Anthropic Margins, Compute Spend, and Losses: S-1 Filing
Full research prompt
Document Anthropic's gross margin, inference and training compute spend, cloud commitments (AWS, Google TPU, Microsoft/Nvidia deals), operating loss, and net loss. Itemize the non-cash and accounting charges inside the net loss, such as stock-based compensation, fair-value changes on convertible instruments or warrants, legal settlements (e.g., the copyright settlement), and impairments. Separate operating loss from these charges and note which numbers are publicly estimated versus filed.
From Anthropic Revenue and Burn Rate: What the IPO Filing Shows
Anthropic’s public S-1 (filed with the SEC on October 26, 2026) is the first audited view of a frontier lab’s books. The operating business is close to break-even and already cash-generative. The GAAP net loss is almost entirely a mark-to-market on financing instruments whose value rose with Anthropic’s own valuation. The binding constraint is not current cash burn; it is multi-year take-or-pay compute that will hit the income statement only as capacity is delivered.
1. Gross margin: an 85% functional margin that is not an inference margin
The S-1 does not publish a “gross margin.” What it does publish, in the functional note, is cost of revenue of $2,380 million on H1 2026 revenue of $16,286 million — about 85% — versus 12% in H1 2025 ($1,019 million of cost of revenue on $1,163 million of revenue) and about 40% for full-year 2025 ($2,773 million on $4,585 million). Full-year 2024 was negative, about −95%.
That 85% is an accounting allocation, not a pure inference margin. The filing says “substantially all” of cost of revenue is compute and infrastructure for customer-serving workloads, plus some personnel and overhead. Training, research, internal Claude use, and free-tier serving sit elsewhere. Total compute and infrastructure in the same half was $8,945 million, so at least about $6.6 billion of the compute bill — roughly three-quarters — is outside cost of revenue. Research and development was $9,741 million in the half and “primarily” consists of compute for model development, research, and internal usage. Sales and marketing of $3,808 million includes compute for free access.
Two further distortions matter for anyone comparing labs:
- Marketplace revenue is booked gross. About 48% of H1 2026 revenue came through AWS, Google Cloud, and Azure. Channel-partner fees sit in sales and marketing, not as a contra-revenue or cost-of-revenue item. Secondary readings of the filing put those fees at roughly 17% of H1 revenue (on the order of $2.7 billion of the S&M increase). A “gross margin” that ignores a 25–30% reseller cut on nearly half of sales overstates unit economics relative to a direct-sold software company.
- Consumption vs. subscription mix is filed: $13.1 billion (80%) usage-based, $3.2 billion (20%) subscriptions in H1 2026. Subscriptions are the less efficient slice of the inference pool.
Filed vs. estimated. The 85% / 40% / 12% figures are arithmetic on filed cost of revenue, not a company metric. SemiAnalysis estimates, published before and around the filing, are a different construct: blended inference gross margin in the mid-60s by mid-2026 (from −94% in 2024 and about 38% in 2025), API above 80% and potentially as high as 88%, with subscriptions about 10% of revenue but more than 40% of inference compute and still roughly 50% compute margin under realistic utilization. Those are outside estimates, not S-1 lines. Post-filing, SemiAnalysis has also described the filed customer-serving margin as about 85% and incremental economics as roughly 90% gross / 40% EBIT — commentary on the filing, not a second set of audited numbers.
What this means for a competitor. You cannot underwrite Anthropic’s token economics off the 85% alone. The margin expands when the shared compute pool is pointed at paying inference and compresses when the same pool is pointed at the next training run. The filing itself warns that when large new capacity comes online, operating margin “may compress… even if the underlying economics of the business continue to improve.”
2. Inference and training compute spend: one pool, two expense lines
Anthropic runs AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs as one fungible pool and directs it among research, model development, customers, and internal use. Expense is recognized when capacity becomes available, not when the contract is signed, and most of it is contracted 12–18 months ahead. The company does not disclose a dollar split of compute between training and inference beyond the functional allocation above.
