Reading Anthropic’s books

This is an analysis of publicly filed financial information. It is not investment advice or a recommendation to buy, sell or hold any security. Hapax Oy is not authorised to provide investment services. Hapax uses Claude, Anthropic’s AI assistant, in its work. Neither founder holds shares in Anthropic or in any other company discussed here.

Anthropic, the company behind the AI assistant Claude, filed its accounts publicly on 26 October.¹ The S-1 analysis is old hat by now, every newsletter has one, but so is my finance hat, and as someone who spent (too) many years in finance and an unreformed finance nerd, I couldn’t leave this one alone. Until now, nobody outside the big labs has known what a frontier AI provider’s books actually look like, and the question at the end of it matters to anyone paying for these tools. If the economics work, today’s token prices have room to fall. If they don’t, what we’re paying now is an introductory offer.

In brief.

This is the first S-1 (registration statement) from a frontier AI lab, and the first close look at whether what we pay per token is a real price or something investors are subsidising.
The number everyone quotes is the 85% gross margin on serving, but it counts only delivery costs. Training new models, which is what stops them going stale in months, is booked as research and isn’t in that 85%. I think much of it is really maintenance. Add the platform fees and anywhere from none to all of that training bill, and the margin is somewhere between 9% and 68%.
Almost all of the $80.7 billion loss comes from old loans whose value on the books grows every time the company’s valuation rises, and no cash leaves the building. Set that aside and operations were profitable for the first time in Q2 2026.
Existing customers are spending about nine times what they spent a year ago, though the base is small and the threshold is low.
The top five customers account for about a quarter of revenue, but that share is falling fast.
Hundreds of billions in fixed computing contracts sit opposite revenue that customers can cancel by the quarter.
A few other labs have published numbers in different forms. In the first half of 2026 Anthropic’s margins are the highest, and its R&D is a smaller share of revenue, but the accounting differs enough that the margins are not directly comparable.

Divide the reported $80.7 billion net loss by $16.3 billion of revenue and you get minus 496%.² NBC News led with the loss.³ Troy Hooper, co-head of equity capital markets at Mergermarket, told the BBC that investors “do not yet have a clear view of Anthropic’s underlying profitability…”⁴

The $72.4 billion that never left the building

To see what the business earns, the $80.7 billion headline has to come apart. Of that loss, $72.4 billion is one income-statement item, the change in fair value of convertible notes, derivative liabilities and SAFEs (an early-stage funding instrument).⁵

Figure 1

Nine in ten dollars of the loss are on paper

GAAP results Non-cash items, excluded from adjusted Adjusted result and cash flow

The fair-value change is the period’s movement in what the company estimates it will owe holders of notes that can convert into shares. The estimate rises when the valuation rises, and no cash is paid when it is booked. Adjusted net income is the company’s own measure. It reverses that change, stock-based pay and the other non-cash items, and includes interest on the company’s cash and investments.

Source: Anthropic’s S-1, filed 26 October 2026, pages 22, 156 and F-12. The half-year figures are unaudited. Shares issued to note-holders and employees dilute other shareholders, so non-cash items still carry a cost. The chart does not show the compute commitments or the stock-based pay that will be recognised when the registration statement becomes effective.

Figure 1 as a table

Item, first half of 2026	US$ million	Source
Net loss	−80,748	S-1 p. 22
Fair-value change on convertible notes	−72,388	S-1 p. 22, p. 156
Other non-cash items	−11,031	S-1 p. 156
of which, marks on new note agreements	−8,720	S-1 p. 156
of which, stock-based pay	−2,185	S-1 p. 156
of which, donation matching	−126	S-1 p. 156
Adjusted net income	2,671	S-1 p. 156
Interest income, inside adjusted net income	1,111	S-1 p. 22
Operating result	−693	S-1 p. 22
Operating result, second quarter alone	974	S-1 p. 157 (Q2 2026)
Cash from operations	3,929	S-1 F-12, p. 168

Between 2023 and 2025, Amazon and Google lent Anthropic about $11.25 billion through convertible notes, a form of borrowing where the lender can swap the loan for shares instead of being repaid in cash.⁶ Amazon’s swap prices were fixed at the time of lending, at about $2.09 and $5.05 per share on today’s share basis.⁷ By May 2026, new investors in the Series H funding round were paying about $58.90 on the same basis.⁸ Google’s notes have all been converted into shares. Amazon still holds $6.2 billion of principal, convertible into about 1.84 billion shares.⁷

Anthropic chose to carry these notes at what they are worth today, using a US accounting option called ASC 825, the fair value option.⁹ Because Amazon’s share count is fixed, every rise in Anthropic’s valuation makes those cheap shares more valuable on paper, and the increase counts as a loss. At 30 June 2026, the $6.2 billion of principal sat on the balance sheet at $108.6 billion, roughly what 1.84 billion shares are worth at the Series H round price.^10,5 As the valuation rose through the half-year, the liability followed, producing the $72.4 billion charge. Of the company’s $132.9 billion accumulated deficit since inception, $113.2 billion is these fair-value changes.⁵

At the IPO, the notes convert into shares and the charge stops on the converted portion. Anything exceeding Amazon’s 12% ownership cap stays outstanding as a rollover note and keeps being revalued.¹¹ How much converts and how much rolls over depends on the offering size and whether Amazon waives the cap. The filing leaves both blank, along with the pro forma balance sheet, the dilution tables and the per-share figures, so nobody outside the company can compute the dilution yet.^11,12

Amazon receives 1.84 billion shares for $6.2 billion of principal, at prices set years ago, without any cash changing hands. The filing says this “will cause additional dilution to our existing stockholders.”⁷ I think of it as the delayed invoice for cheap capital. Amazon received favourable conversion prices in exchange for lending early, and the accounting measures the growing gap between what those shares are now worth and what was paid for them.

