Anthropic's S-1 turns AI demand into supply

Revenue reached $16.3bn in six months and the June quarter turned a profit

Revenue went from $41m in 2023 to $386m in 2024, $4,585m in 2025 and $16,286m in H1 2026, with Q2 alone at $11,555m and an operating profit of $974m after a $1,667m loss in Q1 (S-1). Q3 2026 appears only as "86 times" the $197m of Q4 2024, roughly $16.9bn, consistent with the $16.8bn and about 45% sequential growth shown at the 15 October investor day (Yahoo Finance, 16 Oct). It is the first quarter in which the dollar add fell ($5.2–5.6bn against $6.8bn in Q2), the number the roadshow will be pressed on (Cape Fear Advisors). The "Run-Rate Revenue as of October 31, 2026" line is blank, as are the price range, share count and proceeds (VC Corner).

The $80.7bn half-year loss ($42.0bn for 2025) is about 90% non-cash: $72.4bn of fair-value charges on notes, SAFEs and derivatives driven by Anthropic's own rising valuation, plus $8.8bn of other expense tied to the Google and Amazon note agreements (S-1). Operating cash flow was +$3.9bn after −$2.6bn in 2025. Compute is leased, so finance-lease liabilities were only $443m at 30 June while leases not yet commenced stood at $321.4bn, up from $50.2bn in December. Cash was $112.4bn plus $19.3bn of short-term investments out of $141.1bn raised, the last $65.2bn at $589.01 per share, a $965bn post-money per the press (Decrypt).

The mix is 80% consumption and 20% subscription, and 48% of revenue runs through the AWS, Google Cloud and Azure marketplaces, recognised gross because Anthropic holds the customer contract, with about $2.7bn of partner fees expensed in sales and marketing (S-1; Reuters via Euronext). Half the "AI revenue" the market tallies for Anthropic is therefore also booked by the hyperscalers. Customers above $100k rose from 950 to 11,900 in a year, net revenue retention is "approximately 900%", and the top five customers are about 25% of nine-month revenue, mostly without long-term contracts (TechCrunch). Remaining performance obligations were only $10.8bn at 30 June, about six weeks of run-rate, so durability rests on usage, not backlog (Tokenpost). Menlo's survey gives Anthropic 40% of enterprise LLM API spend and 54% of coding (Menlo Ventures); OpenAI is raising $30bn at $1.25tn pre-money (TechCrunch, 22 Oct).

The gross margin in circulation is a presentation choice, and the notes disagree on it. The S-1 has no cost-of-revenue line; a functional view sits in Note 8, where "substantially all of the cost of revenue consists of compute and infrastructure expenses associated with customer-serving workloads" (S-1). Analysts using that note put the H1 gross margin at about 85%, cost of revenue $2.38bn on $16.29bn (Clouded Judgement; SaaStr); another reader reports no gross margin at all, with the natural opex table (compute $8,945m, personnel $3,152m, S&M $3,168m, other $1,714m) reconciling exactly to the $693m operating loss (Datastudios). Compute was 55% of H1 revenue and 48% in Q2, so treating all of it as cost of sales gives nearer 45%; the 85% figure assigns three-quarters of compute to model development. The filing warns margins "fluctuate... particularly in periods when substantial incremental compute is delivered" (Clouded Judgement): revenue per compute dollar rose from about $0.40 in Q1 2025 to $2.08 in Q2 2026, and the next tranche pushes it back.

The commitment stack: $417bn, $637bn or $795bn depending on who counts

The S-1 names each contract with a floor and a term but leaves the contractual-obligations table blank, so every published total is a reporter's sum (S-1).

The seven fixed-dollar lines sum to $637bn; with the TPU midpoint and the AMD floor, about $795bn. The other totals measure different things: $265.6bn of purchase commitments at 30 June 2026 (Note 13); $417bn of new or amended minimums since December (Note 17), the Reuters and CNBC headline (S-1); "$430bn locked in, up to $637bn" including cancellable deals (TechCrunch via Yahoo); $518bn over ten years (Yahoo Finance); $542–567bn (Fortune). Against $131.7bn of liquidity and about $33bn of H1-plus-implied-Q3 revenue, the floors are five to six times cash, the ratio behind the line that "for every dollar Anthropic earned in 2025, it committed roughly $113 to future infrastructure" (Yahoo Finance).

