDs as an extension of the KV cache.

• NVIDIA’s Storage-Next, which lets the GPU talk to the SSD directly (skipping the CPU to avoid slowing down the procesS).

RAG at scale - instead of trying to force a huge knowledge base into limited DRAM, Kioxia uses SSDs that are large and fast to be able to search quickly.

This means that the SSD became an active memory system which sits between the GPUs’ tiny and fast HBM, and slow old school bulk storage. It does job that neither of those two can do at the same time - to be fast enough to keep up, but cheap enough to hold enough data.

Kioxia’s has three purpose-built SSD lines which aim to solve problems we covered on the previous slide.

• CM Series handles the KV cache overflow problem. It is high-bandwidth, built for the constant swap traffic between GPU and storage as agentic AI’s context grows.

• GP Series handles the RAG problem. It is ultra-fast, low-latency, positioned to sit right alongside HBM as a cost-effective extension of it.

• LC Series handles the data volume problem. It is built to economically absorb the flood of data AI systems generate and consume.

If you enjoyed the read so far, please like and share!

Share

Sandisk relationship

Kioxia and Sandisk have a 25+ strategic partnership centered around a massive manufacturing and R&D JV. While they collaborate closely on production, they also operate as direct competitors in the market for flash memory and storage products.

Here are the key details of their relationship:

The manufacturing JV

• Kioxia and SanDisk jointly invest in manufacturing equipment installed at Kioxia’s Yokkaichi and Kitakami plants in Japan. They are splitting capex for semiconductor fabrication which is giving both of them ability to manufacture at a scale, something that would be difficult to achieve alone.

• R&D and its costs for NAND flash memory are shared equally, while Kioxia actually owns the physical facilities and controls 100% of the manufacturing operations.

• The JV entities sell 50% of their processed wafers to Kioxia and 50% to SanDisk. However, because Kioxia also possesses its own independent capacity, the overall production capacity at the these facilities is split 60% for Kioxia and 40% for SanDisk. The two companies account for 30% of global NAND production.

Agreement Extension

• On January 29, 2026, the companies announced a formal five-year extension of their joint venture agreements for the Yokkaichi and Kitakami plants, which will now run through December 31, 2034.

• As part of the contract extension, SanDisk is paying Kioxia $1.165 billion in cash. This compensates Kioxia for manufacturing services and guarantees long-term product supply allocation for SanDisk.

Financial picture

Kioxia indeed reminds me of SanDisk. Not just due to its business, but also due to its hyper-growth, both on the business side and in its stock price.

Read more

After looking into the company, I think Penguin is more interesting than it first appears

Penguin looks like a memory company to a lot of people, but it’s a lot more than that. The company has evolved into an AI infrastructure platform that combines memory engineering, AI factory deployments, cluster management, and managed services all under one roof.

Penguin is positioning itself as a builder and operator of AI factories, rather than just selling hardware.

As AI adoption moves beyond model training and into production, enterprise customers need infrastructure that can be deployed, managed, and optimized for inference at scale. That market shift plays directly into Penguin’s strengths.

If management executes, Penguin could become one of the more differentiated infrastructure companies in the AI ecosystem.

The next quarterly earnings is on July 7, 2026. Just enough time to learn more about the company.

If AI infrastructure interests you, I've previously covered $CRDO and $AAOI. Links below.

Table of Contents

1. Company Overview

• Brief history

• What they do

• Revenue structure

2. What PENG Does and How It Wins

• Where is the market going today?

• The problem they solve

• Products

• Why PENG Wins

• Risks

3. Financials

• Revenue

• Gross margin

• Valuation

4. My Thoughts

• What to monitor going forward

• Summary

• Bear/Base/Bull Cases

• Why I’m bullish

Stock Snapshot Today

Company Overview

Brief History

Founded in 1988 as SMART Modular Technologies, the company acted as a specialty memory assembler for the first three decades. They designed and packaged DRAM modules and flash storage for OEMs in industrial, telecom, and networking markets. The company went public in 2017 as SMART Global Holdings (SGH).

The most important step for our story is in 2018 when the company acquired Penguin Computing. Penguin at the time was an HPC and AI infrastructure company which was building clusters for government, research, and enterprise customers.

In 2024, the company was officially rebranded as Penguin Solutions, Inc., flipping the ticker from SGH to PENG.