Anthropic also signed new standby financing agreements in 2026 with Google, Amazon and Broadcom, none drawn at 30 June 2026.¹³ A separate charge of $9.4 billion arose because Google and Amazon had bought shares at $259 per preferred share under commitments made in March and April, weeks before the Series H priced at $589.¹³ After smaller offsetting items, the net charge was $8.8 billion.¹³

Strip all of this away, along with stock-based compensation ($2.2 billion of non-cash expense from shares and options promised to employees) and a $126 million donation-matching commitment, and the first half of 2026 shows $2.7 billion of what the company calls adjusted net income.¹⁴ Operating loss, which measures the business before interest, tax and the financing marks above, was $693 million, about minus 4.3% of revenue. In the second quarter alone, Anthropic posted a GAAP (the standard US accounting rules) operating profit of $974 million, its first profitable operating quarter in the six the filing discloses, and generated $3.9 billion of cash from operations in the half-year.^2,15

Revenue first exceeded total compute spending in Q1 2026, and by Q2 it had outrun all operating costs. Between Q2 2025 and Q2 2026 (Figure 2), revenue rose $10.8 billion while every cost line grew by less. Compute and infrastructure were up $4.2 billion, personnel up $1.5 billion, sales and marketing up $2.0 billion and other costs up $0.9 billion.¹⁶

Figure 2

Anthropic’s first operating profit, Q2 2026

Profit or revenue growth Loss Growth in a cost line Operating result Runs off the scale

Revenue topped compute in Q1 2026, operating costs in Q2

US$ billion a quarter

Gross margin, derived from Note 8 by half-year, was 12% for the first half of 2025 and 85% for the first half of 2026 (unaudited).

The net result was a loss in every quarter, from $6.4bn to $66.3bn, driven by the fair-value change on convertible notes.

Revenue grew faster than everything else

Where you stop counting matters

Compute is bought 12 to 18 months ahead, so in any quarter it costs roughly the same whatever revenue does. Between Q2 2025 and Q2 2026, revenue grew by more than all the costs together, and the operating result turned positive for the first time.

Source: Anthropic’s S-1, filed 26 October 2026, pages 157 and 160. All quarterly figures are unaudited. The bridge groups costs by type (compute, personnel and so on) because the split by function (cost of revenue against R&D) is disclosed only for half-years and full years. Gross profit cannot be shown by quarter.

Figure 2 as a table

US$ million	Q1 2025	Q2 2025	Q3 2025	Q4 2025	Q1 2026	Q2 2026	Change, Q2 2025 to Q2 2026
Revenue	376	787	1,382	2,040	4,731	11,555	+10,768
Compute and infrastructure	944	1,390	2,047	2,945	3,387	5,558	+4,168
Personnel	218	320	489	812	1,295	1,857	+1,537
Bartz copyright settlement	0	0	1,500	0	0	0	0
Sales, marketing and partnerships	24	96	202	246	1,033	2,135	+2,039
Other operating and admin	125	181	233	873	683	1,031	+850
Operating result	−935	−1,200	−3,089	−2,836	−1,667	974	+2,174
Revenue less all compute (Armstrong’s measure)	−568	−603	−665	−905	1,344	5,997	+6,600
Adjusted operating result	−830	−1,019	−1,284	−1,721	−617	2,235	+3,254
Adjusted net income	−786	−955	−1,168	−1,539	−310	2,981	+3,936
Net result	−6,592	−6,377	−12,138	−16,858	−14,439	−66,309	−59,932

The headline loss is mostly a financing artefact, and serving customers is profitable at the gross level.

The cost of actually serving people

Does selling AI tokens to paying customers make money before you count everything else? Yes, and the swing is from minus 95% to 85% in two years. Anthropic’s gross margin would tell you the precise number, but the company stopped reporting it on the face of its income statement during 2025, switching to a layout that groups costs by type. The old functional breakdown survives in Note 8 of the financial statements, and I spent a pleasant evening with it.¹⁷

The implied gross margins, worked out from Note 8 (the share of each revenue dollar left after the direct cost of serving):

2024: about minus 95%.
2025: about 40%.
First half of 2026 (unaudited): about 85%.¹⁷

In 2024, the company lost nearly a dollar for every dollar a customer paid. By the first half of 2026, cost of revenue had fallen to 15% of revenue. The filing defines cost of revenue as “substantially all” the compute costs of running models for paying customers, along with some staff and overhead costs.¹⁸ A January 2026 leak to The Information, a technology publication, had the expected 2025 margin at about 40% and 2024 at minus 94%, which matches the filing closely.¹⁹

About 48% of first-half revenue came through cloud marketplaces operated by Amazon, Google and Microsoft.²⁰ Those platforms charge a fee, but Anthropic records the full revenue and puts the fee in sales and marketing, not in cost of revenue.²¹ With the marketplace fees counted as a cost of serving, the margin falls to about 68%.²²

Anthropic manages all its computing power as a single shared pool across customer-serving and model-development workloads.²³ How much of that pool counts as cost of revenue is a management allocation. In the first half of 2026, cost of revenue accounted for about 27% of total compute spending. The rest went to R&D (research and development) and to free-tier access, which sits in sales and marketing.²³