The financing is where the filing changed the debate. Broadcom is supplier, lessor and standby lender: the $42bn note facility is drawable "solely to finance quarterly rent payments" on Broadcom's own leases, pays 9% in kind (13% if revenue conditions are missed), converts after the IPO at 90% of the 30-day VWAP, and cross-defaults into the leases and the $15bn revolver; it is undrawn and Anthropic says it is "not dependent" on it (S-1; PPC Land). Broadcom also guarantees $60bn of syndicated debt (TechTimes). Amazon's remaining $15bn is released on compute milestones and "may be set off against certain amounts Amazon owes us"; Google's $20bn facility shrinks by up to $10bn if Google buys Class N at the IPO price; AMD committed to buy up to $5bn of Class N on delivery milestones (S-1). In cash the suppliers are small funders, $77bn of undrawn notes against $91bn of third-party equity raised in H1 2026; structurally the loop is explicit, and the filing says so: reliance on CSPs and suppliers "such as Broadcom" who are investors and customers "creates complex dynamics that could give rise to conflicts of interest", with cross-defaults that "could set off a cascading series of adverse consequences".

This landed two weeks into a credit scare: on 14 October CoreWeave fell 15.46% to $65.78 with bonds at 15% after Gartner warned that 80% of neoclouds could disappear by 2030, and its CDS reached about 1,000bp (Motley Fool; CNBC). For the capex bulls the S-1 supplies the missing proof: a compute buyer with customer-funded revenue, positive operating cash flow and take-or-pay contracts into the 2030s. For the bears it shows costs still scaling with revenue, half the sales flowing through investors, a six-week backlog and growth slowing in Q3. The market's net read was that it improves supplier order books and does nothing for leveraged capacity owners' cost of debt.

A $100bn deal lands between Nvidia's results and the Nasdaq-100 fast entry

Price talk is $1.5–2tn on a raise of $100bn or more, after the investor day took the discussed figure from about $2tn to about $1.5tn (Investing.com; Yahoo Finance, 16 Oct). The low end is about 22 times Q3 annualised revenue, Nvidia's multiple without GAAP profit (Tunguz); synthetic perpetuals priced the company at $1.86–1.88tn on 25–26 October (Bitcoin.com). Supply is the point for the AI trade: SpaceX raised $75bn ($85.7bn with the greenshoe) at $1.77tn in June, and Meta announced up to $95bn the same morning as the S-1 (CNBC, June; CNBC, 26 Oct); ANT at $100bn would take 2026's new AI equity to roughly $280bn.

Marketing reportedly begins around 9 November, pricing the week of 16 November, first trade 19 or 20 November (Investing.com; Investing.com/Fox). Nasdaq's fast-entry rule admits a top-40 newcomer after its 15th trading day at a weight capped at three times float; SpaceX entered on 7 July at about 1% on a sub-5% float, with more than $800bn tracking the index (Yahoo Finance). The S&P 500 requires twelve months of seasoning and a profitable trailing year (InvestingLive). Lock-ups are tiered, from non-founder insiders at the earlier of two days after the Q1 release or day 180 to co-founders at the earlier of the Q3 release or day 365, all "subject to timed automatic expirations" (S-1); press describes those as windows from day one and at days 27, 48, 62 and 76 (Yahoo live blog), day counts not verified in the filing, tier percentages blank. A single Class F share gives the founders' LLC a fixed 50.1% of the vote and the LTBT's Class T share elects a board majority; the filing warns this may make the company "ineligible for inclusion in indices", though SpaceX's entry suggests Nasdaq does not mind.

The first-day reaction was a chip rally that faded. SOX, NVDA and AVGO were up about 3% intraday and CoreWeave reversed a Musk-driven premarket drop to +5.7%, but reported closes disagree: one has the Nasdaq +0.3%, NVDA +1.4%, AVGO +1.5%, CRWV −1.7% and NBIS −3.7% (Beinsure), another the Nasdaq +1.3%, NVDA +2.6% and AVGO +2.2% (Dimsum Daily). SMH closed at $541.62, +2.7%, with 96% of its holdings up and the equipment and memory names leading (Carriba data; the +1.16% at 533.34 reported mid-session by Twelve Data was not the close). No sell-side or implied-volatility commentary dated 26 October was found.

What it means for each listed name

Amazon (AMZN). Amazon carries Anthropic at $190.4bn in its Q2 10-Q ($97.9bn of notes at fair value plus $92.5bn of preferred), and one scenario table puts the mark at $296bn at a $1.5tn IPO and $394.6bn at $2tn, through other income that consensus EPS excludes (Amazon 10-Q; Investing.com). The S-1 adds the operating floor: at least $110bn of AWS spend to April 2036, about $11bn a year, non-cancelable, plus $6.2bn of notes converting into 1.84bn non-voting shares under a 12% cap that conflicts with press estimates of a 15–20% economic stake (S-1; Yahoo Finance). Thursday's Q3 (consensus EPS $1.96 on $202.2bn) carries little of this because no pricing event fell in Q3; the IPO price is the next mark (MarketScreener). Paper gain now, realised volatility at the late-January report.

