The February Call, Six Months On
What follows is drawn from an internal thesis dated February 12, 2026. It was not written for publication, and it is not reproduced verbatim. Consistent with our standing public policy, the ratings, price targets, return estimates, and positioning recommendations that appeared in the original have been removed. Everything else is presented as it was argued in February, with its February as-of dates preserved, and then graded against what has been disclosed since.
The grades are our opinions. We assigned them ourselves, we chose the legs, and we say so plainly in the section on grading ourselves too kindly. Where a leg cannot be scored from evidence we have already published, we either flag the single outside source used or we leave the leg unscored.
Key Points
- The February 12, 2026 thesis mapped the AI buildout as a four-layer value chain (hyperscaler capital spending, silicon, neoclouds, and power) and argued the winter selloff was "rotation, not destruction." We grade eight legs of that argument below against our own published record, plus one dated outside filing, flagged.
- On our reading of the evidence, and stated as opinion: the capex, silicon, efficiency, neocloud, and power legs graded RIGHT. The February capex projection of $602 billion proved roughly 20% too low; spending is running near $725 billion for 2026, per company guidance and sell-side aggregations as of our August 30, 2026 notes.
- The software half of the paradox call graded EARLY, not RIGHT. February argued the software selloff had overshot; the software complex then printed its worst quarter since the fourth quarter of 2008 before the agent-revenue evidence turned. Direction increasingly supported, timing not.
- The costly miss graded WRONG: memory. The February silicon survey covered four logic and accelerator names and never identified memory as a constraint or named a memory maker. Memory then became the largest re-rate in the complex, a miss we first documented on our Fourth Leg page.
- The speculative flag on advanced micro-reactors graded OPEN: the risk half of "high-risk optionality" has already shown up, and the binary the February note described remains unresolved. Every grade here is an opinion, and the break conditions that would change them are listed in full.
The Scorecard at a Glance
Each leg below is presented the same way: the February claim, faithfully condensed with its February data and dates intact; the grade; and the evidence, dated and sourced. The grade vocabulary:
| Leg | The February claim, in one line | Grade (our opinion) |
|---|---|---|
| Capex engine | $602B of 2026 hyperscaler capex; structural, not speculative | Right and too low |
| Silicon and Rubin timing | Demand supply-constrained; Rubin ships volume in H2 2026, ahead of consensus | Right |
| Efficiency (Jevons) | Cheaper compute expands demand rather than shrinking it | Right |
| Neocloud layer | Contracted backlogs make hypergrowth and margins visible in advance | Right (one name unscored) |
| Power constraint | Power, not capital or silicon, is the binding bottleneck | Right |
| Speculative flag | Micro-reactors are high-risk optionality with a 2026 catalyst | Open |
| Paradox: the builders | If AI can destroy software, compute demand is by definition enormous | Right |
| Paradox: the software leg | The software selloff far exceeded any reasonable downside | Early |
| Memory | No claim was made. That is the point. | Wrong by omission |
Leg One: The Capex Engine Right
The five major hyperscalers would deploy $602 billion in aggregate capital expenditure in calendar 2026, up from an estimated $354 billion in 2025 and roughly triple the 2024 level. The cycle was structural, not speculative, for three stated reasons: paying customers (enterprise AI adoption at 72% per McKinsey's 2025 survey, up from 55% in 2023), measurable returns, and self-funding balance sheets (more than $400 billion in aggregate annual operating cash flow). Every hyperscaler had reaffirmed or raised guidance; none had pulled back.
The evidence. The projection was directionally right and roughly 20% too conservative. As we published on August 30, 2026 in The Token Tollbooth and Powering Intelligence, combined hyperscaler capital spending is running near $725 billion for 2026, up roughly 77% year over year per company guidance and sell-side aggregations, with street estimates above $1 trillion for 2027. Alphabet alone raised its 2026 guidance to $195 to 205 billion by July, per its second-quarter report as covered in our July 27 note, against the $185 billion the February table carried. The invalidation trigger the February thesis set for itself, two or more hyperscalers cutting guidance by more than 10% in a quarter, has not occurred on any disclosure we have published. Grade: RIGHT, with the caveat that the miss was to the upside.
