Is the AI money real, or going round in a circle?

    The data

    filing SEC filing or company release press named outlet, analyst or company statement; not audited data published third-party dataset calculated our arithmetic on reported figures

    Our reading and calls inference

    This note was written by Claude, which Anthropic makes. Anthropic's figures sit in their own rows, every coverage figure is shown without them, and calls 1, 2, 4 and 5 do not depend on them.

    The call inference

    The loop is real but narrow. The test that matters is whether money from outside the stack grows 50% or more a year until 2030.

    • Circularity is Lucent-sized on paper and smaller in money at risk. Almost all of NVIDIA's 36% is the Ohio guarantee. It is a shortfall guarantee: NVIDIA pays only the gap between a guaranteed value and what a site fetches if OpenAI's affiliate walks away. It starts lease by lease from 2028, ends at 20 years or earlier if OpenAI earns a "satisfactory credit rating", and OpenAI must repay anything NVIDIA pays. Lucent had already lent 6% of a year's revenue to carriers that then failed; NVIDIA has paid out nothing, and most of its other money is equity that never has to be repaid. Its near-term exposure is the $36B of capacity commitments, which shrink only as third parties use the capacity.
    • Anthropic's vendors are paid for their risk, mostly in stock, and have done well so far. Lucent lent at par to start-up carriers and could earn only interest. Anthropic's vendors own equity in a buyer with $16.3B of first-half revenue, positive operating cash flow and $131.7B of cash. The danger is wrong-way risk: in a downturn each vendor loses the customer, the value of its stake and the value of its collateral at the same moment.
    • Broadcom carries the most credit-like exposure. Its rent is partly paid with its own notes, which pay interest in more notes rather than cash for about five years, and its security is its own chips.
    • The GAAP loss is mostly the vendors' gain. Add back the $72.4B rise in the notes and the ~$11.8B charge on Google's and Amazon's spring share agreements, and Anthropic's half-year was about breakeven before tax. Anthropic's other shareholders pay for the vendors' gains, through dilution.
    • About a fifth of the vendors' Anthropic revenue is paid with their own money. Google Cloud and AWS feel it most, since Anthropic is about 22% of each one's backlog. Anthropic's own revenue comes from its customers, not its suppliers; about 48% is sold through the clouds' marketplaces and booked gross, which flatters the headline.
    • Anthropic's part of the loop is better funded than OpenAI's. OpenAI's own plan spends more on compute than it earns through 2030, and the cash it raised runs out by 2028.

    The build already finished is paid for only if Anthropic counts inference

    We count only money from outside the stack, following Craig Moffett's test: "are the revenues coming into the system greater than the costs of building the whole value chain". On our base case it is $93B a year: OpenAI's revenue ($40B), cloud AI bought by customers other than the labs ($30B) and the extra ad revenue AI brings Meta and Google ($23B). Adding Anthropic's $67B gives $160B. The yearly cost of the 2023 to mid-2026 build, as depreciation plus a return on capital, is $135B on the base case.

    Whether the existing build is covered turns on one input, Anthropic's reported revenue. A reader who discounts it, given the conflict above, should read the build as about 70% covered.

    The 2030 bill needs outside money to grow 52–73% a year inference

    J.P. Morgan puts AI capex at $5.5T through 2030. On base-case cost terms that carries a yearly cost of about $950B. Reaching it in the 4¼ years to 2030 takes 72.8% growth a year from $93B, or 52.1% from $160B. The labs grow at those rates today; the ad and cloud rows do not.

    The fund's odds that the 2030 bill is covered on time are kept in Is AI a Bubble?: about 45% counting Anthropic since its S-1, and 20% without. They start from the screen in the data: at this size, four years at 50% a year has happened 0.19% of the time, and only for NVIDIA, the seller in this same boom. The buyer side also looks slow. Solow's remark that computers were everywhere but in the productivity statistics came in 1987, and the surge came about eight years later. If AI's payoff runs on a similar lag, buyers' gains arrive after the bill is due.

    The checkpoint, in Q3 2027 annual figures: outside money of $161B without Anthropic or $243B with it, and $178B for the two labs together, up 52% from $117B now. OpenAI's own target of $70B or more annualised by the end of 2026 is the first test of that path.

    Five calls, each with a date inference

    1. 1. Capex tops operating cash flow at Alphabet or Amazon in Q370% · Oct 28–29
      Wrong if both report Q3 capex at or below Q3 operating cash flow. A yes fires the T1 test, which with T3 is the second of our memo's eight tests, and its bubble call.
    2. 2. None of the big four plans lower capex for 2027 than for 202685% · by Feb 2027
      Wrong if Microsoft, Alphabet, Amazon or Meta guides 2027 capex below 2026 in its Q3 or Q4 report. A yes keeps the suppliers' growth case intact into 2027.
    3. 3. OpenAI and Anthropic together reach $178B annualised by Q3 202765% · Oct 2027
      Wrong if their combined annualised revenue is under $178B by October 2027. Early check: if OpenAI exits 2026 below $60B annualised, this call falls to 45% and call 4 rises to 35%.
    4. 4. OpenAI cuts or pushes back at least $100B of compute because it cannot fund it25% · end-2027
      Wrong if such changes total less than $100B by December 2027. A failure to pay stays at 10% or less. Oracle and CoreWeave are hit hardest.
    5. 5. Credit breaks at Oracle or the neoclouds before any of the big four cuts capex80%
      Wrong if a big-four 2027 capex cut comes first, while Oracle's CDS stays under 300bp and CoreWeave's new debt prices inside SOFR+550.

    All five are judgment with stated reasons, not back-tested rates. They are the fund's calls, and Is AI a Bubble? owns their odds. The evidence for each is in the appendix.

    Who carries the risk inference

    • Oracle, CoreWeave and other neocloudsA big tenant renegotiates while their debt comes due. First tell: Oracle's CDS above 300bp, or CoreWeave's new debt wider than SOFR+550.
    • OpenAI (private)Revenue arrives after the compute bills. First tell: it exits 2026 below $60B annualised, or its next round is priced below $852B.
    • Suppliers: NVIDIA, TSMC, ASML, Broadcom, chip ETFsA capex cut, or losses on NVIDIA's support to customers. First tell: any 2027 capex guide under 10% growth, or NVIDIA's net investment above 50% of its free cash flow.
    • Hyperscalers: Meta, Amazon, Alphabet, MicrosoftReturn on capex falls short, so free cash flow and multiples fall. They earn on compute whichever lab wins. First tell: capex above operating cash flow, or rents turning down.
    • Amazon, Alphabet and Broadcom as Anthropic's vendorsAmazon and Alphabet hold a fifth of their cloud backlog and a large equity mark in one company; Broadcom holds credit exposure secured on its own chips.
    • Software and appsSpend concentrates in a few firms. First tell: Ramp adoption flat for three months.