The maintenance problem

The 85% margin covers serving. The spending that keeps the product alive sits in a different part of the accounts. Training a new frontier model is classified as R&D and sits below the gross-profit line, where it never touches the margin. In most businesses, R&D is a bet on the future. Here, where over 80% of revenue in the quarter to 30 September 2026 came from models released in the previous six months, much of that R&D is maintenance.²⁴

Warren Buffett drew this distinction for traditional businesses, separating the profit after only maintenance spending from the rest and calling the result “owner earnings.” It is how I’d normally judge underlying profitability. At AI labs, the accounts don’t give me the split. R&D in the first half of 2026 was $9.7 billion, about 60% of revenue, and the filing bundles training compute, research compute and researcher salaries into one line.^17,23 If training stops, the product falls behind and the revenue with it. That is why I think much of that training spend is really maintenance, even though the accounting books it as research.

Evan Armstrong, in the newsletter The Leverage, called a frontier model “an asset with a six-month half-life” and compared it to how Netflix accounts for a film people watch for a season and forget.²⁵ Counting all compute and infrastructure as cost of goods, he got a gross margin of about 45%.²⁵ OnlyCFO, a financial newsletter, illustrated a middle ground, putting half of R&D in, which gives about 55%.²⁶

I think the useful measure for these businesses is what I’d call a sustaining margin, meaning revenue minus serving costs, marketplace fees, and whatever training is needed to maintain today’s models. It is the AI equivalent of owner earnings after maintenance. I can bound it from this filing. At the low end, counting all of R&D as maintenance and including the marketplace fees, the margin is about 9%.²⁷ At the high end, counting only the marketplace fees and cost of revenue, it is about 68%.²² Where it sits depends on how much of that $9.7 billion of R&D is maintaining today’s models and how much is building the next generation, a question OnlyCFO puts well: how much R&D is needed just to hold today’s revenue?²⁶

Figure 3

Margin is 85% or 26%, depending on the definition

Gross margin Counted as cost

Each bar splits the same half-year’s revenue between what is counted as cost (beige) and what remains as gross margin (indigo). The five bars are separate definitions, each starting from the full revenue.

Sources: Anthropic’s S-1, filed 26 October 2026, Note 8 (F-39) and pages 22 and 149. The half-year figures are unaudited. Anthropic does not report gross margin; every figure here is a recalculation by the analysts named or by Hapax. The split between cost of revenue and R&D is the company’s own allocation of one shared pool of computing power, and the filing does not separate training from research.

Figure 3 as a table

What counts as cost	Calculated by	Margin, our arithmetic	Their figure	Source
Cost of revenue only (Note 8)	Tomasz Tunguz; Hapax	85.4%	85%	S-1 Note 8, F-39; Tunguz, 'Anthropic S-1 Analysis', 26 Oct 2026
Cost of revenue + marketplace fees	Austin Johnsen; Dolphin Research	68.2% (68.1 to 68.8)	68%	Johnsen, 'Anthropic S-1 Review', Artificial Diligence, 27 Oct 2026; Longbridge Dolphin Research, 27 Oct 2026
Cost of revenue + half of R&D	OnlyCFO	55.5%	55.5%	OnlyCFO, 'Anthropic's S-1: Finance Takeaways', 26 Oct 2026
All compute spending	Evan Armstrong; also Johnsen and Dolphin	45.1%	45%	Armstrong, 'Is Anthropic Worth $2 Trillion?', The Leverage, 26 Oct 2026
Cost of revenue + all of R&D	Hapax	25.6%	25.6%	Our arithmetic from S-1 Note 8, F-39

900% revenue retention, with caveats

Customers who were buying a year ago are now spending about nine times as much. The filing reports net revenue retention, or NRR, of approximately 900% as of 30 September 2026.²⁸ NRR measures how much a group of existing customers spends in a quarter compared with the same quarter a year earlier, combining growth, contraction and departures in a single figure. In subscription software, above 120% is strong; Snowflake, a cloud data company, reported 158% at its own IPO in 2020.²⁹

Anthropic’s 900% uses a $1 threshold, so anyone who spent a dollar in Q3 2025 counts. It excludes Claude Pro and Max subscribers, whose fixed-price plans don’t reflect the usage swings the metric is designed to capture, and the base quarter had revenue of $1.4 billion, early in the company’s commercial life.²⁸ The filing itself cautions that this growth “may reflect early-stage adoption dynamics” and “may not be sustainable over time.”³⁰ I read it as a genuine adoption wave, amplified by the $1 threshold and a small base.

Who is paying

Is the revenue spread widely enough, or does losing a few big accounts change the picture? Over the nine months to 30 September 2026, the top five customers accounted for about 25% of revenue, confirming the BBC’s report that “just five enterprise customers make up a quarter of its current business.”^4,31 But that share is dropping. Revenue from outside the top five rose from about 64% in Q1 2026 to about 79% in Q3, meaning the top five’s share dropped from roughly 36% to 21%, even as each of those five set a revenue record.³¹ In 2025, two unnamed customers each accounted for about 12% of revenue.³² By the nine months to September 2026, none reached 10%.³¹

Customers spending at least $100,000 a year (the filing’s proxy for enterprise adoption) went from about 20 at the end of 2023 to more than 11,900 at 30 September 2026.³¹ Those customers produced about 74% of first-half 2026 revenue. About 250 spent $10 million or more. The long tail is another 40 million customers, each below the $100,000 threshold.³¹

Figure 4

Customers above $100k grew 12.5-fold, top five’s share fell

Customers paying $100,000 or more a year

Share of revenue from the five largest customers

The count is of organisations that paid Anthropic $100,000 or more over the previous twelve months, which the company uses as its measure of enterprise adoption. The vertical scale is logarithmic, so equal steps represent equal multiples: the jump from 20 to 200 takes the same space as 200 to 2,000.