Leg Two: Silicon and the Rubin Timing Call Right
The market leader in AI compute held roughly 85 to 90% share of training hardware and had just printed $51.2 billion of data center revenue in its fiscal third quarter, up 66% year over year. The $602 billion capex pool implied roughly $150 to 180 billion of GPU and accelerator spend. And, against a Street consensus of early 2027, the thesis argued the next-generation Rubin platform, which entered full production announcements at CES 2026, would ship in volume in the second half of 2026.
The evidence. As we published on August 30, 2026 in our note on the quarter: NVIDIA's second quarter of fiscal 2027, reported August 26, 2026, showed total revenue of $96.22 billion, up 106% year over year, data center revenue of $89.0 billion, up 117%, and third-quarter guidance of $108 billion. The Vera Rubin platform reached full production in August 2026 and is operational at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, and Nebius, per the company's disclosures. The February timing call, volume Rubin in the second half of 2026 rather than early 2027, is now a matter of record. Growth accelerated from the 66% the February thesis cited to 117%. Grade: RIGHT.
Leg Three: The Efficiency Question Right
The thesis named model efficiency its single most critical indicator. The DeepSeek debate of early 2026 posed the question directly: do efficiency gains shrink total compute demand, undermining the capex cycle, or expand the addressable market through cheaper inference? The February argument sided with the Jevons interpretation: cheaper compute has always produced more compute consumption, not less. It flagged this as the KPI bearing the closest monitoring.
The evidence. This is the leg where the six-month evidence is most lopsided. As we published on August 30, 2026 in The Token Tollbooth: Google disclosed processing 3.2 quadrillion tokens monthly in 2026, up from roughly 480 trillion in 2025 and roughly 9.7 trillion in 2024, an increase of roughly 330 times in two years. Over the same period, open-model token prices fell to roughly $0.05 to $1.20 per million tokens, down 100 to 1,000 times from 2023 levels, and agentic workloads multiplied token consumption per task by 100 to 1,000 times through chained calls and reasoning iterations. Prices collapsed, volume exploded faster, and disclosed AI revenue grew through the deflation. That is the Jevons pattern the February thesis bet on, playing out in disclosed numbers. Grade: RIGHT.
Leg Four: The Neocloud Layer Right
The thesis argued that contracted backlogs gave the independent AI cloud layer institutional-grade revenue visibility. Its exhibit was Nebius, which had reported that same day, February 12, 2026: fourth-quarter fiscal 2025 revenue of $228 million, up 547% year over year, annualized run-rate revenue of $1.25 billion, 2026 revenue guidance of $3.0 to 3.4 billion with a 40% EBITDA margin guide, and a contracted backlog above $20 billion anchored by Microsoft and Meta. The stock sold off on the print; the thesis argued the guidance and backlog, not the reaction, were the signal. It made a parallel argument for IREN's vertically integrated dual-revenue model.
The evidence. This leg is the one place we reach outside our own published pages, and we flag it: per Nebius Group's second-quarter 2026 results, reported August 12, 2026 and furnished to the SEC on Form 6-K, second-quarter revenue was $582.3 million, up 454% year over year; annualized run-rate revenue reached $3 billion at the end of June, up 598%; the adjusted EBITDA margin printed 41%, against the roughly 40% guided in February; and the full-year revenue guidance of $3.0 to 3.4 billion was reaffirmed, alongside four newly disclosed AI cloud deals the company described as averaging over $1 billion each. From our own August 30 NVIDIA note: Nebius is among the first clouds running the Vera Rubin platform in production. The February argument, that the contracted backlog made the growth and margin trajectory visible in advance, has so far tracked the disclosures almost exactly. Grade: RIGHT for the Nebius argument. The IREN half of the leg we leave unscored: our published pages carry no IREN evidence from the intervening six months, and we hold this retrospective to our own published record plus the one flagged filing above.
