Methodology note: The load-bearing claims in this piece were grouped into 31 sets. Each set was assigned 3 independent verification agents (93 votes in total), instructed to "try hard to refute": go back verbatim to 10-K/10-Q/8-K filings, earnings-call transcripts, official pricing pages and original papers, and check numerators, denominators, time windows, fiscal years, GAAP vs. non-GAAP, run-rate vs. recognized revenue, and committed vs. delivered amounts. Result: none of the 31 sets received three votes of "holds as written" (only 2 sets received one or two such votes each); at the sub-claim level, more than 70 items were ruled dead or barred from use in their original form. For 3 pieces of evidence where "a single source carries a key conclusion," an additional counter-evidence search seat and a methodology audit seat were added (6 seats in total): the share of electricity in total AI data center cost was upgraded to multi-source confirmed, but the inference "so rent in the power layer is thin" was rejected by the methodology seat; "two price curves moving in opposite directions" was ruled direction-disputed and rewritten in a narrower form; "Anthropic's quarterly revenue exceeded OpenAI's" was upgraded to multi-source, but "profit has already concentrated in one company" was downgraded. Evidence grades: 【Multi-source】 ≥2 independent sources in the same direction; 【Single source, verified】 traceable to a primary document; 【Direction disputed】 independent sources conflict; 【Vendor-reported】 self-reported by an interested party, usable for direction but not for magnitude; 【Unverified】. None of the frontier labs is public; all of their financial figures come from leaked documents or media accounts and are never load-bearing. All figures are as of early October 2026 (latest quarterly reports: NVIDIA Q2 FY27 ended 2026-07-26; cloud providers Q2 2026; Micron FQ4 FY26 ended 2026-09-03).
The question "are model makers just sewing wedding dresses for others" has four popular answers: the shovel sellers win (chips and infrastructure); models will inevitably commoditize (prices fall an order of magnitude per year, open source catches up); aggregators take all (whoever holds the user squeezes upstream into a commodity); the frontier labs win, winner-take-all. Add a newer version: power and grid interconnection are the most durable scarcity. The four answers contradict each other, yet each can find numbers to support it. One reason is that they use different rulers.
This piece fixes the ruler first and uses the same one for every layer that follows:
The conclusion of this piece has two parts. The current distribution is fairly clear: profit sits overwhelmingly in the chip and memory layers, and within the chip layer, scarcity rent is migrating toward memory. The long-run landing spot is muddied by a structural fact: money circulates among the five layers. Upstream backstops the model layer with equity and credit, and the model layer's valuation flows back into upstream income statements as book gains. "Which layer makes money" has therefore partly become "who bears the model layer's credit risk."
Christensen's "conservation of attractive profits." In a one-page piece in HBR's February 2004 "Breakthrough Ideas for 2004," Clayton Christensen wrote: "When attractive profits disappear at one stage in the value chain because a product becomes modular and commoditized, the opportunity to earn attractive profits with proprietary products will usually emerge at an adjacent stage." 【Single source, verified】 In the same piece he called it "still a hypothesis" and said space constraints made it overly simplified; the name was proposed by Tensilica CEO Chris Rowen. The widely circulated version without "modular and" is a table-of-contents summary written by HBR editors, not his own words. The idea appears earlier in his book with Raynor, The Innovator's Solution (2003): when product performance is "not good enough" (a performance gap exists), "firms that build their products around proprietary, interdependent architectures enjoy an important competitive advantage against competitors whose product architectures are modular"; once performance overshoots, architectures turn modular and profit migrates to the stage where customers are still not satisfied.
Applied to AI, the theory's implication depends on its premise: if frontier model capability is still "not good enough" for customers, the frontier labs that integrate models with products should capture the profit; if it is already "good enough," models will modularize and profit will migrate to adjacent stages.
Thompson's Aggregation Theory. Ben Thompson's 2015 "Aggregation Theory" holds that the internet drives distribution and transaction costs toward zero, letting whoever holds the user relationship squeeze modular suppliers into commodities. By this theory, profit ends at the user touchpoint, not at any supply-side layer. Related to it, but not the same proposition, is Joel Spolsky's 2002 line "Smart companies try to commoditize their products' complements": if you sell A, you want its complement B to be cheap.
Thompson himself revised his view three times in 2026. The revisions themselves are a reading of how hard this question is:
Statements from interested parties should be read according to where their interests lie. Microsoft CEO Satya Nadella said on the Dwarkesh podcast in 2025-11: "I can make the argument that if you're a model company, you may have a winner's curse ... it's one copy away from that being commoditized", followed immediately by "So I think the argument can be made both ways" 【Vendor-reported】. Microsoft sells Azure, sells Copilot, builds its own models, and is an OpenAI shareholder; the more replaceable models are, the better its position. Anthropic CEO Dario Amodei, on the Cheeky Pint podcast in 2025-08, treated each model as a company ("If you consider each model to be a company, the model that was trained in 2023 was profitable"): training cost $100M in 2023 and brought in $200M of revenue, so each individual model is profitable. He called this a "cartoonish cartoon example," not the company's actual financials 【Vendor-reported】. Benedict Evans judged on 2026-07-09 that models are more likely to become "low-margin commodity infrastructure providers", "At the moment, I think every dynamic we can see points to the latter", and the next sentence is "Clearly, the situation today is transitory"; he also wrote that at present "we don't yet know of a network effect or any other winner-takes-all effect" that would let one company stay ahead 【Single source, verified】.
One frequently cited example is out of date. Meta's 2024-07 "selling access to AI models isn't our business model" was a position from the open-source Llama period; Meta has since launched the Llama API and in 2026-04 released Muse Spark without open weights. It can no longer serve as a current example of "commoditizing the complement." The current example is a different company: on 2026-09-02 NVIDIA signed a definitive agreement to acquire Hugging Face for about $11.9B (plus a retention equity plan of up to about $1B, expected to close in the first half of 2027), and the risk factors in the 8-K say "Demand for open-source foundation models and applications based on them promotes the use of our products worldwide" 【Single source, verified】. The shovel seller wants models to be cheap.
The two theories give opposite answers for AI, but they jointly point to two observable quantities: whether customers consider frontier capability already "good enough," and whether the physical bottlenecks (compute, memory, power) are lifted. Section 11 builds scenarios from these two quantities.
The PC era: components and platforms took about 7 times the profit. Per the original SEC 10-K filings: Intel's calendar-2000 net revenue was $33.726B and operating income $10.395B (30.8%, consolidated, including non-PC businesses); Microsoft FY2000 (ended 2000-06-30) revenue $22.956B, operating income $10.937B (47.6%, whole company); Dell FY2001 (ended 2001-02-02) revenue $31.888B, operating income $2.663B (8.4%, whole company, including servers and services); Compaq's 2000 commercial PC segment operating income $289M (revenue $13.136B, about 2.2%, before allocation of corporate expenses; the segment lost money in both 1998 and 1999), consumer PC segment $170M. Intel plus Microsoft is about $21.3B; Dell's whole company plus Compaq's two PC segments is about $3.1B; the ratio is about 6.8 times 【Multi-source; arithmetic】. This is not a strictly like-for-like comparison: fiscal years are offset by up to about 8 months, Intel includes non-PC businesses, and Dell is the whole company; with alternative definitions the ratio ranges from 6.2 to 7.7 times. Per the restatement in Compaq's following-year 10-K, which allocated corporate expenses to segments, its 2000 PC plus consumer business was left with only $145M (0.7%). 2000 was also the peak of the PC cycle. The direction is robust: the modular box-assembly layer was thin, and the component and platform layers that controlled proprietary standards were thick.