Source: Anthropic’s S-1, filed 26 October 2026, pages 8, 56, 142, 143, 148, 158, 194 and 206; the quarterly figures are unaudited. The dates are those the filing gives, so the gaps between points are uneven. "More than" figures are floors. The filing does not name its customers.

Figure 4 as a table

Date	Customers paying $100,000 or more	Source
31 Dec 2023	about 20	S-1 p. 142
31 Dec 2024	about 200	S-1 p. 142
31 Mar 2025	about 400	S-1 p. 158
30 Jun 2025	more than 600	S-1 p. 148
30 Sep 2025	more than 950	S-1 p. 194, p. 206
31 Dec 2025	about 1,500	S-1 p. 142
30 Jun 2026	more than 6,000	S-1 p. 148, p. 158
30 Sep 2026	more than 11,900	S-1 p. 8, p. 142, p. 206

Period	Top five’s share of revenue	Source
Three months to March 2026	36%	100% less the share outside the top five; S-1 p. 143
Three months to September 2026	21%	100% less the share outside the top five; S-1 p. 143
Nine months to September 2026	25%	as stated in the filing; S-1 p. 56

The filing says many top customers have contracts with near-term expiration dates or none at all, and can reduce or stop spending at any time.³³ Michael Monaghan, a portfolio manager at Founder ETF, told the BBC that “many top Anthropic customers have no long-term contract,”⁴ and 80% of first-half revenue was consumption-based.³⁴

The bill for staying in the game

Anthropic has committed to buying computing power it cannot walk away from. Fixed or minimum purchase commitments at 30 June 2026 stood at about $265.6 billion. Leases signed but not yet started carried roughly $321.4 billion in undiscounted payments.³⁵ Agreements entered into or amended after that date carry about $245 billion in further minimums.³⁶ The filing does not reconcile these into a single total. The named agreements with Microsoft, Google and Amazon run about six and a half to ten years each.³⁷

Tomasz Tunguz, a venture capitalist, put the mismatch in his analysis of the filing: the commitments are fixed, but “revenue is consumption-based with no lock-in, and the payments arrive whether customers do.”^34,38 Long-term supply contracts were part of my finance years. I have never seen commitments at this scale set against revenue that can evaporate by the quarter.

RSUs (restricted stock units, shares promised to employees that vest over time) with an IPO performance condition have not yet been expensed. Had the registration statement been effective on 30 June 2026, about $2.9 billion of accumulated cost would have hit the income statement at once, with a further $23.7 billion to follow over about two years.³⁹ Part of the IPO proceeds will go to covering employees’ tax withholding on those units.⁴⁰ Compute itself is contracted 12 to 18 months ahead and expensed as it arrives, often in “large, discrete steps.” The filing warns that operating margins “may compress” as committed capacity comes online, and that short-term compute, which “typically carries a higher cost per unit,” is expected to compress them further.⁴¹

An estimated 5 to 7% of revenue in the nine months to 30 September 2026 came from customers later found to be violating the usage policy or commercial terms, a share the company expects to fall after enforcement in August and September.⁴² A risk factor says that for a catastrophic AI incident, “no single insurer, and potentially the insurance industry in the aggregate, may have sufficient capacity” to cover it.⁴³ The company also says it delayed the pre-training of a new frontier model for more than a month to harden its defences against model-weight theft.⁴⁴

What the others show

Anthropic’s is the first S-1 from a frontier AI lab. A few others have published financial data in different forms. Two Chinese model makers, Zhipu and MiniMax, have listed in Hong Kong and report under IFRS, the accounting framework used in most of the world outside the US. Zhipu’s audited full-year 2025 gross margin was 41.0%. MiniMax reported 25.4%.^45,46 By the first half of 2026 (unaudited at both), Zhipu’s margin had fallen to 26.4% and MiniMax’s to 17.9%, with R&D running at multiples of revenue at both labs.^45,46 At Anthropic, by contrast, R&D in the first half of 2026 was 60% of revenue, smaller than the gross profit of about $13.9 billion. In 2025, Anthropic’s R&D exceeded its gross profit.

Figure 5

Anthropic’s margin rose to 85% as the Chinese labs’ fell

Full year 2025 First half of 2026

Gross margin is revenue less cost of revenue, as each company defines it. Anthropic’s is derived from a note to its accounts, since it no longer reports gross profit; it places training compute in R&D and marketplace fees in sales and marketing. Zhipu attributes its margin decline to a shift towards pay-per-use API serving, a lower-margin line. MiniMax credits improving infrastructure efficiency for its half-on-half rise.

Sources: Anthropic’s S-1, filed 26 October 2026 (US GAAP), Note 8. Zhipu’s HKEX annual results (IFRS, FY2025 audited) and interim results (H1 2026, reviewed by KPMG). MiniMax’s company releases (IFRS). OpenAI from press reports of its 2025 statements. The accounting standards and the definitions of cost of revenue all differ, so the comparison shows direction and scale, not exact equivalence. Anthropic’s first-half revenue was about 140 times MiniMax’s.