Leg Five: Power as the Binding Constraint Right
Power, not capital or silicon, was the binding bottleneck of the entire stack. US data center power demand would roughly triple from about 25 GW of operating capacity in 2024 toward 76 GW in 2026, and the true constraint was transmission: roughly 350 miles of high-voltage line built per year, interconnection delays of one to seven-plus years, and structural pricing power for generators with existing grid connections. The thesis catalogued a hyperscaler nuclear procurement wave of roughly 10 GW, including the 835 MW Three Mile Island restart targeted for 2027, and argued it marked a structural shift in how the largest technology companies secure energy.
The evidence. As we published on August 30, 2026 in Powering Intelligence: more than half of planned 2026 US data centers have been delayed by grid connections, interconnection queues run four to seven years in PJM, ERCOT, and the Southeast, and the Three Mile Island restart remains targeted for 2027. The procurement wave did not stall at the February catalogue; it broadened past it: a binding 1.2 GW small-modular-reactor PPA between Meta and Oklo, Amazon's $500 million X-energy investment, Google's 500 MW Kairos agreement plus 200 MW with Commonwealth Fusion, a Microsoft fusion PPA with Helion targeting 2028, and a sector-wide letter-of-intent pipeline above 30 GW. The February framing, that the bottleneck confers pricing power on connected generation and pulls hyperscalers into decades-long power contracts, is now the consensus description of the buildout. Grade: RIGHT.
Leg Six: The Speculative Flag Open
The thesis placed Oklo at "the speculative end of the nuclear-for-AI spectrum": pre-revenue, a market capitalization near $11 billion, partnerships accumulating, and a stated binary: if advanced micro-reactors prove commercially viable the addressable market is enormous; if not, "the current valuation has no fundamental support." It named a July 2026 reactor test at Idaho National Laboratory as the critical near-term catalyst, with first reactor operation targeted for late 2027 to 2028.
The evidence. Both halves of "high-risk optionality" have since shown up in the disclosures we published on August 30, 2026 in Powering Intelligence. The risk half: as of July 21, 2026, Oklo traded roughly 70% below its October 2025 peak, and its share count grew roughly 40% during 2026. The optionality half: the binding 1.2 GW Meta PPA, a disclosed customer pipeline near 14 GW, and roughly $2.5 billion of liquidity. On the July catalyst the February thesis flagged: the four reactor criticalities disclosed under the Department of Energy program by its July 4, 2026 deadline were Antares, Valar, Deployable, and Aalo; Oklo's disclosed milestones in the period were approvals of its principal design criteria and preliminary safety analysis. The binary the February note described, commercial viability proven or not, has not resolved, and first power remains a 2027 to 2028 question. We do not grade unresolved binaries. Grade: OPEN, with the observation that the February decision to fence this name off as speculative optionality rather than fold it into the core thesis has, so far, aged well.
Leg Seven: The Market Paradox
The centerpiece. As of February 12, 2026, the software complex and the AI infrastructure complex were selling off together: the IGV software index was down 24.6% for the year to date, its worst start on record, one large CRM platform had lost 43% over a year, and AI infrastructure names were falling alongside the software they threatened. The thesis called this internally contradictory: if AI is powerful enough to destroy software businesses, then demand for AI compute is by definition enormous, which validates the infrastructure buildout; if the buildout thesis is wrong, then AI is not powerful enough to threaten the incumbents. Both narratives could not be true at once. Its resolution: "rotation, not destruction," a repricing of timelines rather than of the destination, and it described the resulting dislocation, in its strongest language, as "a generational entry point." It also argued the software selloff "far exceeds any reasonable downside scenario."