Fiber and railways: builders bear the risk, users take the surplus. Andrew Odlyzko predicted in 2003: "Backbone transport is likely to remain a commodity ... It is probable that backbone revenues will stay low, as the complexity, cost, and revenue and profit opportunities continue to migrate towards the edges of the network." 【Single source, verified; a prediction, couched in probabilistic language】 Global Crossing filed for bankruptcy protection on 2002-01-28 (Chapter 11 reorganization, not liquidation). Thompson wrote in 2025-11 that this fiber "basically existed for free — because the companies who built it went bankrupt", enabling today's nearly free internet. In the British Railway Mania, investors had put in a cumulative ~£250M by the end of 1850, close to half of that year's GDP; Odlyzko estimates losses of about one-third, about £80M. This is his estimate, not a measurement, and other scholars argue that financing during the expansion was rational for investors 【Direction disputed】. Paul David's 1990 study found that factory electrification did not affect manufacturing productivity until the early 1920s, about forty years after the first central power station opened—the gains ultimately went to users who reorganized their production processes, but he was discussing productivity, not the distribution of profit.
Smartphones: one integrator took most of the profit, but that was the peak. Counterpoint estimates that in 2022 Apple, with an 18% share of shipments, took 48% of global handset revenue and 85% of operating profit, all three historical highs; its revenue share has since fallen back to about 40%–46% 【Vendor-reported, research firm estimate】.
What is different this time: GPUs will not leave a cheap legacy. The "builders go bankrupt, users benefit" pattern of fiber and railways depends on one condition: long-lived assets that, after bankruptcy, can be taken over by latecomers at far below construction cost and used for decades. Thompson pointed out the AI difference in the same article: "Chips break down and get superseded by better ones; most hyperscalers depreciate them over five years, and that may be generous." 【Single source, verified】 The accounting life of GPUs is 5 to 6 years, and the economic life may be shorter. If this round of building overshoots, what is left to be cheaply taken over will be fabs, power facilities, grid-connected sites and data center shells, not the compute itself. This is where the historical analogy is most likely to fail.
NVIDIA's ledger. Per the 8-K (2026-08-26): for Q2 FY27 ended 2026-07-26, revenue was $96.221B (+106% year over year), data center $89.023B (+117% year over year), GAAP gross margin 75.0%, GAAP operating income $63.734B, operating margin 66.2% 【Single source, verified】. Annual GAAP series: FY23 gross margin 56.9%, operating margin 15.7% (including the ~$1.35B Arm acquisition termination fee); FY24 72.7%, 54.1%; FY25 75.0%, 62.4%; FY26 revenue $215.938B, 71.1%, 60.4%. FY26 gross margin declined for two reasons: the shift from Hopper boards to Blackwell full-system solutions, and charges for H20 excess inventory and purchase obligations (about $4.3B net for the year) 【Multi-source】. Starting in Q1 FY27, NVIDIA's non-GAAP no longer excludes stock-based compensation, so non-GAAP gross margins in older reports cannot be compared directly with the new basis; GAAP is used throughout here.
Scarcity rent is moving up the stack. Memory makers' margins now exceed NVIDIA's. Micron's FQ4 FY26 ended 2026-09-03 (14 weeks): revenue $54.23B, GAAP gross margin 86.8%, GAAP operating margin 80.7%; full-year FY26 GAAP gross margin 80.7% (prior fiscal year 39.8%), full-year operating margin 74.6% 【Multi-source】. SK hynix's Q2 2026 operating margin was 76%, a record for the fifth consecutive quarter. TSMC's gross margin in the same quarter was 67.7% and operating margin 60.3%.
The thickest part is not HBM. Micron's segment data show that the Cloud Memory business unit, which mainly serves hyperscalers and includes HBM, had a gross margin of 83%, actually lower than the roughly 90% of the core data center business unit and the mobile and PC business unit; management said the rising share of HBM pulled down the former's gross margin, and plans to raise HBM pricing substantially in 2027 to "narrow the profitability gap between HBM and conventional DRAM" 【Single source, verified】. What AI demand has created is shortage-driven price increases across the whole memory market; commodity DRAM and NAND capture thicker rent than HBM itself.
NVIDIA's statements corroborate this. Its formal Q3 FY27 gross margin guidance in the 8-K is 74.0% ±0.5 percentage points; CFO Colette Kress added orally on the call that because "we are experiencing extreme pricing conditions in memory", gross margin will bottom at 71%–72% in Q4 and then return to 72%–73% in FY28, on the condition "as executed price increases take effect in Q1"—that is, NVIDIA's own price increases 【Vendor-reported; earnings call only, not in SEC filings】. Its supply and capacity purchase commitments rose from $119B the prior quarter to $279B, "primarily related to the procurement of memory", spread across multiple fiscal years 【Single source, verified】. Memory price increases are squeezing NVIDIA's gross margin, and NVIDIA intends to pass part of that downstream. So this rent migration may be temporary.
Custom chips have not eliminated the chip layer's profit; they redistribute it within the layer. For Broadcom's quarter ended 2026-08-02, AI semiconductor revenue was $16.7B (+221% year over year), with next-quarter guidance of $21.7B; non-GAAP gross margin was about 75%, down 2.1 percentage points sequentially because of the rising share of AI revenue; next-quarter guidance falls to about 73% because of a rising share of custom chips with higher memory content 【Single source, verified】. Management said Anthropic will become its largest custom chip customer in 2027. AMD's data center segment had an operating margin of about 31% in Q2 2026, and AMD issued OpenAI and Meta each warrants for up to 160 million shares at an exercise price of $0.01, vesting in tranches tied to GPU purchase milestones and AMD share price targets; as of 2026-06-27, none had vested 【Multi-source】. The challengers are trading equity for orders.
Two commonly cited numbers about in-house chips must be read by their definitions. Amazon says its chip business (Trainium, Graviton, Nitro) has an annualized run-rate above $25B; this is an internal AWS attribution based on EC2 usage, not external sales revenue, and cannot be set alongside NVIDIA's or Broadcom's chip revenue 【Vendor-reported】. SemiAnalysis estimated in 2025-11 that at Google's own purchase price, TPUv7's all-in per-chip cost is about 44% lower than a GB200 server; external customers renting through Google Cloud can still save up to about 30%; and it estimated that OpenAI, without yet deploying TPUs, had already obtained about a 30% effective cost improvement across its entire NVIDIA fleet on the strength of the competitive threat. It also argued that NVIDIA chose equity investments rather than price cuts to hold onto frontier labs, to avoid gross margin erosion. These are all an analyst firm's model estimates and interpretations, which NVIDIA has publicly rebutted 【Single source; Direction disputed】. Since Q2 2026 Alphabet has begun recognizing revenue from external sales of a limited number of TPU systems; the amount is undisclosed, and the bulk will be recognized in 2027 【Single source, verified】.
The chip layer's share of total gross profit: high, but falling. Investor Apoorv Agrawal's self-built estimate from 2026-04 shows semiconductors at about 79% of AI-ecosystem gross profit dollars (infrastructure about $40B, applications about $20B, semiconductors about $225B), down from 87% two years earlier 【Single source, verified; practitioner's own estimate, application layer uses run-rate, time windows misaligned】. The direction is that the chip layer's share is being diluted, not continuing to expand.