In 2025, the company redomiciled from the Cayman Islands to Delaware.

To summarize: a memory business acquired an AI cluster company, spent six years running them in parallel, and is now betting that the mix of the two is exactly what the inference era needs.

Global footprint of 2,900 employees across corporate, R&D, and manufacturing sites.

What they do?

Penguin describes itself as “a leading provider of memory and AI infrastructure, powering the AI factories of the future for enterprises, sovereign AI initiatives, and neocloud providers.”

That explains it all. It's a business with memory and AI infrastructure under one roof.

Penguin has three business segments.

Advanced Computing

The first is building and running AI factories for companies that can't or won't do it themselves - governments, banks, universities, and enterprises that need their own AI infrastructure but don't have the ability to build it. Penguin does everything for them: designs it, builds it, deploys it, and manages it on their behalf.

Advanced Computing is built around three brands.

• Penguin Computing is the AI factory business. When a company, university, or government agency wants their own AI infrastructure but can’t build it themselves, they call Penguin. The reason is usually the same. They want to avoid placing sensitive data on someone else’s cloud. It designs, builds, deploys and manages the entire cluster on client’s behalf, which means thousands of GPUs fully integrated with networking, storage, cooling, and cluster management software. Some of their customers are Georgia Tech, Meta, and Deepgram, government agencies and neoclouds.

• Stratus is the fault-tolerant computing business. While regular servers can fail and for most businesses that's just an inconvenience but for a bank processing thousands of transactions per second or a hospital running critical equipment it's catastrophic. This business builds servers which duplicate everything internally, then they run two systems at the same time so if one side fails, the other takes over instantly with zero downtime. That is how they get 99.9% uptime. Their platforms are used by financial services, oil & gas, healthcare, and manufacturing customers.

• Penguin Edge was the edge computing business. They were running compute closer to where data is generated (e.g. inside retail stores, factories), rather than sending everything back to a central data center. It was a low margin and crowded space with competitors like Dell and HPE offering the same thing. Management decided it was consuming resources without contributing to the core thesis and is winding it down.

Integrated Memory

The second is selling specialised memory. This is not the type of memory you find in your PCs, laptops, phones. It’s custom, high-performance memory modules built for specific industrial, government, and data centre applications. This business segment is recovering due to an exploding AI demand for memory. Their newest product is directly focused on the memory bottleneck which is slowing down AI inference at scale.

It operates under the SMART Modular Technologies brand. Their product portfolio consists of specialty DRAM modules, solid-state storage, and CXL-based memory expansion solutions which were built specifically for AI workloads.

Customers include data centers, networking OEMs, and enterprises running AI workloads. They don’t just sell the memory modules. They also handle the entire supply chain around them. They source components, manage inventory, assembling products together, and program the firmware before shipping. For customers this matters a lot. They don’t need the memory modules alone, they need them configured, tested and ready for use.

Optimized LED

The third is a legacy LED lighting business which was inherited through past acquisitions. It’s not related to AI, and is declining in revenue. After years of cost cutting it became profitable.

It operates under the Cree LED brand, selling application-optimized LED chips for general illumination, automotive, video displays, and specialty lighting. Management acknowledges it’s non-core and it receives minimal capital attention. It generates cash and stays out of the way.

Revenue structure

Let's look at the revenues from the last three quarters from the Q2 FY2026 earnings presentation.

It's important to note that Penguin is going through a deliberate transition. On one side, they are winding down low-margin, low-quality revenue, and on the other side doubling down on higher-margin, stickier enterprise and sovereign AI deployments. This creates a situation where some headline numbers look worse than they actually are. Revenue is down, but the quality of what remains is improving - and that's what matters for their future.

Let’s briefly talk about each revenue segment.

Advanced Computing

By comparing Q2 FY2025 and FY 2026 it would seem as the situation is alarming, but it isn’t. It dropped down from $200M to $116M due to management’s deliberate strategic exit, and not due to lack of demand or partnership issues.

There are two parts of this segment that are being removed.

Hyperscalers hardware part. Penguin’s previous focus was selling large hardware deployments to hyperscalers. This business was low margin and dependent on a few huge customers who could walk away at any time. Management decided to stop chasing it and instead diversify their customer mix.