Figure 5 as a table

Company	Period	Gross margin	R&D as a share of revenue	Status and standard	Source
Anthropic	2025	39.5%	138%	audited, US GAAP	S-1 Note 8, F-39
Anthropic	H1 2026	85.4%	60%	unaudited, US GAAP	S-1 Note 8, F-39
Zhipu	2025	41.0%	439%	audited, IFRS	Zhipu (Knowledge Atlas) annual results announcement, HKEX, 31 Mar 2026, p. 1 financial summary
Zhipu	H1 2026	26.4%	223%	unaudited, IFRS	Zhipu interim results announcement (English), HKEX, 31 Aug 2026, p. 1 summary table; review by KPMG under HKSRE 2410, p. 52
MiniMax	2025	25.4%	320%	filed, IFRS	MiniMax full-year 2025 results release
MiniMax	H1 2026	17.9%	255%	unaudited, IFRS	MiniMax first-half 2026 results release (interim condensed income statement)
OpenAI	2025	42.6%	147%	press reports, not filed, US GAAP basis	Audited 2025 statements as obtained by Ed Zitron and verified by the Financial Times, via 36Kr, 17 Jun 2026

Zhipu’s annual results attribute the FY2025 margin decline to a shift in revenue mix, and the interim report shows it continuing, with cloud and API revenue growing from 15% to 86.5% of the total. Pay-per-token serving carries lower margins than on-premise deployment.⁴⁵ MiniMax’s margin against the year-earlier half rose (from 12.1% to 17.9%), which the company attributes to improving infrastructure efficiency.⁴⁶

Anthropic’s first-half revenue was about 120 times Zhipu’s and 140 times MiniMax’s, and what each company puts in cost of revenue differs, so the margins are not directly comparable. The same kind of fair-value charge I described above on Anthropic’s convertible notes appears in both Chinese labs’ accounts, on instruments issued to their own investors. MiniMax’s fair-value charge on financial liabilities was $1.59 billion of a $1.87 billion 2025 net loss. Zhipu’s 2025 loss included RMB 937 million of fair-value changes on investor instruments.^45,46

DeepSeek, also Chinese, has published no financial statements, though a Shanghai listing is reported for 2027. Its only official figure is the March 2025 inference-cost snapshot: $87,072 in daily GPU (graphics processing unit, the specialised chip AI models run on) rental serving $562,027 of theoretical revenue at list prices, equivalent to about an 84% gross margin on rental alone.⁴⁷

OpenAI has not filed publicly. Its 2025 figures, as reported by the Financial Times and other outlets, show an implied gross margin of about 43% and an operating loss of $20.9 billion, with leaked internal plans showing cash burn widening until 2028.⁴⁸ OpenAI sells more to consumers and gives away a larger free tier, so its cost of revenue absorbs more heavy usage on flat-rate plans. Epoch AI’s reconstruction of OpenAI’s 2024 compute spending found roughly $5 billion on R&D and $1.8 billion on inference, a similar ratio to Anthropic’s.⁴⁹

Certain of Anthropic’s cloud suppliers are also its investors, customers and competitors (the filing’s own phrase on p. 140). Amazon, Microsoft and Alphabet all booked gains on their Anthropic stakes in the quarter to 30 June 2026, though Alphabet’s own filing does not name Anthropic as the source.⁵⁰

The cost of reaching a given benchmark score has fallen roughly 13 times a year for three years.⁵¹ Anthropic’s derived gross margin on serving went from minus 95% in 2024 to about 12% in the first half of 2025 to 85% (unaudited) in the first half of 2026, and DeepSeek’s inference figures point in a similar direction. Margins at the Chinese labs are harder to compare, because the accounting frameworks, revenue mixes and margin trends all differ, but R&D runs at multiples of revenue at the Chinese labs and at OpenAI, against 60% at Anthropic.^45,46 The filing warns that operating margins “may compress” as committed compute comes online, and because all compute runs on one shared pool I expect the gross margin to feel it too.^23,18,41 Revenue outpaced committed compute in the first half of 2026, but the filing says that won’t last.⁴¹

My sustaining margin sits somewhere between about 9% and 68%. The 85% gross margin overstates things, because it counts only the cost of serving and leaves out the training that stops models going stale.

Whether today’s token prices can last depends on where in that range the sustaining margin falls, and on how fast the cost of reaching a given benchmark score keeps falling.⁵¹ The filing gives the raw numbers but not the split between maintenance and growth training, which would make it precise. I’ll update this when Anthropic fills in the blanks.

