Anthropic's Economic Gross Margin Is Probably 50–60% — Not the 80%+ Headline
The 80%+ headline is not the margin I would underwrite
The Financial Times reported that Anthropic's gross margin had risen above 80%, but the same report said that figure was calculated before revenue sharing with distribution partners and before model-training costs. For valuation work, I think a normalized 50–60% economic gross-margin range is a more defensible starting point until public audited statements establish a stable accounting boundary.
What the Financial Times 80%+ number actually says
The Financial Times reported in September 2026 that Anthropic expected a second consecutive quarter of positive adjusted operating income and that gross margins were above 80%. That sounds like elite software economics. But the qualifier changes the interpretation: the margin was reported before revenue shared with distribution partners, including Amazon, and before the cost of training Anthropic's models. Reuters repeated the same qualification and also said it could not independently verify the private-company figures.
I do not think the useful criticism is that an 80%+ calculation is mathematically fake. The stronger criticism is that it is too flattering for peer comparison. If a meaningful distribution toll is removed before gross margin, then the number no longer answers the same economic question as the gross margin investors pull from an Adobe, Salesforce, ServiceNow, Microsoft or NVIDIA income statement.
That matters because Anthropic is being discussed at valuations where a few hundred basis points of sustainable margin are worth tens of billions of dollars. At that scale, accounting boundaries are not a footnote. They are part of the thesis.
Anthropic's own margin history points much closer to 50–60%
The cleanest reason I do not want to underwrite 80%+ is that Anthropic's reported margin history has been far more ordinary. The Information reported that Anthropic's overall gross margin was roughly 50–55% at the end of 2023. Later reporting put the company's 2025 gross-margin projection at 40%, about 10 percentage points below an earlier expectation because inference costs were running 23% above plan.
By mid-2025, The Information reported recent gross margins of about 60% across direct and partner sales. A separate report said direct sales were around 60% and moving toward 70%, while cloud-provider sales had at one point been deeply negative-margin. Most recently, the publication reported that Anthropic's optimistic 2026 forecast could reach as much as 63%, compared with 40% in 2025.
Anthropic gross-margin evidence over time
Anthropic historical and forecast gross margins
| Period / evidence | Margin | What it tells us |
|---|---|---|
| December 2023 | 50%–55% | reported overall gross margin |
| 2025 | 40% | later Anthropic projection reported by The Information; inference costs were reported 23% above plan |
| recent 2025 reporting | 60% | reported recent gross margin across direct and partner sales |
| 2026 optimistic forecast | 63% | company forecast reported by The Information |
Those observations do not prove the future margin will stay in the 50s. Anthropic is growing fast, model serving is becoming more efficient, utilization can improve, and custom silicon can lower inference cost. But they do make 50–60% a much more evidence-anchored underwriting range than simply capitalizing the newest 80%+ headline.
Revenue sharing is not a rounding error
Anthropic's distribution relationships are economically valuable. Amazon, Google and Microsoft can put Claude in front of enormous enterprise customer bases, and those channels may help Anthropic grow faster than it could on its own. The channel is not free.
The Information reported that Anthropic shares about 50% of gross profit with Amazon on Anthropic models Amazon resells. The same reporting estimated cloud-partner revenue-sharing payments of about $1.9 billion in 2026 and $6.4 billion in 2027.
Whether those payments sit inside cost of revenue, sales and marketing, or another line in a private management presentation is an accounting classification question. Whether they reduce the economics available to Anthropic shareholders is not. If a customer arrives through a reseller and a large slice of the resulting economics goes to the reseller, I want that burden reflected somewhere in the unit-economics model before I attach a premium software multiple.
Training compute belongs below gross profit — but not outside valuation
There is an important accounting point where I would not be maximally bearish. Model-training compute is usually closer to research and development than to the direct cost of serving the next customer request. The Information has noted that model makers generally classify training that way, and I do not think investors should mechanically stuff every training dollar into cost of revenue just to manufacture a lower gross-margin percentage.
But moving training below the gross-profit line does not make it economically disappear. The Information reported that Anthropic told some investors it would have been around break-even if training costs were included in cost of revenue. That is a useful reminder of the distance between a gross-margin headline and the cash economics required to keep a frontier-model company competitive.
My normalization therefore does not assume all training belongs in COGS. It does something more conservative and more practical: it refuses to treat an 80%+ pre-partner-share, pre-training metric as the economic equivalent of an 80%+ mature-software gross margin. Training remains an operating burden that the valuation eventually has to absorb through R&D efficiency and operating leverage.
The public-company comparison is still uncomfortable
To see why the definition matters, compare the normalized Anthropic range with recent total-company GAAP gross margins from established technology leaders. These are not perfect business-model comps: NVIDIA sells semiconductors, Microsoft mixes software and cloud infrastructure, and the software companies have different service mixes. That is exactly why I prefer an explicitly labeled comparison instead of pretending the figures are interchangeable.
Gross margin: Anthropic underwriting range vs. tech leaders
Gross-margin comparison between Anthropic underwriting range and public technology companies
| Company | Period | Gross margin | Basis |
|---|---|---|---|
| Adobe | FY2025 | 89.3% | GAAP total company |
| Intuit | FY2026 | 80.2% | GAAP total company |
| Salesforce | FY2026 | 77.7% | GAAP total company |
| ServiceNow | FY2025 | 77.5% | GAAP total company |
| NVIDIA | FY2026 | 71.1% | GAAP total company |
| Microsoft | FY2026 | 67.9% | GAAP total company |
| Anthropic | Underwriting range | 50–60% | Grizzly Bulls normalized economic range; not GAAP and not directly comparable |
Adobe's FY2025 gross margin was about 89.3%; Intuit's FY2026 margin was about 80.2%; Salesforce and ServiceNow were in the high 70s; NVIDIA reported 71.1%; and Microsoft's total-company FY2026 margin was about 67.9%. Those companies can support huge valuations partly because a large portion of each incremental revenue dollar survives the direct cost layer.