The DeepSeek natural experiment does not show demand elasticity. On 2025-01-27, DeepSeek-R1 triggered a selloff; NVIDIA fell about 17% in a single day, erasing about $590B of market value, at the time the largest single-day market value loss for any US company 【Multi-source】. Since then its quarterly data center revenue has risen from $35.58B (Q4 FY25) to $89.02B (Q2 FY27). This trajectory is compatible with the Jevons-style reading that "efficiency gains bring more demand," but does not constitute evidence: revenue is price times quantity, and management itself says revenue opportunity per GW rose from about $18B to $25B for Blackwell and then $40B for Vera Rubin, so a large chunk of growth comes from unit value; hyperscaler capex was locked in before DeepSeek; and NVIDIA is itself financing demand (Section 10). There is no counterfactual and no price elasticity data 【Causality unverified】.
Segment operating margins are between 35% and 41%. AWS Q2 2026 revenue was $42.2B (+37%), segment operating income $16.6B, margin 39.4% (prior year 32.9%), including $551M of unrealized gains on energy contract derivatives; excluding that, about 38%–39% 【Single source, verified】. Google Cloud revenue was $24.8B (+82%), operating income $8.81B, margin 35.6% (prior year 20.7%); but the +82% is not organic growth—it includes first-time recognition of TPU system hardware revenue and Wiz, consolidated in 2026-03 【Single source, verified】. Microsoft Intelligent Cloud's FY26 Q4 (ended 2026-06-30) operating margin was 40.6%, flat year over year; gross margin fell from 60.4% to 57.1% (own calculation); Microsoft Cloud's full-year gross margin fell to 66%, and 65% in Q4, which the CFO attributed to a sales mix shift toward Azure 【Single source, verified】. AI infrastructure is pressuring gross margins; operating margins are being held up by expense leverage.
Assumption one: R&D is not allocated to segments. Alphabet had $5.79B of "Alphabet-level activities" this quarter not allocated to any segment; the 10-Q says these are "primarily" shared AI R&D (people and compute for general-purpose models), plus shared corporate costs and the like 【Single source, verified】. Neither Google Cloud's 35.6% nor Google Services' 41.8% bears the cost of training frontier models.
Assumption two: GPUs last 5 to 6 years. Microsoft extended the useful life of servers and network equipment from 4 to 6 years in 2022-07 (expected at the time to add $3.7B to FY23 operating income); Alphabet extended to 6 years in 2023 (reducing actual depreciation that year by $3.9B); Meta extended to 5.5 years starting in 2025 (reducing 2025 depreciation by $2.92B, estimated on assets placed in service before end-2024); Amazon went the other way, first extending to 6 years in 2024, then in 2025 shortening part of its servers back to 5 years citing "increased pace of technology development, particularly in the area of artificial intelligence and machine learning", increasing 2025 depreciation by $1.4B (actual figure per the 2025 10-K) 【Multi-source; SEC primary】. Michael Burry estimated in 2025-11 that hyperscalers will understate depreciation by about $176B over 2026–2028; he holds a short position, and the figure cannot be checked against financial statements 【Unverified; interested party】. Cloud-layer margins depend in part on this accounting assumption, and the economic life of GPUs gives different answers in periods of rebounding versus falling rental prices (Section 5).
Assumption three: gross margins on AI rentals will climb. For the gross margin of the GPU rental business itself, the only concrete figure comes from the media. The Information, citing internal Oracle documents, reported that for the three months ended 2025-08, Oracle's NVIDIA cloud business had sales of about $900M and a gross margin of about 14% (as relayed by CNBC) 【Single source; media】. At an analyst day 9 days later, Oracle gave an AI infrastructure "adjusted gross margin of 30%–40%"; this is a forecast target over the life of contracts, not a realized current figure, and not a response to the report 【Vendor-reported】. Quarterly reports through September 2026 still show no actual value for this margin; Oracle's "Cloud and software" segment margin (direct costs only) fell from 59.6% a year earlier to 54.5% 【Single source, verified】.
Cash flow: it cannot be said across the board to have "turned negative." Per SEC filings: Amazon's trailing-12-month free cash flow through Q2 2026 was −$7.6B; Oracle's FY26 free cash flow was −$23.7B, and −$5.4B for the quarter ended 2026-08-31 (operating cash flow included $11.4B of customer prepayments with a financing component; excluding those, about −$16.8B for the quarter); Alphabet was −$5.9B for the quarter but still +$53.3B over the trailing 12 months; Microsoft FY26 about +$67B; Meta Q2 about +$780M 【Multi-source】. Amazon (trailing 12 months) and Oracle have turned negative; Alphabet is negative only for the single quarter, and Meta's quarter is near zero but positive; Microsoft still has large positive cash flow.
The financing methods are more telling. In Q2 Alphabet received net proceeds of $49.6B from equity financing (a public offering of common stock, a Berkshire private placement, mandatory convertible preferred stock), alongside $20.3B of net debt proceeds in the same quarter 【Single source, verified】. Oracle used up its $20B at-the-market offering program in the quarter and has $288B of data center leases not yet commenced. 2026 capex guidance: Alphabet $195–205B; Amazon about $220B on a cash basis (raised from about $200B, mainly due to memory price increases); Meta $130–145B; Microsoft about $175B for the calendar year (lowered from about $190B, an accounting reclassification after leases were changed to operating leases; the CFO said the actual investment plan is unchanged) 【Vendor-reported】.
Contract backlog is about $2.46T, but it is not independent demand. Microsoft commercial RPO $678B, Alphabet revenue backlog $519.5B, AWS unrecognized long-term commitments about $496B, Oracle RPO $664B, CoreWeave RPO $103.7B: a total of about $2.46T 【Single source, verified; arithmetic】. This number cannot be treated as an "AI order pool": cutoff dates and definitions differ, it includes a large volume of non-AI contracts, the same buyers (OpenAI, Anthropic) appear repeatedly across multiple suppliers, and CoreWeave's customers themselves include hyperscalers. Microsoft's CFO said "All sequential commercial RPO growth was driven by commitments from customers outside of frontier model companies. And RPO increased 25% when excluding OpenAI"; inferring from the two growth rates, OpenAI accounts for about one-third of its commercial RPO 【Inferred】. Oracle's RPO increased by only $26B sequentially, and only about 13% will be recognized within the next 12 months. AWS's 10-Q for the same period disclosed that OpenAI added $100B over 8 years on top of the original $38B, and Anthropic added more than $100B over 10 years, but did not state whether these are fully included in the $496B.
Neocloud profits are eaten by interest. CoreWeave's Q2 2026 revenue was $2.575B (+112%); non-GAAP adjusted EBITDA $1.51B (59%), adjusted operating income only $128M (5%, prior year 16%), GAAP operating loss $49M, net interest expense $640M, GAAP net loss $626M. The gap is mainly the quarter's $1.393B of depreciation and amortization, 54% of revenue 【Single source, verified】. As of the end of June, interest-bearing debt was about $35.1B, operating lease liabilities about $16.3B, and shareholders' equity $5B. Non-convertible senior notes carry coupons of 8.5%–9.75%, plus about $6.6B of 1.75% convertible notes; in September it issued another roughly $3.7–4.2B of 2.875% convertible notes. Management guides Q3 interest expense of $860–940M, above its Q3 adjusted operating income guidance of $200–260M; 2026 capex guidance is $35–39B, about 3 times full-year revenue guidance 【Vendor-reported】. In 2025 Microsoft alone accounted for about 67% of its revenue.