Sources

Anthropic, PBC, Form S-1, filed with the SEC (the Securities and Exchange Commission, the US markets regulator) on 26 October 2026. Confidential draft submitted 1 June 2026, per the EDGAR filing list. The company has applied to list on the Nasdaq Global Select Market under the ticker ANT. The filing covers FY2023 through H1 2026 and is at sec.gov.
S-1, p. 22 (summary financial data). Net loss $80,748 million on revenue of $16,286 million (minus 496%, our arithmetic). Operating loss $693 million (minus 4.3% of revenue, our arithmetic).
Steve Kopack, “Anthropic reveals $80 billion loss, huge spending spree ahead of IPO,” NBC News, 26 October 2026. Press reports of a valuation in the range of $1.5 to $2 trillion appear in Reuters, Yahoo Finance and others; no price range is in the filing.
Kali Hays, “Anthropic’s blockbuster share listing – 5 surprising things we now know,” BBC News, 26 October 2026. The article’s revenue figure is $16.2 billion; the filing gives $16,286 million. Hooper is described as co-head of equity capital markets at Mergermarket. Monaghan is described as a portfolio manager at Founder ETF.
S-1, p. 22 (summary financial data) and p. 156 (non-GAAP reconciliation). The fair-value change on convertible notes, derivative liabilities and SAFEs was $1,045 million in FY2023, $5,434 million in FY2024, $34,308 million in FY2025, and $72,388 million in H1 2026. Of the company’s $132.9 billion accumulated deficit, $113.2 billion relates to these fair-value changes (p. 161). By H1 2026 the charge was entirely from convertible notes, as no SAFEs or derivative liabilities remained outstanding (F-32).
S-1, F-12 (cash flow statement). Cash received for convertible notes: $1,750 million in 2023, $6,050 million in 2024, $3,450 million in 2025, totalling $11,250 million (our arithmetic). The lenders were Amazon Web Services, Inc. and Google LLC (F-39 to F-42).
S-1, p. 164 and F-40 to F-42. Amazon’s notes outstanding at 30 June 2026: $2,186 million of principal at a conversion price of $2.0928 per Class N share (1,044,716,970 shares) and $4,017 million at $5.0478 (795,727,992 shares), totalling $6,203 million of principal and 1,840,444,962 shares. These are post-split prices; Anthropic completed a 10-for-1 stock split on 17 September 2026 (F-14, F-65, p. 277). The pre-split conversion prices were $20.928 and $50.478 (F-40 to F-41). Google’s notes (under the 2023 and 2024 NPAs) had all converted by January 2026 (F-39 to F-40). The filing states that the conversion “will cause additional dilution to our existing stockholders” (p. 38).
S-1, F-46. Series H preferred stock was priced at $589.0095 per share, equivalent to about $58.90 per common share after the 10-for-1 split (our arithmetic: $589.0095 / 10 = $58.90). The $108.4 billion check: 1,840,444,962 shares at $58.90 gives $108.4 billion (our arithmetic), within 0.2% of the $108,566 million carrying value of convertible notes payable on the balance sheet at 30 June 2026 (F-5).
S-1, p. 175. Anthropic “elected the fair-value option under ASC 825” for the convertible notes and SAFEs, “to eliminate the accounting complexity that would otherwise result from separately accounting for the liability and equity components.” The fair-value measurements are Level 3 (no market price; valuation requires significant judgment or estimation) and use a hybrid of contingent claims analysis and a common stock equivalent method (pp. 174-175, F-31). No valuation assumptions (volatility, probabilities, weights) are disclosed (F-32). Ernst & Young flagged the measurement as a critical audit matter (F-3).
S-1, F-5. Convertible notes payable at fair value: $14,116 million at 31 December 2024, $41,677 million at 31 December 2025, $108,566 million at 30 June 2026. Principal outstanding at 30 June 2026: $6,203 million (p. 164). The difference, $102,363 million, is the accumulated fair-value adjustment above principal (our arithmetic).
S-1, pp. 58-59, 164 and 175. Amazon’s ownership cap is 12% of Anthropic’s capitalisation (counting all shares as converted plus equity awards and convertibles), unless Amazon waives it. Notes exceeding the cap are reissued as non-interest-bearing rollover notes with a 10-year maturity and continue to be revalued at fair value (p. 175). How much will convert at the IPO and how much will roll over “will be determined based on the final offering size and Amazon’s election with respect to the ownership cap” (p. 164). The filing warns that it “will continue to recognize changes in the fair value of our convertible notes until they are converted into equity, repaid, or otherwise settled” (pp. 146-147).
S-1, pp. 19, 23, 126-131. The pro forma balance sheet, dilution tables, per-share figures and the principal amount converting at the IPO are all left blank in this filing.
S-1, pp. 163-168 and F-31, F-37. Three 2026 note purchase agreements: Google ($20 billion capacity), Amazon ($15 billion) and Broadcom ($42 billion). None drawn at 30 June 2026. The charge on the NPA line itself was $9,405 million (F-37), arising because Google and Amazon bought Series G-2 preferred shares at $259.1364 under commitments made in March and April, before the Series H round priced at $589.0095 in May-June (F-31, F-46). The net “other income (expense)” line is $8,757 million after smaller offsetting items including warrant rights (+$724 million) and a holdback liability (F-37). The filing contradicts itself in places on the 2026 note terms: Note 9 gives interest rates (Google 9%, Amazon 8%) and maturities while the MD&A tables show interest “N/A” and terms “to be mutually agreed” (F-40, F-42 vs pp. 165, 167).
S-1, p. 156 (non-GAAP reconciliation). Adjusted net income of $2,671 million for H1 2026, excluding stock-based compensation ($2,185 million), non-cash charitable contributions ($126 million), the fair-value change ($72,388 million) and other mark-to-market items ($8,720 million). Includes $1,111 million of interest income on the company’s cash and investments.