    What would change our mind inference

    • V1. Amazon or Alphabet books a quarterly loss on its Anthropic stake. First reading Oct 28–29, then each quarter.
    • V2. Anthropic's IPO prices, or trades for 30 days, below $58.90 a share.
    • V3. Broadcom's notes move to the 13% rate, or Amazon's or Google's draws stall on missed milestones.
    • T4. CoreWeave's new debt wider than SOFR+550, or EBITDA after D&A and interest worse than −$0.523B, in its Q3 report.
    • OpenAI below $60B annualised at the end of 2026; H100 one-year rents down 20% over six months; Ramp adoption flat for three months.

    Appendix

    The argument in full

    Anthropic's vendor deals in full

    Anthropic's S-1, filed Oct 26, sets out the deals in full for the first time. The pattern repeats across them. A supplier promises money, in stock purchases or convertible notes. The money is released as Anthropic takes delivery of that supplier's compute. Anthropic then spends it, and more, on that supplier's chips or cloud. 66 filing

    Per-share figures are after the 10-for-1 split of Sept 17, 2026. Each preferred share converts into 10 common shares, so the Series G-2 price of $259.14 is $25.91 a common share, and Series H-1's $589.01 is $58.90. 68

    VendorMoney inInstrument and termsCompute it supportsWhat it gets back
    Amazon$6.2B of 2023–24 notes outstanding; $5.0B of Series G-2 stock (Apr 2026); $15B of 2026 notes undrawn
    • The old notes pay no interest and convert at $2.09 and $5.05 a share, about 1.84B shares in all.
    • The new notes are drawn in $5B steps as compute milestones are met. They settle in stock at the 30-day average price.
    • Unconverted amounts roll into 10-year notes that Amazon can set off against what it owes Anthropic.
    • Its stake is capped at 12%.
    		At least $110.0B, May 2026 to Apr 2036; about 22% of AWS's backlog
    • The old notes are marked at $108.6B, against $6.2B lent.
    • The G-2 shares are worth about $11.4B at the H-1 price. arithmetic
    • It earns its cloud margin on the compute.
    Google$10.0B of Series G-2 (Mar 2026); $20B of notes undrawn; earlier notes already converted
    • The $20B is drawn through 2030 as targets are met. Google takes either a convertible note on terms agreed at the time or stock at the 30-day average.
    • It can move up to $10B of that $20B into stock at the IPO price.
    • Its stake is capped at 14.99%.
    		At least $114.5B, Apr 2026 to Jul 2033, plus TPU capacity for 2028–31 estimated at $125–150B; about 22% of Google Cloud's backlog
    • The G-2 shares are worth about $22.7B at the H-1 price, against $10B paid. arithmetic
    • It also earns from TPU and cloud sales.
    Broadcom$42B of notes undrawn
    • The notes exist "solely to finance quarterly rent obligations" on the TPU lease.
    • They pay 9% in kind, rising to 13% and converting at 85% if revenue-growth tests are missed.
    • The holder can convert at 90% of the 30-day average price.
    • They are junior secured on the leased equipment, pay 150% on a change of control, and can be passed to a financial partner.
    • They mature in about five years.
    		About $161.2B of lease rent over five yearsA coupon that compounds, a 10% conversion discount and the rent. This is the most loan-like deal on the list.
    AMDUp to $5.0BIt buys Class N stock in steps as GPUs are delivered, from 2027, at the 30-day average price. There is no discount and no coupon.Over $20B of Helios MI450 systemsStock bought at the market price. The only upside is the stock itself.
    MicrosoftUp to $5B pressAnnounced Nov 2025. The terms were not found in the S-1 sections we read.At least $44.9B of Azure through May 2033; about 7% of Microsoft's backlogEquity, and Claude on Foundry and Microsoft 365
    NVIDIAUp to $10B pressAnnounced Nov 2025. The terms were not found in the S-1 sections we read.About $206B of NVIDIA-GPU capacity rented through SpaceXAI ($120.5B), Nscale ($44.6B) and Lambda ($41.3B)Equity, plus chip sales to the neoclouds that serve Anthropic
    Meta's raise (Oct 26), for comparisonNVIDIA, AMD and Broadcom showed non-binding interest in up to $5BPart of a $22.5B forward sale and $12.5B mandatory convertible, on the same terms as other buyersMeta's Q3 capex was $34.9B against $32.8B of operating cash flowMeta stock at the market price, and a customer that keeps buying

    Sources: S-1 note-purchase and supply sections, Level 3 fair-value table and stock tables 66,67,68; our S-1 and Meta reports 70,71; Microsoft and NVIDIA from press 72. Backlog shares are our division of the S-1 minimums by each cloud's Q2 backlog. arithmetic

    Do the vendors benefit if it pays off?

    Yes, and they already have. This is the main way the deals differ from old-style vendor financing. Lucent lent at par and could earn only interest. Anthropic's suppliers hold equity, or notes that turn into equity, so they share in the upside. inference

    • Amazon is the clearest case. Its 2023–24 notes convert at prices set when Anthropic was a fraction of its size. At Jun 30 Anthropic carried them at $108,566M on $6,203M of principal. That is about $59 a share, the same as the H-1 price. 67 filing Amazon's own Q2 report booked a $53.4B gain on its Anthropic holdings. 70 internal notes
    • Google and Amazon got a discount this spring. Their March and April agreements locked in Series G-2 at $25.91 a share. Series H-1 sold at $58.90 in May and June, raising $45.7B. Anthropic booked about $11.8B of fair-value charges on those two agreements before they settled. 67,68 filing, with the charge's direction our reading inference
    • The new money is priced at the market. Amazon's $15B, AMD's $5B and the stock route for Google's $20B all settle at the 30-day average price after the IPO. Those deals carry upside only if the stock rises from there. Broadcom's notes are the exception: they carry a coupon and a 10% discount. 66 filing

    Who pays for the vendors' gains. Anthropic's other shareholders do, through dilution. That is the "loss" in the income statement. In the first half, the notes' $72.4B rise and the ~$11.8B charge on the spring agreements together come to more than the whole $80.7B net loss. Before those two marks, Anthropic's half-year was about breakeven before tax. 67,69 inference