Filed compute and infrastructure (nature presentation):
| Period | Revenue | Compute & infrastructure | Compute / revenue |
|---|---|---|---|
| FY2023 | $41M | $433M | 1,056% |
| FY2024 | $386M | $2,522M | 653% |
| FY2025 | $4,585M | $7,326M | 160% |
| H1 2025 | $1,163M | $2,334M | 201% |
| H1 2026 | $16,286M | $8,945M | 55% |
| Q1 2025 | $376M | $944M | 251% |
| Q2 2025 | $787M | $1,390M | 177% |
| Q3 2025 | $1,382M | $2,047M | 148% |
| Q4 2025 | $2,040M | $2,945M | 144% |
| Q1 2026 | $4,731M | $3,387M | 72% |
| Q2 2026 | $11,555M | $5,558M | 48% |
Revenue per compute dollar went from about $0.40 in Q1 2025 to about $2.08 in Q2 2026. H1 2026 as a whole was about $1.82 of revenue per compute dollar, versus about $0.50 a year earlier and about $0.63 for all of 2025. Q2 2026 revenue grew 1,368% year over year while compute grew about 300%.
The best filed proxy for “inference compute for paying customers” is the cost-of-revenue ceiling of $2.4 billion in H1 2026. The best filed proxy for “training plus research plus internal plus free tier” is the residual of at least $6.6 billion, most of it inside the $9.7 billion R&D line. There is no filed number that isolates pre-training or post-training spend from serving research traffic.
Capex is small because the model is rented, not owned: $482 million of property and equipment purchases in H1 2026, against $8.9 billion of compute expense. Owned PP&E, net, was about $1.2 billion at June 30, 2026.
3. Cloud and chip commitments: floors, ceilings, and estimates that do not add cleanly
These are contractual obligations, mostly not yet in the income statement. The S-1 uses several non-comparable buckets. Adding them double-counts.
Balance-sheet-date aggregates (filed):
- Fixed or minimum guaranteed commitments under compute purchase agreements: about $265.6 billion at June 30, 2026, terms of 1–20 years, up from $54.6 billion at December 31, 2025 (the year-end figure counted only commitments with more than a year left, so the jump is not fully like-for-like).
- Leases signed but not yet commenced, primarily compute and infrastructure: about $321.4 billion of future undiscounted obligations at June 30, 2026, up from about $50.2 billion at December 31, 2025. Some are subject to contractual conditions. Commencement 2026–2028; terms from under a year to 20 years.
- Since year-end 2025, new and amended lease and commercial agreements with aggregate future minimum payment obligations of about $417 billion, commencing 2026–2027.
- After June 30, 2026, further agreements of about $245 billion in minimum payments, some conditional.
- Additional TPU capacity for delivery 2028–2031, commercial terms not finalized: company estimate of $125–150 billion, excluded from the commitments table.
- Short-term non-cancelable minimums of about $21 billion are also disclosed in footnotes. One reading of the footnotes (Wired) puts the all-in “if we follow through” figure near $808 billion. That is a sum of overlapping and estimated categories, not a single contractual total.