Penguin Edge part. Another low margin edge computing business sitting inside Advanced Computing, being wound down and expected to be fully gone by end of FY2026.

These two exits are causing a temporary revenue decline in exchange for a higher quality business going forward.

Today, non-hyperscale AI and HPC deployments (enterprises, neoclouds, and governments) represent more than 40% of Advanced Computing segment revenue, up from roughly 20% a year ago.

The decline is real, but the reason matters. It’s a deliberate pivot rather than demand failure.

Integrated Memory

This part of the business is doing heavy lifting today. Revenue went from $105M in Q2 FY2025 to $172M in Q2 FY2026 or 63% YoY. Now it represents 50% of total company revenue, up from 29% a year ago.

The growth here is being driven by DRAM and flash demand, favorable pricing, and their strong execution across networking and data centre customers. Management guided this segment to grow 65-75% growth. They also added that the growth is constrained not by demand but by supply.

To sum it up - they could grow faster if they could source more products. Similar story across the AI theme.

Their long term opportunity within this segment is CXL. This product targets a fundamentally different and higher-margin market than commodity DRAM modules. We’ll talk more about it in the products section.

Optimized LED

This segment is more or less flat for three quarters now. It’s slowly declining, with minimal investments, and management acknowledged that it’s not a priority of their core operations.

For now, it generates cash and it’s out of the way. It’s unclear if the management will divest it in the future.

What PENG Does and How It Wins

Where is the market going today?

In their investor presentation, management presented four macro shifts they believe are reshaping the AI infrastructure market. These shifts are moving in the direction of Penguin. Understanding them is important because they explain why Penguin's combination of AI factory expertise and memory engineering is relevant right now.

AI Moving from Experimentation to Production

Companies are moving past their initial pilot programs and now going into full on-premise deployments. This is happening because their business need reliable, high performance infrastructure in order to realize return on AI investments.

This benefits Penguin greatly. Penguin is acting as a master architect as it provides end-to-end services to design, build, and manage complete AI factories.

AI Workloads Are Evolving

The AI industry is moving from compute heavy training to inference and agentic AI. Companies are actively using AI in their processes, workflows, and tasks, which requires systems that can process data and respond to users quickly.

Penguin benefits directly from this shift. Inference workloads are memory bound and sensitive to latency, which is exactly where Penguin’s expertise lies. By launching products like the MemoryAI KV Cache Server, Penguin directly addresses the memory bottlenecks that emerge during inference by keeping GPUs fed with data.

Adoption Expanding Beyond Hyperscalers

For a long time, AI infrastructure was almost exclusively tied to hyperscalers. Today that is changing. Enterprises, neoclouds, and governments are all building their own AI factories, and these customers need secure infrastructure which is tailored for their specific needs, and not a generic cloud solution.

This is exactly the market Penguin is pivoting to. It is winning contracts that are sticky and high margin. It is providing the integration, software, and managed services these organisations need to run their own infrastructure. Unlike hyperscalers these companies do not have in-house expertise and rely on Penguin to do all their work.

Increased Demand for Memory-Centric Architecture

AI data center’s constraint is mainly memory capacity, rather than just GPU speed. They are running continuous inference with large context windows and thousands of simultaneous users quickly overflows the limited memory physically attached to GPUs, creating a severe “memory wall”. Penguin is well positioned to benefit from this through its CXL pooled memory appliances.

The problem they solve

The Memory Wall

Today, AI adoption is maturing and workloads are moving from compute-heavy training to inference and agentic AI. These workloads are fundamentally different, 30% compute and 70% memory driven.

During token generation, LLMs need to store huge amounts of temporary data in working memory known as the KV cache. With context windows growing and thousands of users querying the system simultaneously, the KV cache quickly overflows the fast but limited HBM attached directly to the GPUs.

When this overflow happens, systems are forced to offload data to much slower NVMe storage drives. This is creating a “triple penalty” that impacts throughput, accuracy and latency.

Penguin directly resolves this architectural bottleneck by introducing the MemoryAI KV Cache Server. This CXL-based infrastructure feeds data to GPUs up to 10 times faster than standard NVMe storage (NVMes are high-speed SSDs that GPUs are forced to fall back on when they run out of memory) and 3.8 times faster than RDMA (networking technology that pulls memory from other servers over a cable). Instead of just relying on HBM alone, companies can now draw from a much larger memory pool which provides them with faster response times, larger context windows and throughputs that their clusters were built for.