    How much of it is circular

    The "all six" total leaves out Amazon's $6.2B of old notes, which bring vendor money to about $118B. Google's IPO option is not added, because it comes out of the same $20B. arithmetic

    • What is circular. Part of what the vendors will book as Anthropic revenue is paid with money they put in. That is about a fifth, on these figures. It matters most for Google Cloud and AWS, where Anthropic is about 22% of each backlog. inference
    • What is not. Anthropic's own revenue comes from its customers, not its suppliers. About 48% of it is sold through the clouds' marketplaces, and Anthropic books that gross, which flatters the headline. 70 No customer was more than 10% of revenue in the first nine months. filing
    • Sized against the whole stack. At Jun 30 Anthropic had $265.6B of fixed commitments. It has signed about $417B since, which may overlap, and estimates $125–150B more for 2028–31. That is roughly $810B. Vendor money of $103–118B is 12–15% of it. The rest must come from customers, cash on hand ($131.7B at Jun 30) or the IPO. 70 arithmetic

    Lucent then, Anthropic's vendors now

    Lucent, Sept 2000Anthropic's vendors, 2026
    What the vendor gave$6.7B of credit lines and $1.4B of guarantees, 24% of a year's revenue; $2.07B drawn 4$15B of stock bought, $6.2B of notes outstanding, $82B more promised
    Who received itStart-up carriers with little revenueA lab with $16.3B of first-half revenue, $3.9B of operating cash flow and $131.7B of cash 70
    InstrumentLoans, repaid at parEquity at or below market; Broadcom's convertible PIK notes
    Vendor's upsideInterestEquity gains: Amazon's notes are worth about 17 times principal
    Vendor's downsideLoan write-offs when carriers failedStake marks fall, the customer shrinks and the collateral loses value, all at once

    Nortel ran the same playbook as Lucent in 2000. We did not re-check its figures for this page.

    Steelman, counter-case and net view

    • Steelman: this is strategic investing, not propping up a weak buyer. Anthropic has $131.7B of cash and could pay its near-term bills without the vendor money. The vendors are buying a share of the customer's upside. Releasing the money only when compute is delivered keeps both sides honest. The vendors have done well so far, which is why they keep signing. inference
    • Counter-case: the payoffs all point the same way. inference
      • Each vendor's equity, notes, rent and backlog depend on Anthropic growing. A slowdown hits all of them at once. That is wrong-way risk.
      • Broadcom is most exposed. Its rent is partly paid with its own notes, which earn interest as more notes rather than cash for about five years, and its security is its own chips.
      • Amazon's and Alphabet's profits now carry large marks on private stakes. Those reverse in a bust. Amazon's $53.4B Q2 gain shows the scale.
      • Stock released at delivery milestones means a vendor ends up with more of Anthropic exactly as Anthropic uses more compute. The vendors are adding to the same bet.
    • Net view. The vendor financing is real, large and rewarding for the vendors so far. It is not the Lucent failure mode, because the buyer has revenue, cash and positive operating cash flow. It does concentrate the AI bet. AMZN and GOOGL hold a fifth of their cloud backlog and a large equity mark in one company. AVGO holds credit exposure secured on its own chips. This page's earlier calls do not change. Anthropic's part of the loop is better funded than OpenAI's. inference
    Vendor tripwireNowScoredIf it trips
    V1: Amazon or Alphabet books a quarterly loss on its Anthropic stakeAmazon booked a gain in Q2Oct 28–29, then each quarterThe marks start working against reported profit
    V2: Anthropic's IPO prices, or trades for 30 days, below $58.90 a share (the H-1 price)No price range yetAt pricing (S-1/A)Vendor marks and the stock-settled draws turn against the vendors
    V3: Broadcom's notes move to the 13% rate, or Amazon's or Google's draws stall because milestones are missedNo notes drawn yetAnthropic's and Broadcom's filingsGrowth is falling behind the compute plan; wrong-way risk is live
    V4: Vendor money rises above 25% of new compute signedAbout 17% of named vendor dealsWith each new dealOutside capital is thinning, and the loop is tightening

    How the money goes round

    Suppliers put money into the labs. The labs sign contracts with the clouds. The clouds buy chips from the suppliers. Money enters from outside in only one place: people, companies and advertisers paying for AI. Only the green arrow can pay for the build in the end.

    OpenAI's announced cloud deals:

    • Microsoft: an incremental $250B of Azure. 24
    • Oracle: about $300B, as reported by the WSJ; Oracle has not named OpenAI. 25
    • AWS: $38B, plus $100B over 8 years. 26
    • CoreWeave: about $22.4B. 27

    These are headline values signed at different dates, not what is still owed. The clouds' combined backlog is about $2.35T: Microsoft $678B, Oracle $664B, Google $514B and AWS $496B. 6,18,10,13 No cloud reports how much of it the labs owe. Press estimates put it at about half for AWS and about half of Oracle's from OpenAI alone. 58 press estimate

    How much NVIDIA has put back into its customers

    Revenue is from NVIDIA's last four 10-Q and 10-K income statements: $57.0B, $68.1B, $81.6B and $96.2B. 5 Lucent's $8.1B and $2.07B are our sums of the two lines its 10-K reports. filing

    Verdict: Lucent-sized on paper, smaller in money at risk. inference

    The Ohio guarantee is a shortfall guarantee. NVIDIA pays only the gap between a guaranteed value and what the site fetches if OpenAI's affiliate walks away. 3 filing

    • Each lease's guarantee starts when that lease starts, which is from 2028.
    • It ends at 20 years, or earlier if OpenAI earns a "satisfactory credit rating".
    • OpenAI must repay NVIDIA for anything NVIDIA pays.

    Lucent had already lent 6% of a year's revenue to start-up carriers that then failed. NVIDIA has paid out nothing on its guarantees. Most of its other money is equity, which never has to be repaid. Its near-term exposure is the $36B of capacity commitments: they shrink only as third parties use the capacity.

    An earlier version said 48% "against Lucent's 24%". That compared NVIDIA's guarantees plus capacity commitments with Lucent's credit plus guarantees. On guarantees alone the like-for-like figure is 36%.

    The five calls

    1. Capex tops operating cash flow at Alphabet or Amazon in Q370% · Oct 28–29
    Evidence

    Capex as a share of operating cash flow, latest quarter:

    • Alphabet: 115% (Q2).
    • Amazon: 119% (Q2).
    • Meta: 103% (Q3: capex $33.9B, cash flow $32.8B), or 106% counting finance-lease principal.
    • Microsoft: 65% (Q2).