Named counterparties (filed deal terms, as extracted by multiple readings of the same S-1):
| Counterparty | What | Amount | Firmness | Term |
|---|---|---|---|---|
| Broadcom | Google TPU hardware, 5-year master equipment leases (finance leases); Anthropic Direct | ~$161.2B rent, excluding interest | Non-cancellable except default | 5 years |
| Google Cloud | TPUs / cloud, true-up if spend falls short | At least ~$114.5B | Non-cancellable except Google breach | Apr 2026–Jul 2033 |
| Amazon / AWS | Trainium and other compute, true-up; up to 5 GW | At least $110B | Non-cancellable except Amazon breach | May 2026–Apr 2036 |
| SpaceXAI (SpaceX; formerly xAI) | NVIDIA GPU capacity, ~$3.1B/month at full delivery | Up to $120.5B | Cancellable on 90 days’ notice after Jan/Mar 2027 | Into 2029 |
| Microsoft Azure | NVIDIA GPUs | At least $44.9B | Non-cancellable except breach or undelivered tranches | Nov 2026–May 2033 |
| Nscale | NVIDIA GPUs, paid as delivered | Up to ~$44.6B | Limited cancellation | Into 2033 |
| Lambda | NVIDIA GPUs; a large tranche depends on a facility lease | ~$41.3B | Limited | Into 2034 |
| AMD | Instinct MI450 / Helios; AMD to buy up to $5B of Anthropic stock as chips deliver | Expected to exceed $20B | Non-cancellable except AMD breach | First tranche 2027 |
| Additional TPU (Google and others) | Capacity 2028–2031 | Est. $125–150B | Not in the commitments table | Estimate |
Microsoft’s floor was amended upward. The November 2025 announcement was about $30 billion; a confidential draft still showed about $31.4 billion; the public S-1 shows at least $44.9 billion after an October 2026 amendment. September 2026 press accounts of a confidential draft’s “$518 billion” total (Google ~$111 billion, Amazon $110 billion, Microsoft ~$31 billion, Broadcom $161 billion, plus other items, ~80% non-cancelable) are draft figures, superseded in part by the public filing. Do not treat $518 billion, $750 billion, or $808 billion as a single filed obligation.
Supplier financing is part of the structure. Broadcom agreed to lend up to $42 billion in convertible notes (9% paid-in-kind, convertible at a 10% discount) solely to fund Anthropic’s rent. Amazon holds the only outstanding convertible notes ($6.2 billion principal). Google, Amazon, and Microsoft are simultaneously investors, compute vendors, resellers, and (for Google and Amazon) model competitors.
Implication. The income statement will not show these dollars until machines are delivered. Once delivered, most of the Google, Amazon, Microsoft, and Broadcom paper is take-or-pay: Anthropic pays the shortfall if usage undershoots. SpaceXAI is the main pressure valve. A demand air-pocket does not shrink the Broadcom/Google/AWS bill; it turns unused capacity into a fixed operating cost on top of whatever training run is in flight.
4. Operating loss versus net loss: the charge bridge
All figures below are from the S-1 (FY audited; H1 unaudited), in millions of USD.
| FY2023 | FY2024 | FY2025 | H1 2025 | H1 2026 | |
|---|---|---|---|---|---|
| Revenue | 41 | 386 | 4,585 | 1,163 | 16,286 |
| Total operating expenses | 592 | 3,366 | 12,645 | 3,298 | 16,979 |
| Operating loss | (551) | (2,980) | (8,060) | (2,135) | (693) |
| Interest income | 69 | 106 | 431 | 115 | 1,111 |
| Change in FV of convertible notes, derivative liabilities, and SAFEs | (1,045) | (5,434) | (34,308) | (10,942) | (72,388) |
| Other income (expense), net | (2) | — | (8) | — | (8,757) |
| Net loss | (1,531) | (8,309) | (41,965) | (12,969) | (80,748) |
H1 2026 walk from operations to net loss (filed):
- Operating loss: $(693)
- Interest income: +$1,111
- Fair-value change on convertible notes, derivative liabilities, and SAFEs: $(72,388) — “primarily driven by an increase in the fair value of our outstanding convertible promissory notes resulting from an increase in our enterprise valuation.” Amazon rollover notes only: $6,203 million principal ($2,186 million under the 2023 NPA, $4,017 million under the 2024 NPA), carried at $108,566 million at June 30, 2026, up from $41,677 million at December 31, 2025. Conversion prices $2.0928 and $5.0478; about 1.84 billion Class N shares issuable, subject to a 12% ownership cap unless Amazon waives it. Fair-value option under the convertible accounting election; the charge stops when the notes convert at the IPO.
- Other expense, net $(8,757), itemized in the notes as: recognition and change in fair value of 2026 note purchase agreements (Google, March 2026; Amazon, April 2026) $(9,405); holdback liability $(39); change in fair value of warrant rights +$724; other $(37).