Alphabet (GOOGL). Google holds the largest firm commitment, $114.5bn to July 2033, plus an estimated $125–150bn of 2028–31 TPU capacity, a $20bn undrawn facility capped at 14.99% and an option to buy $10bn of Class N at the IPO price (S-1). Its ~14% stake is carried at about $124.3bn and re-marks to $210–280bn at the IPO range (AI Weekly; Investing.com); UBS projects Anthropic alone at $76bn of Google Cloud revenue in 2027 (ppc.land). Alphabet has the cleanest mix of contracted revenue, equity upside and a Q3 report that clears before the roadshow; the cost is one customer approaching half a cloud segment.

Broadcom (AVGO). The best-documented beneficiary and the most entangled: $161.2bn of TPU leases, the $42bn standby notes, a $2.5bn escrow, and Hock Tan's statement that Anthropic becomes Broadcom's largest XPU customer in 2027 and 2028 within AI semiconductor guidance of $115bn for FY27 and $230bn for FY28 (S-1; Broadcom Q3 call). Mizuho's $435 target names Anthropic as the fourth hyperscale ASIC customer and Piper Sandler initiated at $460 (Yahoo; TechTimes). Broadcom books revenue on the hardware sale; the rent is the SPV investors' income and the residual-value and guarantee exposure is Broadcom's. Into pricing the order book dominates; below $1.5tn the circular framing returns and AVGO carries it.

Nvidia (NVDA). Nvidia appears as a "strategic partnership" giving Grace Blackwell and Vera Rubin access through Microsoft, SpaceXAI, Nscale and Lambda; none of its fifteen mentions describes an investment, so the "up to $10bn" is not visible as equity (S-1; Microsoft). GPU-backed capacity at the three landlords sums to about $206bn but on the weakest terms: SpaceXAI cancellable on 90 days' notice from early 2027, Nscale paid only on accepted capacity, $34.9bn of Lambda contingent on a lease. Nvidia's 17 November results fall two days before the first trade and are the real referendum on GPU demand; the filing's Trainium, TPU and Helios volumes are the share-loss case written down by the customer.

Marvell (MRVL). Not mentioned. The custom-silicon winners the filing names are Broadcom and AMD, and Mizuho called the Broadcom relationship "potentially negative for MRVL" (Yahoo); the indirect link is Amazon's $110bn Trainium commitment, where Marvell is a second-order name (Motley Fool). Marvell closed at $228.51 on 23 October (Carriba data; a Benzinga piece written before the 13–22 October fall quoted about $283 and +233% year to date, which is stale), so the Marvell page's levels stand.

Arista (ANET). Zero mentions, and the day's sympathy lists went to Akamai, Broadcom, Astera and CoreWeave rather than networking (Investing.com). The one hook is "Anthropic Direct", the filing's own leased data centres with directly procured equipment, a growing share of compute with a blank dollar figure (S-1). ANET traded $172.05–179.56 on 26 October (Robinhood). The S-1 is neither the catalyst bulls wanted nor a threat; Nvidia on 17 November and the pricing week matter more.

Microsoft (MSFT). The November 2025 deal shows up as an Azure minimum of $44.9bn from November 2026 to May 2033, about $6.9bn a year, up from the $30bn announced, with Microsoft's investment capped at $5bn and a $3.2bn gain already disclosed in July (S-1; Microsoft; CNBC mirror). Microsoft is the least dependent cloud as a share of revenue and the least exposed to the stake mark: the lowest-beta way to own the contract and the one the IPO moves least.

AMD. The July 2026 Helios agreement has no fixed amount but exceeds $20bn if the first 2027 tranche of MI450 racks is fully taken, with an option for about the same again, and AMD committed to buy up to $5bn of Class N on delivery milestones (S-1). AMD rose about 2.5% intraday on filing day (Reuters via SRN). The revenue is 2027 and beyond, so AMD is least affected by the November–January flow calendar and most affected by whether the tranche is taken.

CoreWeave and the neoclouds (CRWV, NBIS). The S-1 does not mention CoreWeave; Anthropic's GPU landlords are SpaceXAI, Nscale and Lambda. The filing validates the business model while confirming the marquee contracts went elsewhere, and does nothing for a cost of debt at 15–16% with CDS near 1,000bp (CNBC). Musk's post that "200+ neoclouds are going to see guppy-level death and consolidation" set the premarket tone, the filing produced a +5.7% squeeze, and by one count CRWV closed −1.7% and NBIS −3.7% (Reuters via Investing.com; Beinsure). A $100bn primary competes for the same marginal AI dollar; the neoclouds are the "most leveraged sympathy trade" in both directions (Investing.com).