    Meta's Q3 already counts for the test. For Alphabet or Amazon to fall back below 100%, capex would have to drop, or cash flow rise, by 13–16% in a single quarter. Meta guided full-year 2026 capex to $135–145B. 5,14 filing

    Wrong if
    Both Alphabet and Amazon report Q3 capex at or below Q3 operating cash flow.
    Then
    T1 of our Oct 11 bubble memo reads bear. Its rule is 2 of 4 above 100% on Q3 data. With T3 already bear (CCC spreads above 13%), that makes the second of the memo's eight tests to fire, which is the memo's bubble call. 56
    What it means
    This is a financing signal, not a demand signal. On its own it says little about the hyperscalers' earnings. If capex guidance also softens, the chip suppliers are next in line.
    2. None of the big four plans lower capex for 2027 than for 202685% · by Feb 2027
    Evidence
    • Backlogs: Microsoft's commercial RPO is $678B, up 84%; Google Cloud's is $514B; AWS has about $496B. 6,10,13 filing
    • Over 60% of the data-centre capacity due in 2027 was not under construction in June (J.P. Morgan, via the WSJ), which pushes spending into 2027. 48 press
    • In telecom, capex kept rising through 2000 after high-yield telecom spreads reached 1,200bp in November. It fell only in 2001. 51,52 Fed papers
    Wrong if
    Any of Microsoft, Alphabet, Amazon or Meta guides 2027 capex below 2026, in its Q3 or Q4 report.
    Then
    Suppliers lose their growth first.
    What it means
    A yes keeps the suppliers' growth case intact into 2027.
    3. The two labs reach $178B annualised by Q3 202765% · Oct 2027
    Evidence
    • OpenAI was at about $12B annualised in July 2025, "$20B+" for 2025, and close to $50B in September 2026. That is 2.5× in nine months. 31,23,29
    • Anthropic's quarterly revenue went from $11.6B in Q2 to $16.8B in Q3, up 45%. 33 press
    • If OpenAI ends 2026 at $70B and Anthropic holds at $67B, the two are at $137B. From there they need only 30% more in nine months.
    • Against that, the screen above shows how rare 50%-a-year growth is at this size. And two measures have stalled: big-firm spend per employee and token spend, both flat in August–September on Ramp's card data. 41 primary data
    Wrong if
    The labs' combined annualised revenue is under $178B by October 2027. The two report on different bases: OpenAI net of partner sales, Anthropic gross. 29
    Early check
    If OpenAI exits 2026 below $60B annualised, cut this call to 45% and raise call 4 to 35%.
    4. OpenAI cuts or pushes back at least $100B of compute commitments by the end of 2027 because it cannot fund them25%
    Evidence
    • OpenAI's plan, as the FT reported it in September: revenue rises from $36B in 2026 to $350B in 2030. Compute spending over the period is $856B, against $840B of revenue, with −$278B of free cash flow. 28
    • The $122B it raised in March, at $852B, is "on track to exhaust" by 2028. 22,28
    • Funding is still open: Bloomberg reports talks to raise $30B or more at $1.4T before new money. 30 press
    • The compute plan has already moved once, from about $600B in February to $856B. A restructure is the normal way such plans get resized, not a default.
    Wrong if
    Changes forced by OpenAI's finances total less than $100B by December 2027. Moving work to its own chip or between clouds does not count. The odds and their reasoning are in Is AI a Bubble?; this page earlier put 60% on a wider event that counted any restructuring.
    What it means
    Oracle and CoreWeave are hit hardest. NVIDIA's Ohio guarantee starts only as leases begin, so a delay before 2028 costs NVIDIA growth, not guarantee payouts.
    5. Credit breaks at Oracle or the neoclouds before any hyperscaler cuts capex80%
    Evidence
    • Oracle's five-year CDS was 295bp on Oct 19 after briefly touching 300bp (CNBC). 55
    • Oracle sold $20B of stock through an at-the-market program last quarter. Its net debt is 3.2× EBITDA and its free cash flow is negative. 19,5 filing
    • CoreWeave's 2031 and 2032 notes yielded about 15% on Oct 14. Its $35.1B of debt is 34% of its $104B backlog. 54,21
    • The hyperscalers carry little net debt, but they now need outside capital too. Meta's filing describes a need to "raise hundreds of billions of dollars" over several years, in "an increasingly challenging environment for raising capital". 16 filing
    Wrong if
    A big-four 2027 capex cut comes first, while Oracle's CDS stays under 300bp and CoreWeave's new debt prices inside SOFR+550.
    What it means
    Lenders to the debt-funded builders take the first loss. The second-order hit is suppliers' order books.

    Moffett's four claims, tested

    Craig Moffett made the sharpest version of the bear case on Stratechery on Oct 22. The interview is paywalled. We worked from quotes in a Stansberry write-up and could not read the transcript itself. 44 secondary

    ClaimStrongest formWhat the data sayVerdict
    Much of the revenue just moves between layersNVIDIA funds customers who buy NVIDIA; labs pay clouds with suppliers' moneyGuarantees are 36% of NVIDIA's revenue, 1% without Ohio. Amazon put up $50B of OpenAI's $122B round, $35B of it contingent on an IPO or milestone 36true, concentrated
    Outside money must cover the whole stackCount only consumers, companies, governments and advertisers69% covered without Anthropic, 118% with; the 2030 bill needs 52–73% growth a yearnot yet, open
    Deflation shrinks the revenue"I think there's going to be a lot of price deflation": most work does not need the best modelUnit prices are falling fast: Epoch puts it at 13× a year since 2023 for the same capability, and token prices fell 43% in ten weeks this summer 40,39. Volume is growing faster: Google's tokens rose 6.7× from I/O 2025 to I/O 2026 11. Rents on old chips are rising: H100 one-year contracts rose 38% from Oct 2025 to Mar 2026, and the Silicon Data H100 index rose from $2.00 to $2.88 an hour from December to October 38,37not yet
    The 1999 fibre got used, and the builders still went bustTechnology real, capex unpaidThe same split shows today. The hyperscalers have cash and the borrowers do not: Oracle at 295bp, CoreWeave at about 15%the real risk

    Telecom comparison: internet transit prices fell 16% in 2000, 40% in 2001 and 50% in 2002, and averaged about 61% a year over 1998–2010 (DrPeering, an informal series). 53 Today's rising rents for old chips are the clearest difference from 2000. One caveat: rising rents could also reflect short supply of memory and power rather than demand alone. inference