This structure was deliberately outsourced by the hyperscalers. Nadella said in the same interview, "I didn't want to get stuck with massive scale of one generation", because rack power and cooling change with every generation and a generation of hardware depreciates over four or five years; when short on capacity, "we will take leases, we will take build-to-suit, we'll even take GPUs-as-a-service" 【Vendor-reported】. The depreciation risk of a single hardware generation and the concentration risk of a single customer were left with the neoclouds and their creditors.
There is no single curve for GPU rental prices. The common claim that "H100 fell from $8 an hour to $1.70, down 80%, then rebounded 60%–90%" is computed by stitching together different indices; two of three votes ruled it dead. By definition: hyperscalers' on-demand list prices had a median of about $7.92/hour in January–May 2024 and about $6.26 in the second half of 2025, a decline of only about 21% on a like-for-like basis (Silicon Data). SemiAnalysis's one-year contract price hit a low of $1.70 in 2025-10 and returned to $2.35 in 2026-03 (+38%). SemiAnalysis's own combined spot and contract index was essentially flat at about $2.8 between 2025-10 and 2026-04. Silicon Data's normalized neocloud H100 index was $2.81 on 2026-10-05, about +40% from 2025-12 【Single source each, verified】. CoreWeave reported "an approximately 25% increase across SKUs" in July, and signed an A100 contract running to 2029 (the A100 is a chip released in 2020) 【Vendor-reported】. What can be said: contract rental prices have rebounded about 40% from the 2025-10 low, and 2026 is a seller's market; it cannot be said that rental prices "crashed 80%."
Electricity is a small share of total cost, and this is multi-source. Epoch AI's 2026-05 illustrative model: a 1 GW, all-GB200 data center has an annualized total cost of ownership of about $8.5B, of which servers are about $5B (60%) and electricity about $600M (about 7%, at the average US industrial rate of 8.34 cents/kWh) 【Single source, verified; model estimate】. Other estimates found by the counter-evidence search seat all fall in the same range: SemiAnalysis estimates electricity at about 9%–11% of an H100 cluster's cost, and one GPU cloud provider, using higher power prices, computed 14%–17%; no source says electricity is a major cost of AI data centers 【Upgraded to Multi-source; range about 7%–17%】. Doubling the power price adds only about $600M to annualized cost, while changing IT equipment life from 7 years to 3 years adds about $5B.
But the inference "rent in the power layer is thin" was rejected. The methodology audit seat pointed out that a small cost share does not imply small rent; quite the opposite—precisely because electricity is negligible in the overall ledger, buyers are insensitive to power prices, and whoever controls scarce power supply has room to raise prices. A rough calculation on Epoch's own annualized basis: if a 1 GW campus is energized one year late, the opportunity cost of idle compute is about $6–7B, roughly 10 times the year's electricity bill; SemiAnalysis estimates a GW-scale AI cluster can generate $10–12B of revenue per year, and says companies choose behind-the-meter gas generation not because it is cheaper but because it comes online sooner, deliberately paying a premium to buy time 【Single source; estimate】. Nadella also said in 2025-11 that he had chips on hand but no energized data halls to plug them into. So "being able to get power right away" has a very high shadow price.
Where this rent lands can be seen only in fragments in financial statements, and most of it is not in the hands of regulated utilities:
The conclusion: the money power companies receive through electricity bills is a small share of the AI value chain; the scarcity rent of "having power at all" is real and sizable, and flows in a dispersed way to turbine makers, developers holding interconnection queue positions and energized sites, and landlords of powered data halls. Whether its total is smaller than the chip layer's is very likely so in direction, but there is no direct measurement 【Inferred】. Another piece on this site, "AI's Hardware Shortage and Power Shortage: Real, and For How Long?", discusses the scale and duration of the shortages themselves; here we look only at who takes the rent.
Revenue: two definitions, one overtaking. None of the frontier labs is public; all figures below come from company statements, leaked documents or media accounts. Anthropic's official announcements: run-rate about $9B at the end of 2025, above $30B in 2026-04, and in May "crossed $47 billion"; Bloomberg first reported on 2026-08-17 that it exceeded $65B at the end of July 【Vendor-reported; single source】. Recognized revenue: per a confidential draft S-1 first reported by Reuters on 2026-09-28, $4.59B in 2025 (about $386M in 2024); $4.73B in Q1 2026; about $11.5–11.6B in Q2, a preliminary, unaudited figure 【Single source; leaked】. OpenAI: 2025 revenue about $13.07B (leaked audited financial statements, verified by the FT); per the WSJ in 2026-08, Q1 and Q2 2026 revenue were $5.7B and $6.7B respectively 【Single source】.
Hence "Anthropic's quarterly revenue exceeds OpenAI's for the first time." The counter-evidence search seat confirmed that Bloomberg and the WSJ each saw investor materials, upgrading the Anthropic side to multi-source; but the other end of the comparison (OpenAI's $6.7B) still has only the WSJ as a source. The two companies' recognition methods also differ: Anthropic recognizes revenue sold through the AWS and Google Cloud channels on a gross basis (the leaked draft shows about 47% of 2025 revenue went through these two channels), while OpenAI recognizes the Microsoft channel on a net basis after revenue sharing. On headline figures, Anthropic is about 1.7 times OpenAI; adjusted by the channel revenue-share ratio in the draft, about 1.6 times; adjusted by the roughly 27% overstatement alleged by the OpenAI side, about 1.3 times. The lead does not flip under any definition, but its size is smaller than the headline 【R3 audit: revenue divergence upgraded to Multi-source; magnitude Direction disputed】.
Profit: only one company, one quarter, one adjusted measure. The WSJ reported that Anthropic recorded its first adjusted operating profit excluding stock-based compensation in Q2 2026, about $559M, roughly 5% of revenue; the company told some shareholders it expects Q3 to also be positive—"two consecutive quarters" was a forecast at the time of reporting, not a fact, and all three votes ruled it dead 【Single source】. The leaked draft shows Anthropic's 2025 operating loss was $8.06B (2024: $2.98B) and its GAAP net loss about $42B, of which about $34B was a non-cash revaluation of convertible financing instruments; its two largest customers each accounted for about 12% of 2025 revenue 【Single source; leaked】. OpenAI's operating loss including stock-based compensation widened from $9.3B in Q1 2026 to $12.3B in Q2, about 1.8 times that quarter's revenue (WSJ) 【Single source】. The FT reported in 2026-09 that OpenAI's investor materials forecast cumulative negative free cash flow of $278B over 2026–2030 (the May version said $305B), with revenue forecast to go from $36B in 2026 to $350B in 2030 【Single source; company forecast】.
The audit seat downgraded the conclusion here: revenue has clearly diverged among the leaders, so "the model layer as a whole is sewing wedding dresses for others" has to be tested company by company; but it cannot yet be said that profit has concentrated in any one company. Anthropic's profitability is a company-defined adjusted measure that excludes stock-based compensation and other items, with undisclosed methodology; OpenAI's loss figure includes stock-based compensation; the two sides are not measured with the same ruler.