- Tax provision: $(21)
- Net loss: $(80,748)
Quarterly operating income turned positive in Q2 2026 at +$974 million, after a Q1 2026 operating loss of $(1,667) million. Operating cash flow was +$3,929 million in H1 2026 (versus $(1,513) million in H1 2025 and $(2,637) million in FY2025). Cash and short-term investments were $131.7 billion at June 30, 2026 ($112.4 billion cash, $19.3 billion short-term investments).
What sits inside operating loss (cash and non-cash)
These are operating items. They are already in the $(693) million and the $(8,060) million. They are not additional charges below the operating line.
| Item | FY2025 | H1 2025 | H1 2026 | Character |
|---|---|---|---|---|
| Stock-based compensation | $1,165M (footnote rounds the year to $1.2B; cash-flow add-back $1,158M) | $286M (cash-flow add-back $279M) | $2,185M | Non-cash; inside personnel. Personnel was $3,152M in H1 2026, so cash compensation was roughly $1B. Unrecognized SBC on unvested RSUs was about $26.6B at June 30, 2026 — a future operating charge, not a current one. |
| Non-cash charitable contributions | $541M | — | $126M | Non-cash; inside opex. |
| Bartz copyright settlement | $1,500M | — | — | Cash settlement, recorded entirely in FY2025 operating expenses. Not a fair-value or impairment charge. |
Bartz. Bartz et al. v. Anthropic settled for $1.5 billion, the largest known U.S. copyright class settlement. Final approval was July 20, 2026. The S-1 says $300 million was paid in 2025 and $1.2 billion remained in current liabilities at June 30, 2026. Installments run through September 2027. The company excludes it from adjusted operating income as not indicative of ongoing operations. Other copyright matters (music publishers, Canadian actions, Reddit, Bartz opt-outs) are disclosed with no loss estimate.
Impairments. The income-statement bridge and the non-GAAP reconciliation do not identify a material impairment charge in FY2025 or H1 2026. No filed impairment number belongs in this loss walk.
Company-defined adjustments (non-GAAP, filed reconciliation)
Adjusted operating income excludes SBC, non-cash charitable contributions, and the Bartz settlement. Adjusted net income also excludes the two fair-value lines.
| Add-back | FY2025 | H1 2025 | H1 2026 |
|---|---|---|---|
| Net loss | (41,965) | (12,969) | (80,748) |
| SBC | 1,165 | 286 | 2,185 |
| Non-cash charitable contributions | 541 | — | 126 |
| Bartz settlement | 1,500 | — | — |
| FV of notes, derivatives, SAFEs | 34,308 | 10,942 | 72,388 |
| FV of certain mark-to-market instruments | 3 | — | 8,720 |
| Adjusted net income (loss) | (4,448) | (1,741) | 2,671 |
| Adjusted operating income (loss) | (4,854) | (1,849) | 1,618 |
| Adjusted operating margin | (106)% | (159)% | 10% |
H1 2026 GAAP operating loss of $693 million plus $2,185 million of SBC plus $126 million of charitable contributions equals the $1,618 million adjusted operating profit. Interest income of $1,111 million then more than covers the GAAP operating loss before any fair-value add-back. The $2,671 million adjusted net profit is the company’s measure, not cash earnings and not GAAP.
5. Filed versus estimated — a clean split
Filed (S-1, October 26, 2026), or arithmetic strictly on filed lines:
- Revenue, compute and infrastructure, operating loss, net loss, interest income, and the fair-value lines in the table above.
- Cost of revenue and the functional split (R&D $9,741M, S&M $3,808M, G&A $1,050M in H1 2026); the ~85% “gross margin” is derived, and the S-1 does not label it as such.
- SBC, Bartz ($1.5B in FY2025 opex; $1.2B liability at June 30), non-cash charitable contributions.
- Other-expense components ($9,405M note-purchase marks, $724M warrant gain, $39M holdback).
- Amazon note principal $6.2B vs. carrying value $108.6B.
- Operating cash flow $3,929M; capex $482M; cash and short-term investments $131.7B.
- Purchase-commitment floor $265.6B and uncommenced-lease obligations $321.4B at June 30; the $417B / $245B / $125–150B subsequent and estimated buckets as the company states them.