    The bull case at full strength

    • Demand outside the labs is broad and on the record. filing
      • Microsoft: 100,000 Foundry customers and over 30M paid Copilot seats. 7,6
      • Google: "nearly 500" cloud customers each processed over a trillion tokens last year, and Cloud revenue grew 82%. 10,9
      • AWS: growing 37%, with its AI business above a $25B run-rate and growing at triple digits. 12
      • Palantir: US commercial revenue grew 149% in Q2. 42
      • Cursor: passed $4B annualised in June. 43 press
    • The best lab is turning sales into cash. press, unaudited
      • Anthropic told investors it had positive free cash flow in Q3 and an adjusted operating margin above 11%. Bloomberg reports an 85% gross margin. 33
      • Over 1,000 customers each pay it more than $1M a year (Anthropic, April). 35
      • The S-1 now shows first-half operating cash flow of +$3.9B and an adjusted operating margin of +10%. 70 filing
    • Ads pay without any lab. Meta's ad revenue grew 26% in Q3. On its Q2 call, Meta credited its ranking models with 15.7% more conversions on Facebook. 14,15 company
    • Chips last longer than the bear case assumes. Oracle reports 97.9% GPU utilisation, with older GPUs renewed or resold at a 20% premium. Most of those GPUs are four or more years old (call, via DCD). 20 secondary With seven-year chips and 6% capital, coverage is 82% without Anthropic and 141% with it.

    What the bull case concedes. filing / primary data

    • Microsoft's backlog grew 25% excluding OpenAI, against 84% in total. 7
    • On Ramp's card data, 80% of business spend on the two labs comes from 1% of firms. Adoption is still rising, but more slowly. 41
    • Microsoft has not updated its $37B AI run-rate since April. 8

    Ed Zitron's case, checked against his sources

    • About $110B of AI compute demand, 90% or more from two labs. 46 This is his estimate and he names no source. On $110B, existing-build coverage would be 81%. Agrees with our base case.
    • The clouds' AI growth is the labs. He cites three brokers: 45
      • Barclays: the labs are 73% of Amazon's AI revenue.
      • UBS: 59% for Amazon.
      • Wells Fargo: 70% or more of Microsoft's.
      This is why our cloud row counts only the non-lab part. Agrees.
    • Oracle's backlog is long-dated. He gives 12% due within a year, from a BofA note. Oracle's 10-Q says about 13% of $664B, roughly $86B. 18 Close, and confirmed by the filing.
    • Hidden liabilities. He gives over $1.35T of off-balance-sheet debt for four hyperscalers, citing Nikkei; the later Nikkei figure, which adds Oracle, is $1.65T. 47 These are mostly leases not yet started and purchase commitments. Our cost figure leaves them out. Makes the gap worse.
    • Where the data push back. An overbuilt market should see rents fall, and H100 rents are rising. And his ledger leaves out the extra ad revenue AI brings Meta and Google.

    The earlier version cited Fred Duan on 2027 depreciation and Gavin Baker on Anthropic's cash flow. We could not find an original for either, so both are dropped; the cash-flow claim rests on the FT, Bloomberg and Axios reports. Brad Gerstner's line on All-In is that AI revenue must go "from $200 to $450 to $800 or $1 trillion just to keep up". He gives no year. 50

    Chamath's AI ROI case, checked against primary data

    "Deep Dive: Why AI is booming, but productivity isn't" by Chamath Palihapitiya, Social Capital, Sept 25, 2026. 59 Only the public part was read; its practical guide is paywalled.

    His stance. He is bullish on the value but says it is not in the productivity numbers yet. Firms bolt AI onto old processes instead of first deleting the steps that should not exist. In his words: "AI lowers the cost of doing things, but does nothing about whether they were worth doing in the first place." claim

    Steelman. AI costs buyers almost nothing per seat, so price is not what holds adoption back. The constraint is managers redesigning work, which moves at the speed of large companies, not of the models. When firms do redesign, the gains are large but go to a few winners. If he is right, demand is real but slow, and very lopsided.

    His claimWhat the original showsVerdict
    The median firm spends $12.50 per employee a month on AI, so it breaks even if staff are 0.15% more productiveRamp's own June post gives a median of $46 per employee a month in April (middle half $3–352). We could not find $12.50 in Ramp's published data. On $46, break-even is 0.54%, about 13 minutes of a 40-hour week. 62 arithmeticFigure not found; the point still holds
    Ramp tracks more than 70,000 businessesThe Ramp AI Index on FRED: 55.9% of businesses paid for AI in September. That is up 0.6 points on August, against 2.4 points from February to March. 63 dataConfirmed; adoption is slowing
    515 startups got the same tools; the half shown how others reorganised made 1.9× the revenue, almost all from the top 10%Kim, Kim and Koning, SSRN, Mar 30, 2026. Treated firms found 44% more AI uses, had 1.9× the revenue and needed 39.5% less outside capital, with no change in hiring. The gains sat at the 90th percentile and above. 60 paperConfirmed
    At P&G, one person with AI matched a two-person team without itDell'Acqua and others, NBER w33641, Apr 2025. In a field trial of 776 P&G professionals, "individuals with AI matched the performance of teams without AI". 61 paperConfirmed
    US productivity grew about 1.5% a year in 1973–95 and 3.3% in 1995–2003BLS nonfarm output per hour, on FRED: 1.50% a year from Q4 1973 to Q4 1995 and 3.30% to Q4 2003. Since Q4 2019 it has grown 2.06% a year, to Q2 2026. 64 arithmeticConfirmed
    MIT economists find automation pays when a person checks once, at the endNo paper is named, and we did not find one.Unsourced; not used
    Solow, 1987, and Gates, 1995, quotesWell-known quotes, not re-checked here.Not re-checked
    • Where it agrees with our calls. The gains are lopsided, and that matches Ramp's finding that 1% of firms make 80% of lab spend. 41 Outside money therefore rests on a few heavy buyers. That is why call 3 is 65% and not higher. inference
    • Where it sharpens the timing risk. Solow's remark came in 1987, and the productivity surge started about eight years later. The $950B-a-year bill is due by 2030. If AI's payoff runs on a similar lag, buyers' gains arrive after the build has to be paid for. That is the main reason we put the chance of on-time coverage at about 45% with Anthropic since its S-1 (40% before) and 20% without. inference
    • Where it conflicts. If the limit is how firms organise work rather than price, cheaper tokens bring more use. That softens Moffett's deflation worry for buyers. It does not soften it for sellers, and our ledger measures sellers' revenue. inference
    • Tripwire it adds. The Ramp AI Index flat for three months running, already in the risk table. September is 55.9%.