The builders half Right
The evidence. The infrastructure side of the paradox resolved the way February said it would, and faster. Capex accelerated past the February projection to roughly $725 billion, NVIDIA printed a $96 billion quarter with a $108 billion guide, token volumes rose 330-fold over two years, and the power procurement wave broadened, all per the August 2026 notes cited in the legs above. Nothing we have published in six months records a hyperscaler retreat. The core logical move, that pricing software destruction concedes the compute demand case, has been borne out in the disclosed numbers. Grade: RIGHT.
The software half Early
The evidence. Here the record requires the harder grade. The February claim was not just directional; it was a claim about magnitude, that the selloff had already overshot. It had not finished. As we published on August 31, 2026 in The Fourth Leg, the IGV index closed the first quarter of 2026 down 24%, its worst quarter since the fourth quarter of 2008, with roughly $2 trillion of software market value erased, per index data and press aggregations. A reader acting on the February argument sat through the worst software quarter in eighteen years before the evidence turned. It did then turn: Salesforce disclosed Agentforce annual recurring revenue of $800 million, up 169%, on February 25, 2026; Palantir reported 85% revenue growth on May 4, 2026; Snowflake's May 27, 2026 print was followed by a 36% single-session gain; and by June our own published audit still marked the complex "hated, and inflecting" rather than recovered. Direction increasingly supported; timing wrong enough to be expensive. Grade: EARLY, and we think calling this one RIGHT would be exactly the kind of self-scoring this page exists to avoid.
Leg Eight: The One the Survey Missed Wrong
There was no claim. The February silicon survey covered four logic and accelerator names: the training leader, AMD, Intel, and Qualcomm. Memory appears in the document only as a component specification inside other companies' products, 192 GB of HBM3e cited as a competitive feature of one accelerator, 768 GB of LPDDR as a feature of another. No memory maker is named anywhere in the thesis. The ten-KPI monitoring dashboard contains no memory metric. A document that mapped the AI value chain across four layers, and built a dashboard to catch constraint shifts, surveyed the silicon layer and passed on memory entirely.
The evidence. Memory became the largest re-rate in the complex. As we documented on August 31, 2026 in The Fourth Leg: memory re-rated from late 2025 into 2026, Micron reached a market capitalization near $1.05 trillion, roughly a sevenfold re-rate within a year, SanDisk was up roughly 465% year to date by early June 2026 per market trackers, and Bank of America's May 2026 Global Fund Manager Survey, per press summaries, recorded 73% of managers calling long semiconductors the most crowded trade in markets. By August 26, 2026, the constraint had migrated into the income statement of memory's largest customer, with NVIDIA guiding gross margin to a 71 to 72% trough partly on memory costs. The Fourth Leg page owns this in the words we will keep using: "We watched the same supply-chain checks everyone else did, and we did not act on them before the move." The February thesis is where that miss began. Its framework was pointed at exactly this kind of constraint shift, and its dashboard was not wired to see the one that mattered most. Grade: WRONG, by omission, and the omission was the costly kind.
The Case That We Are Grading Ourselves Too Kindly
Steelmanned, as we do for every framework we publish:
We chose the legs, and we graded our own paper. Six of nine chips above read RIGHT. A skeptic should notice that the legs were defined after the fact by the same desk being graded, and that carving the paradox into two chips let us bank a RIGHT alongside the EARLY. A different editor could cut the same document into legs that grade worse.
The capex leg was closer to consensus than the chip implies. By February 12, 2026, every hyperscaler had already guided capex up publicly. Projecting that disclosed guidance would land is a weaker call than the RIGHT chip visually suggests. The genuinely non-consensus claims in the document were the paradox resolution and the Rubin timing, and on the paradox we were early on the half that was hardest to hold.
The evidence base is our own record. We scored against pages we ourselves published in August 2026, written with full knowledge of how the spring played out. The discipline of using published, dated pages cuts both ways: it prevents invented evidence, but it also means the grader and the evidence share an author. We flagged the one outside source used; a reader should still discount accordingly.