Gross margin: almost entirely unverified. The Information reported that neither company met its own gross margin forecast for 2025: OpenAI's adjusted gross margin was about 33% (about 40% in 2024, original target 46%), Anthropic's about 40% (internal expectation about 50%, because inference costs were 23% higher than expected); but OpenAI's leaked 2025 financials from June show a gross margin of about 43%, on a different basis 【Unverified; Direction disputed】. The FT reported Anthropic's gross margin above 80%; this figure is before revenue share paid to channel partners such as Amazon, and excludes training costs 【Single source】. The "40%–50% inference gross margin" relayed by Evans includes server depreciation but not next-generation training costs that are "currently far larger than revenue" 【Unverified; he gave no source either】.
Commitments: what the model layer owes upstream. The leaked draft shows Anthropic has roughly ten-year cloud, compute and infrastructure obligations to six partners totaling about $518B, about 80% non-cancelable (Broadcom-related equipment leases $161.2B, Google $111.1B, Amazon $110B) 【Single source; leaked】. OpenAI's compute spending figures exist in several versions on different bases and cannot be strung into a time series: the roughly $1.4T Altman cited in 2025-11 (about 8 years of commitments); about $600B through 2030 told to investors in 2026-02 (CNBC); $750B reported by the WSJ in 2026-07; about $856B for 2026–2030 in July investor materials 【Single source each】.
The "model commoditization" and "frontier premium" camps have argued for two years. Part of the disagreement comes from the two camps looking at different curves; but the claim that "frontier prices are rising, not falling, symmetric with fixed-capability prices" does not hold. The curves are taken one at a time below.
Curve one: the lowest price for a fixed capability is collapsing. Epoch AI measured in 2025-03 that the per-token price to reach a given benchmark score falls by 9 to 900 times per year, with a median of about 50 times; GPT-4-level GPQA performance falls about 40 times per year. Epoch's 2026-09-22 report switched to "total cost of running the benchmark" and found that the cost of a given level of performance has fallen about 47% per quarter since 2023, about 13 times per year (with different aggregation methods, the quarterly decline ranges from 43% to 58%, i.e., about 9–30 times per year); prices fall fastest right after a model becomes the strongest, and the rate of decline halves two years later; this pattern appears on only 3 of 5 main benchmarks 【Multi-source】. Gundlach et al. at MIT FutureTech found the price of a given level of performance falls 5–10 times per year. a16z's 2024-11 "LLMflation" put it at 10 times per year, and the author acknowledged the method is "far from perfect" 【Multi-source; magnitude ranges from 5 times to more than 50 times per year, direction consistent】. This is the line the commoditization narrative describes.
Curve two: doing a task with the strongest model is getting more expensive. The abstract of the same MIT paper states: "the price of running frontier models is rising between 3× to 18× per year due to bigger models and larger reasoning demands" 【Single source, verified; sample of only 3 benchmarks over 19 months】. The main driver of the increase is the surge in tokens consumed by reasoning models, not a general rise in per-token prices. This is the line the frontier-premium narrative describes.
List prices themselves: the top tier is reaching higher, but prices have tiered. OpenAI's pricing page (input/output per million tokens, standard tier): GPT-5 $1.25/$10 → gpt-5.4 $2.5/$15 → gpt-5.5 $5/$30; as of 2026-10, the top is gpt-6-astra at $10/$50, while the same generation's mainstream tier, gpt-6-sol, costs only $2/$10, and there is also gpt-6-luna at $0.10/$0.50. Anthropic's top tier (Fable 5/5.1, Mythos 5/5.1) is $10/$50, lower than the $15/$75 of Opus 4/4.1 in 2025 【Multi-source; official pricing pages】. So "flagship prices rise monotonically with each generation" does not hold; what holds is "the top tier is reaching higher, while a cheaper mainstream tier has appeared."
Open weights lag by about 4 months; closed models take most of the revenue. Epoch measured in 2026-05 that since 2026 the strongest open-weight models have lagged the closed frontier by about 4 months on average 【Multi-source】; its earlier measurement for 2023-01 to 2025-10 averaged about 3 months. The two time windows differ, so one can only say the recent range is slightly larger, not that "the gap is widening." An SSRN working paper by Frank Nagle and Yue (introduced on the Linux Foundation blog), using OpenRouter data, says closed models account for about 80% of usage and 96% of revenue, at an average price about 6 times that of open models, costing enterprises an extra $20–48B per year (preferred estimate $24.8B); OpenRouter accounts for only about 1% of global API spending and skews toward developers 【Single source, verified; skewed sample】. An enterprise survey by Menlo Ventures (an Anthropic investor) estimates enterprise LLM spending shares at Anthropic 40%, OpenAI 27%, Google 21%, with the open-source share falling from 19% to 11% 【Vendor-reported; investor's estimate】.
Putting the two lines together, the model layer's profit depends on a demand question: does customers' money stay on the second line (paying ever-higher per-task bills for the strongest capability), or slide toward the first line (using much cheaper "good enough" capability)? Cursor's experience offers a micro sample (Section 8): it turned its gross margin from negative to slightly positive by switching the base of its in-house model to the open-source Kimi K2.5.
The incumbent giants have not been disrupted; they have become steadier. Alphabet's Q2 2026 Search & other revenue was $63.27B (+17%), Google Services segment operating margin 41.8% (prior year 40.1%), and Pichai said AI Mode has more than 1 billion monthly active users; but this margin excludes the $5.79B of Alphabet-level costs that are mainly shared AI R&D, so it is overstated 【Single source, verified】. Meta's advertising revenue was $59.36B (+27.5%); consolidated operating income was $18.78B, −8% year over year, but the quarter included $2.4B of legal expenses and $1.18B of layoff severance; excluding those, about +9%. The cost growth drivers listed in the 10-Q are, in order, employee compensation, data centers and third-party cloud services, legal expenses, and third-party AI token costs; tokens are one of four and listed last, so "AI costs are eroding Meta's profit" does not hold 【Single source, verified】. Microsoft's Productivity and Business Processes segment had a Q4 operating margin of 57.9%, with more than 30 million paid M365 Copilot seats (unit price undisclosed). Apple Services revenue was $30.74B with a gross margin of 75.6%; capex for the first nine months was $6.8B, below $9.47B in the same period a year earlier.
Apple is the sample of an "asset-light aggregator." On 2026-01-12 Apple and Google issued a joint statement that the next generation of Apple Foundation Models will be based on Gemini; Bloomberg reported before signing that Apple was close to a deal worth about $1B per year, and the formal terms are undisclosed 【Single source; media】. Compare the figure found by the court in the antitrust case: in 2022 Google paid Apple more than $20B to be the default search engine. The two payments are from different years and flow in opposite directions, so they can only be compared in order of magnitude: the party holding distribution gets far more from model makers than it pays them.
AI-native applications: fast growth, thin gross margins, constrained by upstream. Bessemer's roughly 10 fastest-growing "Supernova" AI application companies have an average gross margin of about 25%, often negative 【VC sample basis】. Per TechCrunch in 2026-04, citing people familiar, Cursor "operated at negative gross margins until recently", reaching a slim positive gross margin only through its in-house Composer plus calls to cheaper models such as Kimi, and per a single source only large enterprise customers are gross-margin positive while individual developer accounts still lose money 【Single source】. Composer 2 is built on Moonshot AI's open-source Kimi K2.5, and Cursor co-founder Aman Sanger acknowledged "It was a miss to not mention the Kimi base in our blog from the start" 【Multi-source】. The circulating claim of a "gross margin of −23% for the quarter ended January 2026" could not be traced to any original source and is not usable.