If Anthropic can genuinely converge toward 70–80% while maintaining explosive growth, the upside case becomes much easier to defend. If the durable economics settle closer to 50–60%, the company can still be extraordinary — but the valuation needs much more revenue growth and much more operating leverage to do the same work.
Why I land at 50–60%
- The range is anchored to the reported 50–55% overall margin at the end of 2023 and more recent reporting around 60%, rather than to one newly reported 80%+ management metric.
- It gives Anthropic credit for substantial inference-efficiency improvement from the weak 2025 margin period.
- It treats partner distribution economics as a real cost to shareholder economics even when the accounting presentation places those payments outside the headline gross-margin calculation.
- It does not force all model-training spend into cost of revenue; training remains an R&D-like operating burden below gross profit.
- It is deliberately a range. Anthropic is private, its business mix is changing quickly, and there is no audited public filing that lets outside investors reconcile one stable gross-margin definition across periods.
That is what I mean by normalized economic gross margin: not a substitute accounting standard, but the margin range I would use to stress-test the valuation until public disclosures give investors something better.
At a $965 billion valuation, the distinction is enormous
Reuters reported that Anthropic was valued at about $965 billion in May 2026 and that its annualized revenue run rate had climbed above $65 billion by mid-2026. Reuters also reported discussions around a possible IPO valuation of roughly $2 trillion. These are dated private-market and run-rate figures, not audited same-period financial statements, so the multiples below are illustrations rather than conventional public-company valuation ratios.
| Valuation case | Revenue multiple | Gross-profit multiple at 60% | Gross-profit multiple at 50% |
|---|---|---|---|
| $965B private valuation | 14.8× | 24.7× | 29.7× |
| $2.0T illustrative IPO case | 30.8× | 51.3× | 61.5× |
At the May private valuation, a 50–60% margin range turns roughly 14.8× revenue into about 24.7–29.7× normalized gross profit. At a $2 trillion valuation, the same exercise becomes roughly 51.3–61.5× normalized gross profit.
None of that proves Anthropic is overvalued. A company growing this quickly can outrun a very high starting multiple if the growth lasts and margins expand. It does mean the margin definition carries enormous valuation weight. An 80–90% software business and a 50–60% AI infrastructure-heavy business should not receive the same revenue multiple merely because both sell software through an API.
The Grizzly Bulls view
Anthropic may become one of the most important software companies in the world. Claude is growing at a remarkable pace, enterprise distribution is broadening, and inference efficiency can improve dramatically. I am not bearish on the product because I am skeptical of one margin headline.
I am bearish on the shortcut: taking an 80%+ figure that explicitly excludes partner revenue sharing and training, placing it beside mature software margins, and then using that comparison to make a near-trillion-dollar valuation look ordinary.
The useful question is not whether Anthropic can print an 80%+ margin under a particular management definition. It is how much revenue remains after recurring inference and distribution economics, how much recurring model-development spending is required to defend the franchise, and how much of today's spectacular growth ultimately converts into cash for shareholders.
Until audited public statements establish a stable accounting boundary, 50–60% is the range I would underwrite. At $965 billion — and especially in a $2 trillion IPO scenario — that distinction is not cosmetic. It is the valuation.
Methodology and limitations
- Separate reported Anthropic margin figures by scope instead of assuming every press report uses one stable gross-margin definition.
- Treat the Financial Times / Reuters 80%+ figure as reported but preserve its explicit exclusions for partner revenue sharing and model training.
- Use 50–60% as a Grizzly Bulls underwriting normalization anchored to historical reported margins, current partner economics and recent company forecasts. It is not a claimed current audited GAAP result.
- Keep model training conceptually below gross profit as an R&D-like operating cost while still treating it as economically relevant to valuation.
- Calculate public-company comparison margins from audited full-year revenue and cost-of-revenue figures where practical; NVIDIA's 71.1% figure is directly reported in its FY2026 filing.
- Use the rounded $65 billion annualized revenue run rate, $965 billion May 2026 private valuation and $2 trillion possible IPO case only as dated valuation illustrations.
Anthropic is private. Much of the financial evidence comes from press reports citing people familiar with company disclosures rather than audited public filings. Reuters explicitly said it could not independently verify the September 2026 profitability figures. The Information sources may be paywalled. Peer companies differ materially in business mix and reporting periods. This study should therefore be read as a valuation framework built from the best available public evidence, not as a restatement of Anthropic's financial accounts.
Sources
Download the derived figures
The public files contain only the derived underwriting metrics and peer-margin references used in this article. They do not redistribute private-company source documents.
Research data
- Download CSVCSVCSV downloadTabular public study data for spreadsheet analysis, independent checks, and new charts.Data snapshot September 16, 2026 · Reuse with attribution to the canonical study.
- Download JSONJSONJSON downloadStructured public study data for programmatic verification while preserving the published field names and research context.Data snapshot September 16, 2026 · Reuse with attribution to the canonical study.
Citation and reuse
When citing the 50–60% range, describe it as a Grizzly Bulls normalized economic underwriting range, not as Anthropic's reported GAAP gross margin.