The miss is bigger than one row. The memory omission is not just a ninth chip. It is evidence about the framework itself: a value-chain map with a ten-KPI dashboard missed the largest single re-rate inside its own subject matter. If the dashboard could miss memory, the honest question is what the current dashboard is missing now, and no grade on this page answers that.
"Generational entry point" is not yet a fact. The February document's strongest phrase remains an opinion. Six months of supportive prints do not settle a claim of that size; the fiber and telecom cycle the thesis dismissed looked validated at month six too, and the circular-financing structures we documented in our August 30 NVIDIA note, a disclosed $108.5 billion maximum guarantee exposure and more than $500 billion of arranged third-party capital, are precisely the late-cycle features a 1999 analogy would predict. The paradox logic itself, "both cannot be true," is rhetorically clean but not airtight: software margins can erode and infrastructure can overbuild in the same decade. Markets have mispriced two things at once before.
Six months is the friendly horizon. Grading a cyclical thesis at the moment of maximum momentum flatters it. The February thesis set its own invalidation triggers precisely because cycles look best right before they do not. We will regrade this page when the evidence changes, and the section below is the contract for that.
What Would Change These Grades
These grades are opinions, and opinions with no exit conditions are marketing. The conditions below map to the break conditions we already publish and track on The Conditions · Coverage:
Capex, from RIGHT to WRONG. Two consecutive quarters of hyperscaler capex guide-downs. This is already a tracked break condition on our published frameworks, and it would retroactively convert the February capex leg from prescient to peak-cycle.
Efficiency, from RIGHT to WRONG. Token volume growth falling below token price deflation for four consecutive quarters. The Jevons argument is arithmetic, not faith: if volume stops outrunning deflation, February's most critical KPI resolves against the thesis.
Silicon, from RIGHT to weaker. A gross-margin print below the guided 71 to 72% trough, or a cut to the roughly 70% fiscal 2028 growth outlook, both tracked conditions from our August 30 note. Either would mean the demand the February thesis extrapolated is not holding at disclosed economics.
Neocloud, from RIGHT to WRONG. A miss against the reaffirmed $3.0 to 3.4 billion full-year revenue guidance, or backlog concentration resolving badly. Visibility was the February claim; guidance is where visibility either cashes or does not.
Software, from EARLY to WRONG, or to RIGHT. If agent revenue lines decelerate sharply or net revenue retention starts showing broad seat loss, the market's first-quarter verdict was correct and EARLY becomes WRONG. If the disclosed inflection compounds through fiscal 2027, EARLY hardens toward RIGHT. Both conditions are already published on our Fourth Leg framework.
The speculative flag, from OPEN to graded. First power on or off the 2027 to 2028 schedule, or a safety event anywhere in the advanced-reactor class, both tracked conditions on our energy framework. Resolution either way earns the leg a real grade.
Memory stays WRONG. No subsequent evidence regrades a miss. The row is permanent, the way the losers stay on The Record.
Public notes like this one are a fraction of what we write. Members receive our full coverage notes, quarterly outlooks, and thematic frameworks as they publish, including the retrospectives where we grade them later.
The founding rate is $99.95 per year and closes September 29, 2026. Membership details
The next note lands in member inboxes first. Founding rate locked at $99.95 per year; no payment today.
First member mailing: mid-September 2026. Unsubscribe anytime. Membership details
Sources
- Internal Three Douglas thesis, "The AI Infrastructure Supercycle," dated February 12, 2026; on file, unpublished, condensed here with ratings, targets, and positioning removed. All February figures are as stated in that document as of February 12, 2026 and were deliberately not refreshed.
- Three Douglas Research, "NVIDIA's $96 Billion Quarter: What the Print Settled, and What It Didn't" (August 30, 2026): fiscal Q2 2027 results, Rubin production status, guarantee and financing disclosures, break conditions.