There are two cases of upstream supply cutoffs. In 2025-06, according to Windsurf CEO Varun Mohan's one-sided account, Anthropic cut off nearly all of its first-party capacity for Claude 3.x models with less than five days' notice, at a time when OpenAI was reported to be acquiring Windsurf; Windsurf was subsequently broken up, with Google taking a non-exclusive license for about $2.4B and hiring the CEO, and Cognition acquiring the rest 【Single source; party's own account】. On 2026-08-14 SpaceX completed its acquisition of Cursor, paying about 389 million SpaceX Class A shares, implying an equity value of $60B, about 20 times the roughly $3B run-rate reported by the media; after closing, OpenAI announced it would stop providing models to Cursor 【SEC primary; cutoff Multi-source media】.
Model makers are also tightening spillover from their own subscriptions. Starting in 2026-01, Anthropic technically blocked third-party tools impersonating Claude Code; in February it clarified in compliance documentation that subscription OAuth is limited to Claude Code and Claude.ai (the consumer terms themselves have prohibited unauthorized automated access since 2024-02 and were not changed); starting 2026-04-04 it first disabled subscription quotas for OpenClaw and said this would extend to all third-party harnesses 【Multi-source】. GitHub announced on 2026-04-27 that for Copilot "the current premium request model is no longer sustainable", switching all plans to usage-based billing from June 【Single source, verified】.
Model companies are eating the application layer. Anthropic self-reported in 2026-02 that Claude Code's run-rate exceeded $2.5B, with enterprises accounting for more than half, while the company's total run-rate was $14B at the same time 【Vendor-reported】. OpenAI said in August its advertising business had reached about a $1B run-rate; CFO Sarah Friar told investors that enterprise revenue now exceeds consumer revenue 【Vendor-reported; relayed by attendees】. Existing SaaS companies' AI upsell has delivered limited results: Salesforce's Agentforce ARR exceeded $1.5B for the quarter ended 2026-07-31, but starting this quarter the definition expanded to include Slackbot and others; total revenue +11%, about 6% organic growth after excluding acquisition contributions 【Single source, verified】.
Application-layer companies are not acquired at low valuations (Cursor at about 20 times run-rate), but their profit and control are constrained by upstream supply. Current AI application-layer profit lands mainly with those who already held the users.
Putting the numbers from the previous five sections on the same ruler (latest-quarter GAAP operating margin, segment basis): memory (Micron 80.7%, SK hynix 76%) > chip design (NVIDIA 66.2%) > foundry (TSMC 60.3%) > cloud segments (Microsoft Intelligent Cloud 40.6%, AWS 39.4%, Google Cloud 35.6%) > neocloud (CoreWeave −1.9%) > frontier labs (OpenAI's operating loss about 1.8 times revenue, per media reports). In the application layer, segment margins of the giants holding distribution are between 40% and 58% (and exclude AI R&D), while AI-native application gross margins range from negative to 25%.
Some caveats: fiscal periods are offset by up to about two months; Google Cloud and Services do not absorb AI R&D; cloud segment margins depend on 5.5–6-year depreciation assumptions; lab figures are media-reported. The caveats do not change the ranking. As of October 2026, the current profit distribution of the AI value chain is a line tilted upstream: the closer to physical scarcity (memory, advanced chips, foundry), the higher the margin; the closer to the model, the bigger the loss. This is the part of the "shovel sellers win" narrative that stands up. What it does not answer is what sustains this line and for how long.
How much downstream revenue is needed to hold this line up. These calculations are all "napkin math," and they get bigger every year. Sequoia's David Cahn in 2024-06 took NVIDIA's run-rate revenue forecast, multiplied by 2 (GPUs are about half of total data center cost) and by 2 again (end users need a 50% gross margin), arriving at the "$600B question"; in 2026-07, on his personal Substack, the same method gave $1.5T, meaning "how much lifetime end-user revenue is needed to pay back one year of capex," and he put the cumulative figure since ChatGPT at about $3T (which he called a rough estimate) 【Single source, verified】. Bain estimated in 2025-09 that about $500B of annual capex in 2030 would require about $2T of annual revenue, and that even assuming companies move all on-premises IT budgets to the cloud and reinvest the savings from AI, there would still be a shortfall of about $800B; its 2026-09 update estimates annual AI infrastructure spending could reach $1.5T by 2031, which, with capex at about 25% of revenue, requires an annual market of nearly $6T, while existing consumer and enterprise AI can contribute only $1.2–1.8T, a gap of $4.2–4.8T 【Consulting firm basis】. Both Bain versions use the same capex-to-revenue ratio; the gap widened because the capex forecast was raised about 3 times. These numbers are not forecasts but "how much is needed"; they show that for upstream margins to be sustained, downstream revenue must grow far faster than it does today.
If you look only at cash, the model layer is indeed working for upstream. But once equity, credit and accounting gains are drawn in, the money forms a loop.
Cash flows upstream. Microsoft recognized $24.1B of revenue from its commercial arrangements with OpenAI in FY26 (including revenue share, mostly Azure compute payments), with $6B receivable at period end 【Single source, verified】. The common claim that "Microsoft received more from OpenAI than OpenAI's revenue for a year" does not hold: the windows differ (Microsoft's fiscal year versus a calendar year), and the OpenAI revenue it was compared to is a media figure. The more accurate statement: this money is mainly paid out of OpenAI's financing, and Microsoft is itself one of the funders. Anthropic's $518B, roughly 80% non-cancelable ten-year obligation is a larger sum in the same direction.
Equity and credit flow downstream. NVIDIA's CFO said on the 2026-08 earnings call "we've invested nearly $50 billion in the frontier AI labs", and expects that lab demand supported by NVIDIA's balance sheet will "contribute toward roughly a quarter of our business next year"; she pre-emptively noted "we know some will call this circular financing. We see it differently" 【Vendor-reported; forward-looking】. NVIDIA's 10-Q says: "We believe AI clouds and AI model makers ... currently lack the ability to secure long-term infrastructure contracts and investment-grade financing capacity", so NVIDIA provides land, power, data center shells and capacity guarantees; it has $36B of cloud service purchase commitments to AI clouds (generally 6 years), and if AI clouds cannot sell the committed capacity, NVIDIA agrees to buy it 【Single source, verified】. The 2025-09 letter of intent to invest "up to $100B" was non-binding; what actually materialized was $30B of equity in OpenAI's 2026 financing, and Jensen Huang said in 2026-03 that investing $100B is "probably not in the cards" 【Multi-source】. In 2026-08 NVIDIA also provided a residual value guarantee for the PORTS-Pike campus that SB Energy leases to OpenAI (about 4.25 GW, 20 years), with a cumulative payment cap of $105B—this is a contingent guarantee against a default shortfall, taking effect in installments from 2028, declining over the lease term, terminating once OpenAI obtains a satisfactory credit rating, with OpenAI required to reimburse NVIDIA for amounts actually paid; it is not credit that has already flowed out 【Single source, verified】.