    Tripwires, with today's reading

    LineNowScoredIf it trips
    T1: capex above operating cash flow at 2 of the big four on Q3 data, or 2027 capex growth guided under 10%1 of 4 (Meta 103%)Oct 28–29Call 1 resolves; with T3, the memo's bubble call
    T2: NVIDIA's net equity investment above 50% of free cash flowH1: 50.3% ($35.2B / $69.9B); Q2 alone 40.3% 56NVIDIA Q3, about Nov 18 (date not yet confirmed)Suppliers' growth at risk
    T3: Oracle CDS above 300bp, or CoreWeave debt wider than SOFR+600; or IG above 110bp, or CCC above 13%CCC 13.40% (Oct 22): tripped. Oracle 295bp. IG 0.89%weekly to Dec 31Already bear; call 5 moving
    T4: CoreWeave new debt wider than SOFR+550, or EBITDA after D&A and interest worse than −$0.523B2031–32 notes at about 15%CoreWeave Q3, about Nov 9–16 (date not yet confirmed)Neocloud stress confirmed
    OpenAI annualised revenue at year endAbout $50B (Sept)Jan 2027Below $60B: call 3 to 45%, call 4 to 35%
    H100 one-year rents fall 20% over six monthsIndex $2.88/hr, risingmonthlyDeflation case gains; suppliers exposed

    T1–T4 are the lines in our Oct 11 AI capital cycle memo (RPT-101). Its bubble call needs 2 of its 8 tests to read bear on Q3 or later data by Mar 31, 2027. 56 Earnings dates other than Microsoft's are from the earnings calendar, not the companies. 57

    Method: the coverage ledger and the 2030 bill

    Does outside money cover the cost? The ledger

    We count only money from outside the stack, following the test Craig Moffett set out on Stratechery: "are the revenues coming into the system greater than the costs of building the whole value chain". 44 The figures below are annual rates for Q3 2026. Each input sits in its own row, so a reader can swap in their own.

    Outside money ($B a year)LowBaseHighBasis
    OpenAI274050
    • Low: Q2 recognised revenue of $6.7B, × 4 (WSJ). 32
    • Base: more than $40B annualised in August (Bloomberg).
    • High: close to $50B annualised in September, on OpenAI's own net basis (FT). 29
    press
    Cloud AI from customers other than the labs203040
    • AWS: about $8.5B of non-lab AI revenue in 2026 (Barclays).
    • Microsoft: about 30% of its $37B AI run-rate, since Wells Fargo puts the labs' share at 70% or more. 8,45
    • Google: $0–20B, which it does not disclose.
    broker estimates via Zitron
    Extra ad revenue that AI brings Meta and Google023460%, 25% or 50% of their combined revenue growth, about $93B a year:
    • Meta: Q3 grew $13.4B on the year, × 4.
    • Alphabet: about 40% of its Q2 growth, × 4.
    Neither company reports how much of that growth AI caused. This is the widest input. 14,9 assumption
    Total without Anthropic4793136
    Anthropic676767Q3 revenue of $16.8B × 4. Anthropic gave the figure to investors and the FT, Bloomberg and Axios reported it; it is unaudited. It is on a gross basis that includes sales through cloud partners, so part of it may double-count the cloud row. 33 press
    Total with Anthropic114160203
    Yearly cost of the build in serviceBullBaseBear
    Cash capex 2023 to Q2 2026: Microsoft, Alphabet, Amazon, Meta and Oracle 5$1,134.9B
    Share of that capex spent on AI70%70%100%
    Chips' share of the spend, and how long they last (buildings last 20 years)60%, 7 yrs60%, 6 yrs60%, 5 yrs
    Cost of capital6%8%9%
    Yearly cost: depreciation plus return, as a level annuity$113B$135B$225B

    Each bar uses the base outside-money total. Across all three outside-money cases the range is 21–120% without Anthropic and 51–180% with it. The cost leaves out leases, power, staff and the labs' own running costs, so coverage is overstated. Meta alone has $356B of signed leases that have not yet started. 17 arithmetic

    Verdict: the existing build pays only if Anthropic counts. Whether the build is covered today turns on one input: Anthropic's reported revenue. A reader who discounts it, as the conflict note below says to, should read the build as about 70% covered. inference

    The 2030 bill and the growth it needs

    J.P. Morgan Global Research puts AI capex at $5.5T through 2030. 49 analyst On the base-case cost terms, that spend carries a yearly cost of about $938B, which we round to $950B. arithmetic

    Closing the gap in the 4¼ years to 2030 takes the following annual growth:

    • 72.8% a year from $93B, without Anthropic.
    • 52.1% a year from $160B, with Anthropic.

    What the table means. The bill is covered on time under these conditions: inference

    • With Anthropic: outside money doubles next year, then keeps 70% of each year's growth rate the year after.
    • Without Anthropic: it grows about 150% next year with 70% kept, or 200% with half kept.

    The labs grow at those rates today. The ad and cloud rows do not.

    The fund's odds of the bill being covered on time, kept in Is AI a Bubble?, are about 45% with Anthropic (40% before its S-1) and 20% without. inference The reason is the one fact below, a screen of US-listed companies taken from the earlier version of this page. We did not re-run it.

    The sample is 2,632 company-quarters from 161 companies, 2004–2026. It counts only companies still listed, so it is a frequency, not a probability for OpenAI. The one company of this size that grew this fast was the seller in this same boom.

    The checkpoint. Annual figures for Q3 2027 that would put outside money on the required path:

    • $161B without Anthropic.
    • $243B with Anthropic.
    • $178B for the two labs together, growing at the 52% the whole ledger needs, from $117B now. This is call 3.