The Complexity of Building an AI Factory

Building a production-grade AI data center is way more than just buying GPUs. You have to achieve optimla performance through a finely tuned balance between compute, storage and power. Setting it up wrong creates bottlenecks which can destroy returns on what is already an expensive investment.

AI adoption moved beyond hyperscalers and now many companies, neoclouds, universities, governments are trying to have their own AI infrastructure, but unlike hyperscalers, they don’t have internal engineering teams to be able to design or deploy these kinds of environments.

Penguin acts as the key architect focused on helping solving both problems.

Instead of just selling commodity hardware, Penguin provides end-to-end services which include design, build, deployment, and ongoing management of the entire infrastructure.

Penguin is responsible for the full physical stack. Its work ranges from rack integration to network configuration. Through this work it removes the burden of complexity and allows organisations to focus entirely on their AI outputs.

Sovereignty and Data Privacy

Today, there’s a lot of focus on data sovereignty and data privacy. Governments, healthcare institutions, banks, defense companies, they all possess highly sensitive proprietary data and they cannot host them on public clouds.

These entities are driving demand for “sovereign AI”. What does it mean? It means they want localized, on-premise infrastructure that will guarantee absolute data privacy and strict compliance with respective regulations. However, to build such an infrastructure you need an infrastructure partner capable of designing and operating a massive data center from the ground up.

Penguin deliberately targets this anti-hyperscaler/sovereign AI market by building highly secure AI factories that remain under the customer’s total control. Their ability to execute is validated by their deep strategic partnership with South Korean telecommunications giant SK Telecom. Together, they successfully designed, deployed, and now manage the Haein supercomputer in Seoul. Haein is a massive sovereign AI cluster featuring over 1,000 NVIDIA Blackwell GPUs.

Unplanned Downtime

Hardware failures and unplanned downtime are incredibly costly, especially in a production environment. When AI clusters work, a silent failure or a degradation of a single GPU can disrupt training runs, corrupt data or halt company’s inference operations. For continuous agentic workflows, having 24/7 system availability is a critical requirement.

Through its Stratus brand Penguin provides fault-tolerant edge and data center computing platforms. These platforms are engineered to guarantee 99.9% availability for mission-critical applications. These systems are designed to continue operating through hardware failures by providing built-in redundancy and automated failover, minimizing downtime without requiring human intervention.

Speed to Deployment

Time-to-market is a critical competitive advantage in the AI race today. To build an AI factory it takes months of intensive work, including engineering, testing, troubleshooting. Penguin’s CEO confirmed that the “time to deployment is now directly tied to time to first token”, meaning the faster you can get your AI factory built and running, the faster your AI starts generating value.

Penguin is able to compress the deployment time by moving the engineering burden to their own facilities. Penguin performs in-factory builds, rack integration, and burn-in performance testing before the system ships to the customer. This process ensures the hardware arrives tested and ready to be activated and used.

GPU Underutilisation

While it’s simply to get GPUs, that’s only the first step. The key problem is how to maximize their utilization. Without a very effective orchestration software, GPUs can often hit communication bottlenecks, suffer from unbalanced workloads, and sit idle for long periods of time. Poor cluster management means that companies are operating their very expensive AI infrastructure at a fraction of its value and capacity.

Penguin solves this massive inefficiency by acting as the “brain” of the operation with its proprietary ClusterWareAI management software. This software provisions resources, optimizes data comms, and minimizes latency drops.

Liquid Cooling

AI accelerators generate an extreme amount of heat. Traditional passive air cooling is completely inadequate for this level of thermal density. To manage this power and thermal issue is a massive engineering problem. A lot of companies do not have an idea how to safely implement or integrate liquid cooling.

Penguin is bypassing thermal constraints through their engineering expertise in advanced liquid cooling integration. It handles the full physical stack - designing liquid cooling manifolds and power distribution directly into the cluster architecture from day one.

They have also done full immersion cooling deployments. Shell is a confirmed named customer where Penguin is powering their high-performance data centers including immersion cooling technology.