- Three Douglas Research, "The Token Tollbooth: Where AI's Value Settles When the Waves End" (August 30, 2026): token volumes, token pricing, 2026 capex aggregates.
- Three Douglas Research, "Powering Intelligence: The AI Buildout's Next Bottleneck Is Electricity" (August 30, 2026): grid delays, interconnection queues, nuclear and SMR agreements, Oklo disclosures and market data, DOE program criticalities.
- Three Douglas Research, "The Fourth Leg: Where the AI Trade Goes After Memory" (August 31, 2026): the memory re-rate and the documented miss, IGV and software drawdown data, agent revenue disclosures, June 2026 crowding audit.
- Three Douglas Research, "The Three Waves of AI and the $250 Billion Question" (July 27, 2026): Alphabet Q2 2026 results and capex guidance.
- Nebius Group N.V., second-quarter 2026 results, reported August 12, 2026 and furnished to the SEC on Form 6-K: revenue, annualized run-rate revenue, adjusted EBITDA margin, reaffirmed guidance. This is the single fact source in this note drawn from outside our published pages, and it is used only for the neocloud leg.
- Market performance and survey figures (IGV quarterly performance, Micron and SanDisk market data, Bank of America May 2026 Global Fund Manager Survey) as previously published and source-flagged in our Fourth Leg note; index data, market trackers, and press summaries are secondary sources.
Important Disclosures
Not investment advice
This commentary is published by Three Douglas, LLC ("Three Douglas Research") for informational and educational purposes only. It does not constitute investment advice, a research report subject to any exchange or regulatory standard, an offer, or a solicitation to buy or sell any security. Nothing here is tailored to any reader's circumstances, objectives, or risk tolerance. Consult a qualified financial advisor before making investment decisions.
Archival content
The February 2026 document graded here was an internal thesis dated February 12, 2026 and was not written for publication. It is presented in condensed form, not verbatim. The ratings, price targets, return estimates, and positioning recommendations that appeared in the original have been removed, consistent with our standing policy of publishing no ratings, targets, or sizing. The February arguments are reproduced in order to be graded, not reissued: nothing in the condensed February text reflects a current view or recommendation, and its figures are as of February 12, 2026 only. The grades themselves are opinions of Three Douglas Research, assigned by the desk being graded, and readers should weight them accordingly.
Positions
Three Douglas, LLC, its members, and affiliated persons may hold long or short positions in securities discussed, including NVIDIA, Microsoft, Alphabet, Amazon, Meta, Nebius, IREN, CoreWeave, AMD, Intel, Qualcomm, Micron, SanDisk, Salesforce, Snowflake, Palantir, Vistra, Constellation Energy, and Oklo, and may transact in them at any time without notice. Assume we are talking our book; read accordingly.
Forward-looking statements
This commentary contains forward-looking statements, estimates, and scenario analyses, including projections of industry capital spending, compute demand, power infrastructure, and software revenue trajectories, both in the February material and in the grading commentary. All are inherently uncertain, represent assumptions as of the dates indicated only, and may prove materially wrong. Figures attributed to company disclosures, surveys, and street estimates may be revised by their sources. We undertake no obligation to update any statement, though this page's grades will be revisited if the listed conditions trigger.
Risk of loss
Investing in securities involves risk, including possible loss of the entire investment. Securities of companies discussed here are volatile and have experienced significant drawdowns within recent periods, including drawdowns documented in this very note. Concentration in a single sector amplifies risk. Past performance is not indicative of future results.
Accuracy
Information is drawn from sources believed reliable, principally our own published notes of July 27 to August 31, 2026 and the dated filings and secondary sources listed above, but is not guaranteed as to accuracy or completeness. February figures are reproduced from the internal document as of February 12, 2026 and were not refreshed. Errors and omissions are possible; corrections will be made if identified.