SMH and QQQ. For the semiconductor ETF the filing is a demand confirmation that produced a +2.7% close at $541.62, back above the $538 shelf that broke on 22 October, with 96% of holdings up (Carriba data); that lifts the SMH reversal scorecard from one to five of ten, a low forming rather than confirmed, and the filing does not by itself end a correction, and the next inputs are Nvidia on 17 November and the pricing week. For the Nasdaq-100 the S-1 adds a dated mechanical flow: fast entry around 11–14 December at up to three times float, bought by more than $800bn of index money pro rata from the other hundred names, overlapping the December reconstitution and the reported day-27 lock-up window (Yahoo Finance; marketclock). Index-level, the filing is neutral into pricing and a supply headwind in mid-December.

Conclusion

The S-1 settles one argument and opens another. Demand is no longer the question: a company with $16bn of half-year revenue, a $974m operating quarter and take-or-pay contracts into the 2030s is the customer the capex cycle was waiting to produce, and the suppliers it names in dollars now have an order book a public investor can read. What it does not settle is whether that order book is self-funding: costs still scale with revenue, the gross margin depends on which footnote you believe, the backlog is six weeks, growth slowed in the quarter the roadshow leads with, and the financing is a closed loop the filing describes in its own risk factors. The cleanest signal from here is not the filing but the price ANT trades at against the $1.5–2tn talk, because that one number sets Amazon's and Alphabet's Q4 marks, Broadcom's convertible strike, the Nasdaq-100 weight and the lock-up sellers' incentive at once. Until pricing, the reference-price effect rewards contracted suppliers and stake holders; after the debut, a $100bn primary, a mid-December index buy and a staggered insider release turn demand into supply, and positions built for the first phase should be out, or hedged, before the second.

Related trades

The ANET put needs no S-1 decision: what changed is that the 27 November expiry now sits ten days after Nvidia and a week after ANT's first trade, with the breakeven of $173.74 (strike less premium) where the stock is, so it is a bet on those two events, not on Anthropic. The Amazon put is the position the S-1 most directly supports: the $110bn floor is the first contracted, non-cancelable revenue line from an AI lab to a cloud, the 29 October print will not include a stake mark, and a March 2027 expiry collects premium across the IPO, the fast entry, both January lock-up windows and the Q4 report with the first public mark; the risk is a debut below $1.5tn turning that mark into a non-operating loss. The Broadcom spread gains from the order book and from two sell-side targets above its $410 upper strike; it should be the first position to take profit on if AVGO reaches that strike during the roadshow, because after the debut Broadcom carries the circularity narrative. The Marvell page loses a catalyst rather than a thesis. The SMH and QQQ pages are unchanged in level and changed in calendar; both should carry the 17 November, 16–20 November and 11–17 December dates.

Three new ideas follow, each defined-risk because the notes give no implied-volatility or skew data and the pricing week will move option prices as much as stock prices.

The first is bullish Alphabet through a January 2027 call spread, entered only after the 28 October report and the Fed decision have passed. The reasoning is the filing's own: the largest firm commitment at $114.5bn with a further $125–150bn of TPU capacity, a ~14% stake carried at $124.3bn that re-marks to $210–280bn at the IPO range, a $10bn option to buy more at the IPO price, and a Q3 report that clears before the roadshow. The trigger is GOOGL holding its post-earnings level into the 9 November start of marketing; the exit is pricing week, before the stake mark becomes a two-way bet on ANT's aftermarket.

The second is bearish CoreWeave through a January 2027 put spread. The filing names three other GPU landlords and no neocloud, CoreWeave's bonds yield 15–16% with CDS near 1,000bp, a $100bn primary absorbs capital that might have refinanced it, and the SpaceXAI contract, cancellable from January 2027, shows what Anthropic's GPU demand looks like when it is not take-or-pay. The trigger is a close below the 14 October close of $65.78 after ANT prices, or a failed rally into the pricing week; a spread rather than a naked put, because the +5.7% intraday squeeze on 26 October is what a crowded short looks like.

The third is a hedge rather than a view: a QQQ put spread expiring in December 2026 or January 2027, sized against the short-volatility book (the ANET put, the AMZN put and the AVGO spread). The mid-December overlap of fast entry, the annual reconstitution and the day-27 lock-up window is the one dated event in this calendar where index money sells the other hundred names to fund ANT, and the QQQ page's $700 floor is where that flow would show. The trigger is a Nasdaq-100 close below $700 after the ANT debut, or ANT trading below its IPO price in its first week; if neither happens by the fast-entry date, the hedge expires as the cost of having been long the AI trade through its largest-ever supply event.