Hardware Agnostic

Enterprises don’t want to be locked into a single chip vendor. At the same time, if you mix NVIDIA GPUs, AMD accelerators, and different processors in the same cluster, this creates serious management issues. You have different software stacks, different drivers, different performance profiles all running under one roof.

ClusterWareAI solves this issue. It’s hardware-agnostic by design. It’s a single management layer that orchestrates NVIDIA and AMD hardware under one unified interface. It integrates open-source tools with Penguin’s own software.

For customers this is a real value add. They can pick the best silicon for their workload and Penguin can handle the complexity of making it all work together.

Products

The best way to cover the Penguin product portfolio is to move away from describing them as a traditional hardware component manufacturer and instead frame their offerings around their vertically integrated “AI Factory Platform”.

The Full-Stack AI Factory Platform consists of:

Layer 1 - ClusterWare Software

This is the software brain of the entire platform.

The software brain of the entire platform. ClusterWareAI is a hardware-agnostic management layer that turns bare-metal hardware, networking, and software into a single unified environment.

When a new server is added to the cluster, it automatically installs the operating system, configures the networking, and registers it as ready for work. There’s no manual setup required. It manages NVIDIA and AMD hardware under one interface and continuously monitors cluster health, automatically isolating failing components before they can interrupt active workloads.

Their moat is the operational data collected over four billion GPU runtime hours across 99,000 deployed GPUs.

Layer 2 - MemoryAI Platform

Penguin’s flagship product is the industry’s first production-ready CXL KV Cache server - pooling 3TB of DDR5 main memory with up to eight 1TB CXL expansion cards into an 11TB seamless active memory pool per node.

Instead of forcing GPUs to offload KV cache data to slow NVMe storage, the system keeps it in fast CXL memory. It is feeding data up to 10x faster than NVMe and 3.8x faster than RDMA alternatives.

Already in production at a Tier One financial institution with a second order confirmed from a generative AI company. Currently the only company offering this solution at scale.

Layer 3 - Advanced Computing Systems

ACS represents the physical hardware foundation optimized for high-performance computing and scalable AI workloads. This layer handles the physical hardware integrating AI accelerators like NVIDIA Blackwell and AMD EPYC with high-speed networking and storage into fully assembled, rack-scale systems.

Dense AI racks generate a lot of heat. From day one, Penguin engineers the power distribution and liquid cooling directly into the cluster design including full immersion cooling.

This layer rests on Penguin’s moat for their engineering expertise and conformal coating that hermetically seals memory modules so they can operate without corrosion while fully submerged in dielectric cooling fluid.

This layer also houses Penguin’s Stratus brand. Stratus provides specialized, fault-tolerant computing platforms that are engineered to guarantee 99.9% availability for mission-critical applications.

Layer 4 - OriginAI Architectures

OriginAI is Penguin’s portfolio of pre-validated AI factory blueprints.

Instead of engineering a data center from scratch, their customers get pre tested configurations covering compute, storage and networking. Configurations are all assembled and burn-tested in Penguin’s own facilities before anything ships.

Using Penguin’s AI factory blueprints results in faster deployments. Hardware arrives on data center site ready to run.

The blueprints are built on decades of real-world HPC deployment experience. Every failure mode, every performance tuning decision is added into the design before the customer ever touches it. This makes OriginAI more than just a reference design. It represents an institutional knowledge shipped in a rack.

Layer 5 - End-to-End Services

Penguin has a comprehensive professional support which spans across the entire infrastructure lifecycle - design, building, deployment. Penguin takes responsibility for the AI factory and allows their customers to focus entirely on their AI software and models rather than infrastructure management. This layer is focused on neoclouds, enterprises, governments that need production ready AI factories, but lack knowledge to build and run them.

Layer 6 - Partner Ecosystem

Penguin doesn’t manufacture everything itself. It partners with the best suppliers (Dell for servers, NVIDIA and AMD for chips) and focuses its own resources on the software, memory expertise, and integration that sits on top. SK Telecom’s $200M investment opens doors into Asian sovereign AI markets. SK hynix is co-developing next-generation memory solutions with Penguin’s SMART Modular division.

Heritage Businesses

Integrated Memory - SMART Modular Technologies
Designs and assembles specialty DRAM, flash storage, and CXL memory solutions for data centres, networking OEMs, and government customers. This is the heritage business that made MemoryAI possible. It generates steady cash flow and supply chain depth that pure-play AI infrastructure companies don’t have access to.