It is not just NVIDIA. Amazon's $50B commitment to OpenAI was fully paid in 2026; for Anthropic, it invested $8B in convertible notes over 2023–2025, and another $10B of preferred stock in Q2 2026, of which $5B came from a financing facility of up to $20B released according to AWS compute delivery milestones 【Single source, verified】—equity flowing down and compute flowing up are written into the same contract. NVIDIA and Microsoft committed in 2025-11 to invest up to $10B and $5B respectively in Anthropic, while Anthropic committed to purchase $30B of Azure compute 【Multi-source】. Broadcom finances Anthropic's deployments through a financing platform with Apollo and Blackstone (first tranche $35B) 【Single source; verified by only one vote】.
Valuations return to upstream profit as book gains. Amazon's Q2 2026 operating income was $27.5B, while its fair-value markup on Anthropic preferred stock was about $50.5B ($62.8B for the first half), about 3 times the AWS segment's $16.6B operating income in the same quarter; this is a pretax, non-operating item, for which about $15.9B of discrete income tax was recorded in the first half 【Single source, verified】. Of NVIDIA's $141.4B of first-half GAAP pretax income, $24.1B was other income, mainly unrealized gains on equity investments, about 17% (about 11% for Q2 alone) 【Single source, verified】. Microsoft recorded a $6.5B pretax net gain on its OpenAI investment in FY26 (mainly a dilution gain from OpenAI's restructuring), plus a $3.2B gain on its Anthropic investment in Q4, the latter kept in non-GAAP 【Single source, verified】. Alphabet's $77.5B of unrealized gains on non-marketable equity in Q2 came from one unnamed private company, widely inferred by the market to be Anthropic, but the name does not appear anywhere in the 10-Q, and the quarter's gains also include SpaceX, so it is not a clean example.
The same revaluation leaves mirror images on both sides: Anthropic's roughly $34B non-cash loss in 2025 came from revaluing convertible instruments, and on Amazon's side it is a gain.
This loop shows three things. First, comparing layers by net income double-counts the model layer's valuation. Second, "model makers are sewing wedding dresses for others" holds for cash but is reversed for risk: upstream is backstopping the model layer's credit, and about a quarter of NVIDIA's business next year, the bulk of cloud providers' backlogs, and the bulk of Amazon's pretax profit this quarter are all tied to the solvency and valuations of a handful of labs (Alphabet also booked a large gain this quarter from one unnamed private company). Third, this loop is coming under pressure: the WSJ reported in late 2026-08 that NVIDIA paused some AI cloud revenue-sharing deals due to internal antitrust concerns and pushback from partners, and NVIDIA responded that the model is "still in place and continues to evolve" 【Unverified; media account】; Jensen Huang said in March that equity investments are probably done, and on the August call the form of support shifted to "selective credit enhancement" for about 2 GW for another frontier lab.
So "which layer the profit lands in" has to be at least partly replaced with another question: who ultimately bears the model layer's credit risk. This site's "Is This AI Capex Boom Another 1999? A Physical for an Analogy" examined this analogy from the angle of financing structure; what this piece adds is the accounting and credit links between layers.
Theory cannot give a point forecast, but the two theories in Section 1 jointly point to two observable variables. Crossing them yields four scenarios. For each scenario, the signals that would show up first are listed.
Variable one: whether frontier capability is "good enough" for customers. Christensen's premise is a performance gap; Thompson's September inference is "good enough." Observable proxies: whether enterprises are willing to pay a premium for the strongest models (per-task cost rising while spending share does not fall), or migrating to cheap "good enough" models (open-source share recovering, mainstream tiers scaling up, application companies switching bases).
Variable two: whether the physical bottlenecks are lifted. When the scarcity of memory, advanced packaging and power access eases. Observable proxies: whether memory makers' margins fall back below NVIDIA's; the direction of H100/B200 contract rental prices; whether PJM capacity prices can clear below the cap; whether turbine slot reservations convert into firm orders.
The reading as of October 2026 sits closer to between scenarios 1 and 2: the physical bottlenecks persist (record memory margins, rebounding rental prices, capacity prices hitting the cap), while evidence of "good enough" has only just appeared (Thompson's single September event, Cursor switching to an open-source base, OpenAI launching a cheaper mainstream tier) and cannot yet be confirmed. The money loop makes this judgment more fragile: if a frontier lab's financing runs into trouble, the first to be hit is the upstream that backstops it.
First, today's profit is upstream, and it gets thicker the closer it gets to physical scarcity. Memory > chips > foundry > cloud segments > neocloud > model layer. This rests on primary SEC data and is the firmest conclusion of this piece. It does not support "the chip layer's share is still expanding": the practitioner-estimated chip-layer share of gross profit fell from about 87% to about 79%, and NVIDIA's gross margin is being squeezed by memory price increases.
Second, scarcity rent is migrating within the chip layer. From GPUs to memory, and it lands on shortage-driven price increases in commodity DRAM/NAND, not on HBM itself. NVIDIA plans to pull its gross margin back with its own price increases, so this migration may be temporary.
Third, the cloud layer's high margins are conditional. R&D not allocated, 5.5–6-year depreciation, AI rental gross margins climbing: each of the three assumptions has counterexamples. Cash flow has already diverged: Amazon and Oracle are negative, and Alphabet has started raising money in the equity market.
Fourth, the model layer has diverged internally, but profit has not yet appeared. The revenue leader changed places within a year, and the only profit is one quarter's adjusted figure. "The model layer as a whole is sewing wedding dresses for others" has to be examined company by company.
Fifth, the "wedding dress" runs half the wrong way. Cash flows from the model layer to upstream, equity and credit flow from upstream to the model layer, and the model layer's valuation then returns to upstream income statements as book gains. The question most worth pressing is no longer which layer has the highest margin, but who is bearing the model layer's credit risk.
Sixth, the long-run landing spot depends on two observable variables: whether the frontier is "good enough," and whether the physical bottlenecks are lifted. Theory cannot give the answer, but it does say what to watch.
Ordered by strength of evidence, strongest first.
What to watch. If any of four things happens, the judgments in this piece should change: first, memory margins falling back below NVIDIA's, indicating that physical bottlenecks are starting to lift; second, a frontier lab reporting GAAP operating profit, or conversely, a down round—the latter would leave a reversing entry in upstream income statements; third, a recovery in enterprise spending share for open-source or mainstream-tier models, which is the direct reading of "good enough"; fourth, NVIDIA's PORTS-Pike guarantee or AI cloud purchase commitments actually being triggered, the first hard signal that the money loop is under stress.
For this issue, each of the 31 sets of load-bearing claims was assigned 3 independent verification agents (two batches of low-risk theoretical quotations and historical figures used a smaller model; the rest used Opus-class), through three lenses—verbatim, counter-evidence, and definitions—with instructions to try hard to refute. Result: none of the 31 sets received three votes of holds-as-written; at the sub-claim level, more than 70 items were ruled dead or barred from use in their original form. For 3 single-source load-bearing pieces of evidence, a counter-evidence search seat and a methodology audit seat were added.
The most important original-draft statements that verification overturned or rewrote:
These corrections share one direction: almost every number that made some layer look "more extreme"—steeper rental price crashes, thinner power rent, higher frontier prices, more concentrated profit—became somewhat milder once checked against the original source.