S-1, p. 157 (Q2 2026 GAAP operating income of $974 million); F-12 and p. 168 (operating cash flow of $3,929 million for H1 2026).
S-1, pp. 157-160 (quarterly results of operations, all unaudited). Q2 2025 to Q2 2026: revenue $787 million to $11,555 million (+$10,768 million); compute and infrastructure $1,390 million to $5,558 million (+$4,168 million); personnel $320 million to $1,857 million (+$1,537 million); sales, marketing and partnerships $96 million to $2,135 million (+$2,039 million); other operating and admin $181 million to $1,031 million (+$850 million); operating result minus $1,200 million to plus $974 million. Revenue less all compute first turned positive in Q1 2026 ($4,731 million minus $3,387 million = $1,344 million).
S-1, Note 8 (Supplemental Operating Expense Information), F-39. Note 8 gives operating expenses by function: cost of revenue, R&D, sales and marketing, and general and administrative. The implied gross margins are derived: (revenue minus cost of revenue) divided by revenue. Cost of revenue was $752 million on revenue of $386 million in FY2024 (minus 95%); $2,773 million on $4,585 million in FY2025 (about 40%); $2,380 million on $16,286 million in H1 2026 (about 85%). These are our calculations; Anthropic does not report gross margin as a metric. The FY2023, FY2024 and FY2025 figures are audited (F-4); the H1 2025 and H1 2026 figures are unaudited (F-39 header).
S-1, F-39, Note 8 footnotes (cost-of-revenue definition).
The Information, January 2026, as summarised by Tiger Brokers and other outlets: Anthropic had cut its 2025 gross-margin forecast to about 40% from 50%, with 2024 at minus 94%. The Information’s reporting is paywalled; these secondary summaries match the filed figures closely.
S-1, p. 56 and F-16. About 48% of H1 2026 revenue came through cloud marketplaces (the filing names Amazon Bedrock, Gemini Enterprise Agent Platform and Microsoft Foundry on p. 206), up from 11% in 2023.
S-1, F-21. Anthropic is the principal and recognises revenue on a gross basis. The marketplace fee is booked in sales, marketing and partnerships, not in cost of revenue.
Austin Johnsen, “Anthropic S-1 Review,” Artificial Diligence, 27 October 2026; Longbridge Dolphin Research, “Anthropic Sheds the Cash Guzzler,” 27 October 2026. Channel fees increased by about $2.7 billion between H1 2025 and H1 2026 (S-1, p. 149). Including them as a cost of serving gives a margin of about 68% (derived; the exact fee total is not disclosed). Cape Fear Advisors, “Anthropic S-1: Gross Revenue and a Q3 Stated as 86 Times,” 26 October 2026, also discusses the gross-booking question.
S-1, p. 145: the company manages compute “on a total basis rather than by individual workload.” Total compute and infrastructure was $8,945 million in H1 2026; cost of revenue was $2,380 million (Note 8), so at most about 27% of compute expense went to serving paying customers. The remainder includes R&D compute, research and internal-use workloads, and free-tier compute, which sits in sales and marketing (F-39, footnote 3).
S-1, p. 143.
Evan Armstrong, “Is Anthropic Worth $2 Trillion?,” The Leverage, 26 October 2026. His margin of about 45% treats all compute and infrastructure ($8,945 million) as cost of goods against $16,286 million of H1 2026 revenue.
OnlyCFO, “Anthropic’s S-1: Finance Takeaways,” 26 October 2026. The illustrative figure: “With 50% of R&D included in COGS, H1 2026 gross margin would be about 55.5%.” The counterpoint: the right question is how much ongoing R&D is needed to sustain current revenue, not where to draw the line in one period.
Our arithmetic. Revenue $16,286 million minus cost of revenue ($2,380 million) minus R&D ($9,741 million) minus estimated marketplace fees (about $2,700 million, the increase disclosed at p. 149, as the low end of the fee range) leaves about $1,465 million, roughly 9% of revenue. With fees at the $2,820 million ceiling the figure is about 8%. R&D includes researcher salaries and overhead as well as compute, so this overstates the charge from training; it also represents the extreme case where all of R&D is maintenance.
S-1, p. 2-3 (defined terms) and p. 10. NRR of approximately 900% as of 30 September 2026. The definition uses a $1 revenue threshold in the base quarter and excludes Claude Pro and Max subscribers. Q3 2025 revenue was $1,382 million (p. 157).
Snowflake, Inc., Form S-1, filed August 2020 (NRR of 158% at 31 July 2020). Snowflake later reported 178% in its FY2022 10-K (at 31 January 2022), about 16 months after the IPO.
S-1, p. 138.
S-1, pp. 8-9, 56, 142-143, 148, 194, 206. Customers with $100,000 or more of trailing-twelve-month revenue: approximately 20 (end-2023), approximately 200 (end-2024), approximately 1,500 (end-2025), more than 11,900 (30 September 2026). Revenue from these customers was approximately 74% of H1 2026 revenue (p. 148). Approximately 250 customers had $10 million or more (p. 9). Approximately 40 million customers each had under $100,000 (p. 142). The top five represented approximately 25% of nine-month revenue; none exceeded 10% individually (p. 56). Revenue from outside the top five was approximately 64% in Q1 and 79% in Q3 (p. 143); the top-five shares (roughly 36% and 21%) are our arithmetic. Revenue from each of the top five reached a record high in Q3 (p. 143). The 12.5 times figure is our arithmetic (more than 11,900 over more than 950; both are floor figures).
S-1, F-15. In 2025, two unnamed customers (“Customer A” and “Customer B”) each represented 12% of revenue.
S-1, p. 56. “Many of our top customers are not subject to long-term contracts with us or have contracts with near-term expiration dates and are able to reduce or in some cases cease spending at any time and for any reason.”