Optimised LED - Cree LED
Application-optimised LED chips for general illumination, automotive, video displays, and specialty lighting. Strategically non-core. Runs on a capital-light outsourced manufacturing model.

Why PENG wins

Software & Orchestration

The foundation of Penguin’s software moat is ClusterWareAI, a hardware-agnostic operating system that unifies massive bare-metal AI clusters under a single control plane. Its true competitive advantage is an operational data moat built on over four billion GPU runtime hours across more than 99,000 deployed GPUs, powering predictive maintenance algorithms that autonomously isolate failing nodes before workloads crash. This software stickiness prevents catastrophic downtime, maximizes expensive GPU utilization, and ensures customers are never locked into a single silicon provider.

Memory Engineering & Innovation

Backed by a 30-year engineering heritage from its SMART Modular division, Penguin possesses a unique capability to solve the AI inference “memory wall” where expensive GPUs sit idle waiting for data. Their flagship MemoryAI KV Cache Server is the industry’s first production-ready CXL appliance, pooling 11 terabytes of active memory to feed GPUs up to 10 times faster than standard NVMe storage.

Full-Stack Physical Infrastructure

Penguin dramatically accelerates time-to-market because they are moving the complex engineering load into its own facilities using OriginAI reference architectures that scale flawlessly from 64 to over 24,000 GPUs. To be able to handle theis heat of accelerators, Penguin is designing custom direct-to-chip liquid and full immersion cooling systems for server racks. This physical moat is heavily reinforced by conformal coating that hermetically seals memory modules. This allows them to operate safely while being fully submerged in dielectric cooling fluid without corroding or degrading.

Customer Validation

Penguin acts as the architect for enterprises and sovereign companies that demand localized and highly secure AI factories but lack massive internal engineering teams. Their operational excellence is validated by a $200 million strategic investment from SK Telecom and the successful deployment of South Korea’s 1,000-GPU Haein supercomputer from initial order to a live production environment in a very short time frame. Deep integration into global technology supply chains, including Elite Partner status with NVIDIA and a distribution agreement with CDW, cements Penguin’s position as a trusted, end-to-end platform operator rather than a commoditized hardware reseller

Risks

Here are some of the risks to be aware of:

Supply Chain Constraints
Penguin faces significant risks related to the limited availability and fluctuating costs of critical materials and components, especially in the highly constrained memory market. Management mentioned during earnings calls that they are “fighting every day for supply.” Any failure to secure components could delay new product deployments and negatively impact the overall business.

Customer Concentration & Revenue Timing
Penguin relies heavily on a small number of significant customers, making it vulnerable to customer churn and discounting pressures. This is one of the reasons they are trying to diversify their revenue and rely less on hyperscalers.

When securing new clients, Penguin must first negotiate a master agreement from zero, then wait for it to be converted into an official purchase order, and then go through the difficult process of sourcing materials and setting up the hardware. On top of that, actual revenue recognition is frequently tied to the customer’s final decisions regarding physical delivery completion and successful system launch.

Margin Compression
Penguin’s gross margins are facing pressure from the strategic pivot we covered earlier, and from the wind-down of its Penguin Edge business, which is expected to be fully discontinued by the end of FY2026. Additionally, management shared that new AI customer wins typically begin with lower margin upfront hardware sales, which can sometimes compress overall profitability.

Competition & Technological Shifts
Penguin operates in the fast-paced AI sector and faces intense competition alongside the constant threat of rapid technological change. The company’s future financial success depends heavily on its ability to avoid development delays and successfully commercialise new solutions before facing market obsolescence.

If you have read the article this far, consider liking and sharing it. Thanks!

Share

Financials

Revenue

Penguin's revenue is not a clean growth story. It is a business going through a deliberate and somewhat messy transition. It’s shedding low quality while building higher quality revenue underneath for the long term. Before we dive into numbers, it’s important to understand the context.

Read more

A very late update this time. The main reason is writing new articles. The summary of this article is concentration. Let’s dig into it.

To clarify, I’m not putting this post behind a paywall to “FOMO” anyone in. I don’t share buy/sell signals or make recommendations. I’m simply sharing my portfolio to provide transparency to paid subscribers - nothing more, nothing less.

Let’s begin.

Read more