    OpenAI's own year-end target is $70B or more annualised. 30 It is the first test of that path. press

    Data tables for every chart
    Conflict, limits and changes

    Conflict note, limits, and what changed

    • Conflict. This page was written by Claude, which is made by Anthropic. Anthropic's figures are shown in their own row. Each coverage figure is given without them, and calls 1, 2, 4 and 5 do not depend on them.
    • Limits.
      • The probabilities are judgment with stated reasons, not back-tested rates.
      • OpenAI's figures and Anthropic's Q3 figure are press reports of company statements, not audited. 34 Anthropic's figures to Jun 30 now come from its public S-1. Its first-half numbers are unaudited, and its 2023–25 numbers are audited. 65
      • The S-1 is long, and we read the sections on notes, supply, fair value and stock. We did not find Microsoft's or NVIDIA's investment terms, so those come from press.
      • The FT, WSJ, Bloomberg, Stratechery and broker notes were paywalled. Where that is so, the source list names the readable copy we used.
    • What changed on Oct 26.
      • Added a section on Anthropic's vendors, built from the public S-1. It covers what each vendor puts in, the terms, the compute behind it and what the vendor gets back. It also covers the self-funded share, how vendor money compares with the ~$810B stack, a Lucent comparison, a steelman, a counter-case, a net view and tripwires V1–V4.
      • Checked two claims against the filing. "Vendor money is about $110B of ~$800B" holds: the filing gives $103B, and press reports bring it to $118B. "Most of the GAAP loss is a fair-value gain for note holders" also holds: the notes account for $72.4B of the $80.7B loss, and the outstanding notes are all Amazon's.
      • The five calls and the coverage figures above are unchanged.
    • What changed on Oct 25.
      • The probability panel from two AI models is removed. In its place are five calls with dates and the facts that would prove them wrong.
      • Capex is restated to $1,134.9B.
      • The cost model is shown with every input.
      • NVIDIA's comparison with Lucent is now like-for-like: 36%, not 48%.
      • OpenAI's contracts are restated to about $710B. The $281B and $319B were not announced deal sizes.
      • OpenAI's growth now uses dated points only.
      • Oracle's net debt to EBITDA is 3.2×, and 13% of its backlog is due within a year.
      • Meta's 15.7% is dated to Q2, on Facebook.
      • Epoch's figure is 13× a year.
      • Unsourced investor quotes are dropped.
      • Earnings dates are flagged as unconfirmed.
      • Added Chamath's AI ROI piece, checked against Ramp, FRED and the two papers it draws on.
    • What changed on Oct 26, later.
      • Rebuilt chart-first. Reported facts now come first as 19 charts drawn from one data file; our reading, odds and calls sit in a separate section; the full argument, method, tables and sources moved to this appendix.
      • The two summaries of Oct 25 and Oct 26 are merged into one set of findings and one call.
      • No fact, figure or source was removed. The calls, odds and coverage figures are unchanged.
    Sources

    Sources

    Every link below was opened during the Oct 25 check unless it is marked paywalled. A paywalled link is the original, and the readable copy we used follows it.