- Named deal floors and caps (Broadcom ~$161.2B, Google at least ~$114.5B, Amazon at least $110B, Microsoft at least $44.9B, SpaceXAI up to $120.5B, Nscale up to ~$44.6B, Lambda ~$41.3B, AMD expected to exceed $20B).
- Q2 2026 operating income of $974M. Consumption/subscription mix and 48% marketplace share.
Estimated or third-party, not a filed line:
- SemiAnalysis inference margins (mid-60s blended, API >80%, up to 88%, subscription compute share >40%).
- Any single “total commitments” number — $518B (September draft, via Reuters), $542–567B, $637B (sum of seven named deals, Tunguz arithmetic), $657B, $750B, or ~$808B. The filing leaves the all-in total structured as separate buckets, some estimated, some conditional, some overlapping with the June 30 notes.
- The exact dollar of compute inside cost of revenue (at most $2.4B) and the exact training-versus-inference split (not disclosed).
- Q3 2026 revenue. The S-1 leaves the quarter blank and says revenue was 86 times the $197 million of Q4 2024, which implies on the order of $17 billion. Press reports of $16.8 billion for Q3, and of ~$33 billion for the first nine months, are outside the H1 financial statements.
6. How to read this if you are underwriting the business
The cash engine and the GAAP loss are different objects. In H1 2026 the company lost $693 million on operations, earned $1.1 billion of interest on a $132 billion cash pile, and generated $3.9 billion of operating cash flow, while booking $81 billion of non-cash valuation marks that more than explain the $80.7 billion net loss. Those marks are real dilution — Amazon’s $6.2 billion of principal is a claim on equity carried at $109 billion — but they are not compute spend and they end at conversion.
The economic risk is the delivery schedule. Compute was 48% of revenue in Q2 only because revenue outran capacity that had already been ordered. The next tranches (Microsoft from November 2026, SpaceXAI ramping toward ~$3.1 billion a month, Broadcom lease commencements, AWS and Google true-ups through the mid-2030s) are largely fixed once delivered. Margin will oscillate with when that capacity lands relative to token demand, not with whether the fair-value charge prints again.
A competitor or a public-market investor should underwrite three numbers separately: customer-serving compute (the ≤$2.4 billion H1 cost-of-revenue pool, before reseller fees), the training residual (most of the other ~$6.6 billion of H1 compute, inside R&D), and the undiscounted obligation stack ($265.6 billion of minimum purchase commitments plus $321.4 billion of uncommenced leases at June 30, plus what was signed after). Conflating the 85% functional margin with company-level profitability, or the $80.7 billion net loss with cash burn, is the misread the filing is built to invite.
Recent Findings Supplement (October 2026)
Anthropic’s October 26, 2026 S-1 filing (its first detailed public financial disclosure ahead of a planned Nasdaq IPO under ticker ANT) provides the primary new data on H1 2026 results (ended June 30, 2026), separating operational performance from large non-cash accounting items tied to rising valuation and prior financing instruments.[1][2]
This filing updates prior estimates with audited/unaudited figures, showing revenue growth outpacing compute costs enough for a Q2 operating profit while the GAAP net loss is dominated by fair-value adjustments.[3][4]
Revenue Growth and Operating Results
Anthropic reported $16.286 billion in revenue for H1 2026 (more than 14x the $1.163 billion in H1 2025 and exceeding all of 2025’s $4.585 billion), driven primarily by usage-based consumption (80%+ of revenue).[5][6]
- Operating loss narrowed sharply to $693 million (from $2.135 billion in H1 2025); Q2 alone flipped to $974 million operating income as revenue more than doubled QoQ to $11.555 billion while compute/infrastructure costs rose more modestly to $5.558 billion.[3][7]
- Adjusted operating income reached $1.618 billion (10% margin) after excluding stock-based compensation and other items; operating cash flow was positive $3.929 billion.[4][8]
For competitors or entrants: The data shows a viable path to operating leverage once scale is achieved, but highlights heavy reliance on usage-based revenue that can fluctuate seasonally or with customer budgets.[9]