Theory and history: Christensen, HBR "Breakthrough Ideas for 2004" (https://hbr.org/2004/02/breakthrough-ideas-for-2004), Christensen–Raynor–Verlinden "Skate to Where the Money Will Be" (https://hbr.org/2001/11/skate-to-where-the-money-will-be), Thompson "Aggregation Theory" (https://stratechery.com/2015/aggregation-theory/), "The Benefits of Bubbles" (https://stratechery.com/2025/the-benefits-of-bubbles/), "Agents Over Bubbles" (https://stratechery.com/2026/agents-over-bubbles/), "Anthropic's Safety Superpower" (https://stratechery.com/2026/anthropics-safety-superpower/), "Frontier Overhangs" (https://stratechery.com/2026/frontier-overhangs/), Spolsky "Strategy Letter V" (https://www.joelonsoftware.com/2002/06/12/strategy-letter-v/), Evans "Ways to think about token pricing" (https://www.ben-evans.com/benedictevans/2026/7/9/ways-to-think-about-token-pricing), Nadella interview (https://www.dwarkesh.com/p/satya-nadella-2), Amodei interview (https://cheekypint.substack.com/p/a-cheeky-pint-with-anthropic-ceo), Odlyzko 2003 (https://www-users.cse.umn.edu/~odlyzko/doc/itcom.internet.growth.pdf), Odlyzko on the Railway Mania (https://www-users.cse.umn.edu/~odlyzko/doc/hallucinations.pdf), Intel 2000 10-K (https://www.sec.gov/Archives/edgar/data/50863/000091205701503434/0000912057-01-503434.txt), Compaq 2000 10-K (https://www.sec.gov/Archives/edgar/data/714154/000089056601000112/0000890566-01-000112.txt)
Chips and memory: NVIDIA Q2 FY27 CFO commentary (https://www.sec.gov/Archives/edgar/data/1045810/000104581026000073/q2fy27cfocommentary.htm), NVIDIA 10-Q (https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000075/nvda-20260726.htm), NVIDIA PORTS-Pike 8-K (https://www.sec.gov/Archives/edgar/data/1045810/000104581026000069/nvda-20260817.htm), NVIDIA Hugging Face 8-K (https://www.sec.gov/Archives/edgar/data/1045810/000104581026000078/nvda-20260902.htm), Micron FQ4 FY26 (https://www.sec.gov/Archives/edgar/data/723125/000072312526000018/a2026q4ex991-pressrelease.htm), SK hynix 2Q26 (https://news.skhynix.com/en/q2-2026-business-results/), TSMC 2Q26 (https://www.sec.gov/Archives/edgar/data/1046179/000104617926000451/a2q26e_withguidancexfinal.htm), Broadcom Q3 FY26 (https://www.sec.gov/Archives/edgar/data/1730168/000173016826000076/avgo-08022026x8kxex99.htm), AMD 10-Q (https://www.sec.gov/Archives/edgar/data/2488/000000248826000123/amd-20260627.htm), SemiAnalysis TPUv7 (https://newsletter.semianalysis.com/p/tpuv7-google-takes-a-swing-at-the), Agrawal "The Economics of Generative AI" (https://tailwinds.substack.com/p/the-economics-of-generative-ai-two)
Cloud and neocloud: Amazon Q2 2026 10-Q (https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm), Alphabet Q2 2026 10-Q (https://www.sec.gov/Archives/edgar/data/1652044/000165204426000071/goog-20260630.htm), Microsoft FY26 10-K (https://www.sec.gov/Archives/edgar/data/789019/000119312526323660/msft-20260630.htm), Meta Q2 2026 10-Q (https://www.sec.gov/Archives/edgar/data/1326801/000162828026050705/meta-20260630.htm), Oracle Q1 FY27 10-Q (https://www.sec.gov/Archives/edgar/data/1341439/000119312526389274/orcl-20260831.htm), CoreWeave Q2 2026 (https://www.sec.gov/Archives/edgar/data/1769628/000176962826000362/coreweave2q26earningspress.htm), Oracle GPU gross margin report (https://www.cnbc.com/2025/10/07/oracle-stock-nvidia-chip-margins.html), SemiAnalysis "The Great GPU Shortage" (https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity)
Power: Epoch AI data center cost breakdown (https://epoch.ai/data-insights/ai-datacenter-cost-breakdown), PJM 2028/29 capacity auction (https://www.prnewswire.com/news-releases/pjm-capacity-auction-procures-138-318-mw-of-generation-resources-as-work-continues-to-address-growing-electricity-demand-302825613.html), Vistra Q2 2026 (https://www.sec.gov/Archives/edgar/data/1692819/000169281926000017/vistra-20260630xearningsre.htm), GE Vernova Q2 2026 (https://www.sec.gov/Archives/edgar/data/1996810/000199681026000147/gevpressrelease2q26.htm), Digital Realty Q2 2026 (https://www.sec.gov/Archives/edgar/data/1297996/000110465926086270/dlr-20260723xex99d1.htm)
Model layer and pricing: Anthropic Series H (https://www.anthropic.com/news/series-h), Anthropic–Google–Broadcom (https://www.anthropic.com/news/google-broadcom-partnership-compute), Anthropic draft S-1 report (https://fortune.com/2026/09/29/anthropic-ipo-s-1-prospectus-income-statement/), WSJ revenue report, syndicated (https://finance.yahoo.com/technology/ai/articles/openai-q2-growth-trails-anthropic-102250259.html), OpenAI cash flow projection report (https://finance.yahoo.com/technology/ai/articles/openai-projects-burning-278-billion-230655542.html), Epoch price trends 2025 (https://epoch.ai/data-insights/llm-inference-price-trends), Epoch "The Plunging Price of Thought" (https://epoch.ai/publications/the-plunging-price-of-thought), Gundlach et al. (https://arxiv.org/abs/2511.23455), OpenAI pricing page (https://developers.openai.com/api/docs/pricing), Epoch open-closed gap (https://epoch.ai/data-insights/open-closed-eci-gap), Nagle (https://www.linuxfoundation.org/blog/revealing-the-hidden-economics-of-open-models-in-the-ai-era), Menlo Ventures (https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/)
Applications and money flows: Microsoft–OpenAI 2026-04 amendment (https://blogs.microsoft.com/blog/2026/04/27/the-next-phase-of-the-microsoft-openai-partnership/), Bessemer State of AI 2025 (https://www.bvp.com/atlas/the-state-of-ai-2025), Cursor gross margin report (https://techcrunch.com/2026/04/17/sources-cursor-in-talks-to-raise-2b-at-50b-valuation-as-enterprise-growth-surges/), Windsurf cutoff (https://techcrunch.com/2025/06/03/windsurf-says-anthropic-is-limiting-its-direct-access-to-claude-ai-models), Salesforce Q2 FY27 (https://www.sec.gov/Archives/edgar/data/1108524/000110852426000187/crm-q2fy27xexhibit991.htm), Sequoia "AI's $600B Question" (https://sequoiacap.com/article/ais-600b-question/), Cahn "AI's $1.5T Question" (https://dcahn.substack.com/p/ais-15t-question), Bain 2026 Technology Report (https://www.bain.com/insights/new-innovation-is-required-to-fund-ais-6-trillion-buildout-technology-report-2026/)
Related research on this site: for the financing structure of this buildout compared with 1999, see "Is This AI Capex Boom Another 1999? A Physical for an Analogy"; for the scale and duration of the hardware and power shortages themselves, see "AI's Hardware Shortage and Power Shortage: Real, and For How Long?"; for whether frontier capability is still improving quickly, see "Have Scaling Laws Hit a Wall? An Argument Without a Definition".