    1. NVIDIA 10-Q, Q2 FY27, commitments and contingencies note ($36B AI-cloud capacity, $20B leases to reassign, $25B equity commitments, $279B supply). sec.gov R17
    2. NVIDIA 10-Q, Q2 FY27 (guarantees $3,529M; equity securities $51.2B non-marketable, $42.8B marketable). sec.gov
    3. NVIDIA 8-K, Aug 17, 2026 (residual value guaranties up to $105B, PORTS campus, Pike County, Ohio). sec.gov
    4. Lucent Technologies 10-K405, FY ended Sep 30, 2000 (credit commitments, guarantees, revenue). sec.gov
    5. Financial statements via a market-data vendor, compiled from filings: capex, operating cash flow, revenue, Oracle net debt to EBITDA, and NVIDIA's quarterly revenue. Filings behind them: MSFT 10-K FY26, GOOGL 10-Q Q2, AMZN 10-Q Q2, ORCL 10-Q FQ1 FY27.
    6. Microsoft FY26 Q4 release (commercial RPO $678B, +84%; over 30M paid Copilot seats). sec.gov
    7. Microsoft FY26 Q4 earnings call ("RPO increased 25% when excluding OpenAI"; 100,000 Foundry customers). microsoft.com
    8. Microsoft FY26 Q3 release and call (AI run-rate over $37B, +123%). sec.gov, call
    9. Alphabet Q2 2026 release (Google Cloud $24.8B, +82%; revenue $119.8B). sec.gov
    10. Sundar Pichai, Q2 2026 remarks (Cloud backlog $514B; nearly 500 customers over 1T tokens). blog.google
    11. Google I/O 2026 keynote (about 480T to over 3.2 quadrillion tokens a month). blog.google
    12. Amazon Q2 2026 release (AWS +36.7%; AI run-rate over $25B). sec.gov
    13. Amazon 10-Q, Q2 2026 (performance obligations about $496B; OpenAI and Anthropic expansions). sec.gov
    14. Meta Q3 2026 release (revenue $64.6B; ads +26%; 2026 capex $135–145B). sec.gov
    15. Meta Q2 2026 call transcript (15.7% conversion uplift on Facebook). investor.atmeta.com PDF
    16. Meta PRE 14C, Oct 21, 2026 ("raise hundreds of billions of dollars"). sec.gov
    17. Meta 10-Q, Q3 2026, commitments note (leases not yet commenced $355.7B). sec.gov
    18. Oracle 10-Q, FQ1 FY27 (RPO $664B; about 13% due within 12 months). sec.gov
    19. Oracle FQ1 FY27 release ($20B of stock sold through an ATM). sec.gov
    20. DatacenterDynamics, Sep 11, 2026, on Oracle's call (97.9% utilisation; 20% premium). Secondary; Oracle's own transcript could not be opened. datacenterdynamics.com
    21. CoreWeave Q2 2026 release (revenue $2.58B; backlog about $104B; debt lines summing to $35.1B). sec.gov
    22. OpenAI, "Accelerating the next phase of AI", Mar 31, 2026 ($122B at $852B post-money). openai.com
    23. OpenAI CFO, "A business that scales with the value of intelligence", Jan 18, 2026 ("$20B+" ARR in 2025). openai.com
    24. OpenAI, "The next chapter of the Microsoft–OpenAI partnership" (incremental $250B Azure). openai.com
    25. WSJ via The Decoder (Oracle about $300B; secondary), and OpenAI, Stargate 4.5 GW with Oracle. the-decoder.com, openai.com
    26. Amazon / OpenAI 8-K exhibit, Feb 2026 ($38B expanded by $100B over 8 years). sec.gov
    27. CoreWeave, expands OpenAI agreement by up to $6.5B (total about $22.4B). coreweave.com
    28. Financial Times, Sep 18, 2026, OpenAI projections (paywalled). ft.com; readable: Reuters via Investing.com
    29. Financial Times, Oct 8, 2026, OpenAI run-rate close to $50B, and why the $70B differed (paywalled). ft.com; readable: TechCrunch
    30. Bloomberg, Oct 9, 2026, OpenAI expects $70B or more annualised by year end; raise at $1.4T pre-money (paywalled). bloomberg.com; readable: Bloomberg Línea
    31. The Information via Reuters, Jul 30, 2025 (OpenAI about $12B annualised). Market data
    32. WSJ, Aug 18, 2026, OpenAI Q2 revenue $6.7B (paywalled). wsj.com; readable: SiliconANGLE
    33. Bloomberg, Oct 15, 2026, Anthropic Q3 revenue $16.8B; and Axios. Bloomberg Government, axios.com
    34. WSJ, Oct 20, 2026, Anthropic draft S-1 (H1 $16.3B; loss $80.7B; compute over $600B; paywalled). wsj.com; readable: The Next Web
    35. Anthropic, Google and Broadcom compute post, Apr 6, 2026 (over 1,000 customers at $1M+). anthropic.com
    36. Bloomberg Law, Apr 1, 2026 (Amazon's $50B, $35B contingent). bloomberglaw.com
    37. Silicon Data H100 rental index (live $2.88 on Oct 24; $2.00 on Dec 9, 2025; covers neoclouds only). silicondata.com, baseline
    38. SemiAnalysis, "The Great GPU Shortage", Apr 2, 2026 (one-year H100 contracts $1.70 to $2.35). semianalysis.com
    39. Silicon UK, Aug 12, 2026, Silicon Data token prices $2.04 to $1.16–1.18 per million tokens. Secondary; the Jefferies note is not public. silicon.co.uk
    40. Epoch AI, "The plunging price of thought", Sep 22, 2026 (about 47% a quarter, 13× a year). epoch.ai
    41. Ramp AI Index, October 2026 (adoption 55.9%; 80% of lab spend from 1% of firms). ramp.com
    42. Palantir Q2 2026 release (US commercial +149%). sec.gov
    43. Dealroom, Jun 9, 2026 (Cursor over $4B annualised). dealroom.co
    44. Stratechery, "An Interview with Craig Moffett About SpaceX and Bubbles", Oct 22, 2026 (paywalled, not read). stratechery.com; quotes from Whitney Tilson's Daily. The Oct 24 X summary by @TMTMoats could not be retrieved.
    45. Ed Zitron, "The AI Demand Bubble", Aug 4, 2026 (relays Barclays, UBS, Wells Fargo and BofA notes). wheresyoured.at
    46. Ed Zitron, "Don't Look Up", Aug 11, 2026. wheresyoured.at
    47. Ed Zitron, "The More You Buy, The More You Lose" (citing Nikkei), and Benzinga on the $1.65T figure. wheresyoured.at, benzinga.com
    48. WSJ, Jun 3, 2026, citing J.P. Morgan: over 60% of 2027 capacity not under construction (paywalled). wsj.com; readable: edwardconard.com
    49. J.P. Morgan Global Research, AI capex ($5.5T through 2030). jpmorgan.com
    50. All-In podcast, Brad Gerstner, Sept 2026 (transcript). happyscribe.com
    51. Couper, Hejkal and Wolman, Richmond Fed Economic Quarterly, Fall 2003 (telecom capex peak and fall). richmondfed.org
    52. Antoniewicz, Federal Reserve Board, for the BIS CGFS, 2001 (high-yield telecom spreads 1,200bp in Nov 2000). bis.org
    53. DrPeering, internet transit pricing, historical (informal series). drpeering.net
    54. CNBC, Oct 14, 2026, CoreWeave bond yields (did not load). cnbc.com; readable: investingLive
    55. Oracle CDS at 295bp, Oct 19, 2026, citing CNBC. Secondary; the dealer quote itself was not seen. primexbt.com
    56. RPT-101, AI capital cycle memo, Oct 11, 2026 (internal Notion; T1–T4 definitions; NVIDIA H1 investment over free cash flow). notion
    57. Earnings dates: Microsoft confirmed Oct 28 at microsoft.com. The others are from the earnings calendar and not yet announced by the companies.
    58. Backlog share estimates: The Information on AWS, via Michael Parekh; Oracle about half from OpenAI, per ERP Today. Both are unattributed estimates.
    59. Chamath Palihapitiya, "Deep Dive: Why AI is booming, but productivity isn't", Social Capital, Sept 25, 2026 (public part; guide paywalled). research.socialcapital.com
    60. Kim, Kim and Koning, "Mapping AI into Production: A Field Experiment on Firm Performance", SSRN, Mar 30, 2026. ssrn.com
    61. Dell'Acqua and others, "The Cybernetic Teammate", NBER Working Paper 33641, Apr 2025. nber.org
    62. Ramp, "How much do AI tokens cost businesses?", June 8, 2026 (April data). ramp.com
    63. Ramp AI Index, overall adoption (RAMPAIALL), FRED. fred.stlouisfed.org
    64. BLS nonfarm business output per hour (OPHNFB), FRED; growth rates are our calculation from Q4 levels. fred.stlouisfed.org
    65. Anthropic, PBC, Form S-1, filed Oct 26, 2026 (CIK 2130357). sec.gov
    66. S-1, MD&A, liquidity: note purchase agreements and AMD supply agreement (pp. 163–167): the Amazon 2023 and 2024 notes; the 2026 agreements with Google ($20.0B), Amazon ($15.0B) and Broadcom ($42.0B); Broadcom's terms; AMD's $5.0B. Read in the filing linked above.
    67. S-1, notes to the financial statements, fair value (F-31 to F-32): the Level 3 table, with convertible notes at $108,566M on Jun 30, 2026, a first-half change of $72,388M, and note purchase agreements at $(11,753)M changed and $11,810M settled.
    68. S-1, stock tables and Note 1: Series G-2 at $259.1364 (Google 38.6M shares, Amazon 19.3M); Series H-1 at $589.01, raising $45.7B in May–Jun 2026; the 10-for-1 forward split on Sept 17, 2026.
    69. S-1, MD&A results (pp. 152, 156): net loss by period; first-half operating loss of $693M; other income and expense.
    70. RPT-158, Anthropic S-1 report, Oct 26, 2026 (internal Notion): the provider minimums, the commitment stack, cash flow, marketplace share and backlog shares; also our earnings prep notes for Amazon's $53.4B Q2 Anthropic gain. notion
    71. RPT-159, Meta stock sale report, Oct 26, 2026 (internal Notion; from Meta's filings): NVIDIA, AMD and Broadcom's non-binding interest in up to $5B. notion
    72. Pure AI, Nov 18, 2025: Microsoft up to $5B and NVIDIA up to $10B invested in Anthropic, and Anthropic's $30B Azure commitment. Press. pureai.com