S-1, p. 138. Consumption-based revenue was 80% of H1 2026 revenue ($13.1 billion of $16.3 billion).
S-1, F-58 (purchase commitments) and F-35 (leases). Fixed or minimum purchase commitments of about $265.6 billion at 30 June 2026 (terms of 1 to 20 years). Not-yet-commenced leases of about $321.4 billion in undiscounted payments at 30 June 2026 (partly conditional; from $50.2 billion at 31 December 2025).
S-1, F-65. Agreements entered into or amended after 30 June 2026 carry minimum commitments of about $245 billion, plus about $21 billion of short-term commitments.
S-1, p. 170-171. The named agreements: Microsoft, at least $44.9 billion from November 2026 to May 2033 (about six and a half years); Google, at least about $114.5 billion from April 2026 to July 2033 (about seven years); Amazon, at least $110 billion from May 2026 to April 2036 (ten years). The broader commitment and lease categories in F-58 and F-35 include terms of up to 20 years.
Tomasz Tunguz, “Anthropic S-1 Analysis: How 7 Key Metrics Stack Up,” 26 October 2026.
S-1, F-52. The $2.9 billion is the figure “if this offering had been completed on June 30, 2026”; the trigger is the effectiveness of the registration statement, so the actual charge depends on the date. Unrecognised cost of $23.7 billion over a weighted-average period of about 2.1 years. A further approximately 91 million RSUs worth $5.8 billion were granted after 30 June (F-64).
S-1, use-of-proceeds section. The company intends to use “a portion of the net proceeds to satisfy our anticipated tax withholding and remittance obligations related to the settlement of certain of our outstanding restricted stock units.”
S-1, p. 133-134. Committed compute coming online: operating margins “may compress in those periods.” Short-term compute “typically carries a higher cost per unit” and is expected to “compress our operating margin and Adjusted Operating Margin.”
S-1, p. 145.
S-1, p. 50.
S-1, p. 51. “We recently chose to delay pre-training a new frontier model for more than a month” to harden defences against model-weight exfiltration.
Zhipu (Z.ai Co., Ltd.), results announcement for the six months ended 30 June 2026, HKEX, 31 August 2026 (unaudited; reviewed by KPMG under HKSRE 2410). Revenue RMB 953,892 thousand, gross profit RMB 251,611 thousand (margin 26.4%), R&D RMB 2,131,171 thousand (223% of revenue, our arithmetic), loss for the period RMB 2,071,992 thousand. Cloud and API revenue was 86.5% of the total, up from 15.2%, with a gross margin of 24.6%. Annual results announcement for the year ended 31 December 2025, HKEX, 31 March 2026 (audited): revenue RMB 724,334 thousand, gross margin 41.0% (down from 56.3% in 2024), R&D 439% of revenue. The decline is attributed to “an increase in the proportion of cloud deployment business” (p. 9 of the 2025 annual results). The 2025 net loss of RMB 4,718,167 thousand includes RMB 937,433 thousand of fair-value changes on financial instruments issued to investors. Both reports are under IFRS; page references for the H1 2026 document are approximate.
MiniMax, results for the six months ended 30 June 2026 (company release; unaudited). Revenue $116,573 thousand, gross profit $20,813 thousand (margin 17.9%), R&D $296,870 thousand (255% of revenue, our arithmetic), net loss $357,997 thousand. Full-year 2025 results (company release): revenue $79,038 thousand, gross margin 25.4% (from 12.2% in 2024), R&D 320% of revenue. The 2025 net loss of $1,871,617 thousand includes a $1,589,850 thousand fair-value loss on financial liabilities (85% of the net loss, our arithmetic). Both reports are under IFRS.
DeepSeek, “Day 6: One More Thing, DeepSeek-V3/R1 Inference System Overview,” GitHub (Open Source Week), c. 1 March 2025. A 24-hour snapshot: 608 billion input tokens and 168 billion output tokens served at an estimated cost of $87,072 (H800 rental at $2 per GPU-hour), against theoretical revenue at published prices of $562,027. DeepSeek reported this as a “cost profit margin of 545%.” On a revenue basis that is equivalent to a gross margin of about 84% (our calculation). The disclosure covers GPU rental only, with no training, research, staff or overhead. DeepSeek has published no financial statements.
OpenAI’s 2024 and 2025 financial statements as reported by Ed Zitron and verified by the Financial Times (via 36Kr, 17 June 2026): 2025 revenue of $13.07 billion, cost of revenue $7.5 billion (implied gross margin about 43%, our arithmetic), operating loss $20.92 billion. The 33% adjusted gross-margin figure and the cash-burn projections are from The Information, as reported by The Decoder (21 February 2026 and 9 October 2026). OpenAI has not filed publicly; these are press-reported figures, not audited disclosures.
Epoch AI, “OpenAI’s compute spend,” 10 October 2025, based on press-reported investor documents. R&D compute of roughly $5 billion (including about $0.5 billion on final training runs of released models and about $4.5 billion on other research and experiments) and about $1.8 billion on inference. The inputs “should not be considered completely reliable,” Epoch AI notes.
S-1, p. 140 (“certain of” the cloud service providers are also investors, customers and competitors). Amazon reported $53.4 billion in Q2 2026 pre-tax other income “primarily” from Anthropic investments (Amazon Q2 2026 8-K exhibit). Microsoft booked a $3.2 billion Anthropic-related gain in its quarter to 30 June 2026 (FY26 Q4 earnings release). Alphabet reported a large equity-investment gain reportedly related to Anthropic (secondary reporting on Q2 2026 results; Alphabet’s own 10-Q does not name Anthropic).
Epoch AI (Emberson and Roodman), “The plunging price of thought,” 22 September 2026. The cost of reaching a given performance level on a benchmark fell about 47% per quarter (roughly 13 times a year) over three years, measured across five benchmarks. This is cost per question at a fixed score, not per-token price.