Net Loss Breakdown and Non-Cash Charges
The GAAP net loss totaled $80.748 billion for H1 2026 (vs. $12.969 billion in H1 2025), driven overwhelmingly by non-operating, non-cash fair-value adjustments rather than cash outflows.[1][10]
- $72.388 billion from changes in fair value of convertible notes, derivative liabilities, and SAFEs (primarily reflecting higher enterprise valuation).[4]
- $8.720 billion (or ~$8.7 billion) from changes in fair value of other mark-to-market financial instruments.[4]
- Stock-based compensation of $2.185 billion (included in operating expenses; up significantly YoY).[4]
- Non-cash charitable contributions of $126 million.[4]
- The $1.5 billion Bartz et al. copyright settlement (literary works used in training) was recorded as a charge in 2025 ($300 million paid that year; $1.2 billion remaining current liability at June 30, 2026); no incremental charge appears in H1 2026 results.[11][12]
Company-adjusted net income (excluding the above valuation changes, SBC, etc.) was a $2.671 billion profit.[4] No impairments are detailed in the reporting.[4]
Implication: GAAP net loss figures are not indicative of cash burn or operational health; adjusted metrics and cash flow provide a clearer operational view, though valuation-driven charges will recur with further equity raises or valuation shifts.[2]
Compute and Infrastructure Spend
Compute and infrastructure expenses reached $8.945 billion in H1 2026 (nearly 4x the $2.334 billion in H1 2025), covering both training and inference.[5][10]
- Only $2.38 billion was classified in cost of revenue (“substantially all” of which relates to customer-serving workloads), implying a proxy gross margin (revenue less CoR) of ~85% (vs. ~12% in H1 2025); the balance (~$6.565 billion) appears in R&D and sales/marketing (e.g., training, research, free-tier access).[13][13]
- Overall compute spend fell to 48% of Q2 revenue (from much higher earlier percentages).[3]
For market entrants: High fixed/variable compute costs remain a barrier, but the accounting split and improving efficiency (revenue growth outpacing spend) suggest gross-margin-like metrics can appear strong when most training/free usage is expensed elsewhere.[13]
Cloud and Compute Commitments
The S-1 details substantial multi-year minimum commitments (largely non-cancellable except in limited cases), totaling hundreds of billions across providers; these support a multi-chip strategy (AWS Trainium, Google TPUs, Nvidia GPUs via partners, etc.).[14][15]
- AWS: At least $110 billion (through April 2036).[16]
- Google Cloud/TPUs: At least $114.5 billion (through July 2033); additional estimated $125–150 billion for further TPU capacity (2028–2031).[14]
- Microsoft Azure: At least $44.9 billion (through May 2033; recently amended upward from prior ~$30–31 billion figures).[14]
- Broadcom (TPU-related leases): ~$161.2 billion.[14]
- Other Nvidia-powered: Nscale ~$44.6 billion; Lambda ~$41.3 billion; SpaceXAI up to ~$120.5 billion (largely cancellable on 90 days’ notice after early 2027).[14]
Minimum non-cancellable commitments are commonly cited in the $265–430+ billion range across major providers, with broader totals (including estimates and softer deals) reaching $542–567 billion or higher.[17][18]
Implication: These locked-in obligations create significant supplier leverage (and revenue share, with ~48% of H1 revenue flowing through Google/Amazon/Microsoft channels) while representing a major portion of cloud providers’ backlogs; new entrants face similar scale requirements or must negotiate differently.[19]
Cash Position and Context
Cash and short-term investments stood at ~$131.7 billion as of June 30, 2026 (supported by $141.1 billion total capital raised historically).[3]
All figures above derive from the October 26, 2026 S-1 (public estimates prior to filing have been superseded by these filed numbers). No additional post-filing updates or separate research/policy items appear in recent coverage.[20]
These disclosures mark a shift from private estimates to standardized, comparable public data, enabling better benchmarking of AI lab economics.