Safeguard
Industry Analysis•

Anthropic's revenue, losses and margins: reported versus guessed

Anthropic's reported revenue, run rate, margins, losses and cash burn in one place, with each figure marked as reported, projected or estimated.

Safeguard Security Team
16 min read

Anthropic has no audited public accounts yet, so every figure about its revenue and profit comes from reports on investor briefings and a draft prospectus. Revenue has been reported as growing from about $4.6 billion in 2025 to $16.8 billion in the third quarter of 2026 alone, while margin figures range widely depending on how they are defined.

This guide lays the numbers out in order, says where each one comes from, and flags the ones that are projections or outside estimates. The main reports are from Reuters (on a draft prospectus), Bloomberg (on investor documents, republished by Fortune) and the Financial Times (on an investor meeting), plus a financial tracking page last updated on 15 October 2026 that compiles them. Anthropic has not published these results itself, so read this as a map of the reporting, not as a financial statement. Where Anthropic's own announcements give a figure, that is stated.

Key numbers at a glance

Dates are the date of the report or statement, not the period measured. All private company figures are unaudited.

  • 2025 full year revenue: nearly $4.6 billion. Reuters, from a draft S-1, reported September 2026.
  • Run rate, 12 February 2026: $14 billion. Company statement, announcing a $30 billion Series G at a $380 billion post-money valuation.
  • Q1 2026 revenue: $4.73 billion. Bloomberg, republished by Fortune, 15 August 2026, preliminary.
  • Run rate, 6 April 2026: above $30 billion. Company statement, up from about $9 billion at the end of 2025.
  • Run rate, May 2026: above $47 billion. Bloomberg, republished by Fortune, 15 August 2026, and a cloud computing news site, 1 October 2026.
  • Q2 2026 revenue: more than $11.5 billion. Bloomberg, republished by Fortune, 15 August 2026, preliminary, against $787 million a year earlier. Reuters later put it at about $11.6 billion.
  • Run rate, late July 2026: $65 billion. Attributed to Epoch AI.
  • Run rate above $100 billion pace, 18 September 2026. New York Times, via Axios, as listed on the tracking page.
  • Q3 2026 revenue: $16.8 billion, up about 45 percent on Q2. Financial Times, 14 October 2026, relayed by Reuters, shared at a closed investor meeting.
  • Q3 2026 adjusted operating margin: 11 percent. Financial Times, 14 October 2026, non-GAAP.
  • Gross margin before partner revenue share and training: above 80 percent. Financial Times, September 2026.
  • 2025 net loss: $42 billion, of which about $34 billion non-cash. Reuters, 28 September 2026, from the draft S-1.
  • Cash and short term investments, end of 2025: $20.28 billion. New York Post, on the filing.
  • Reported infrastructure commitments: at least $518 billion over roughly a decade. Reuters, relayed 1 October 2026.

Revenue: a very steep line

Reported figures, with the original outlet the tracking page cites:

  • 2025, full year: nearly $4.6 billion (Reuters, from a draft S-1). Reuters is reported to describe this as about twelve times the prior year.
  • Q2 2025: $787 million (Bloomberg, via Fortune).
  • Q1 2026: $4.73 billion (Bloomberg, via Fortune).
  • Q2 2026: more than $11.5 billion, marked preliminary (Bloomberg, via Fortune). Bloomberg's report said deliberations were ongoing and the figures could be revised. Reuters later gave about $11.6 billion.
  • Q3 2026: $16.8 billion, up about 45 percent on the second quarter using the Reuters figure (Financial Times, relayed by Reuters and The Next Web on 15 October 2026).
  • Run rate, late July 2026: $65 billion (attributed to Epoch AI).

Two items from the same page are not reported figures. An annualised third-quarter run rate of about $74 billion is an analyst estimate, attributed to SemiAnalysis via the Financial Times. A year-end 2026 figure of $110 billion is an investor projection reported by the Wall Street Journal.

Run rate deserves a note. It takes a recent period, usually a month or a quarter, and multiplies it up to a year. It is a useful signal in a fast-growing business, but it is not revenue earned. Anthropic itself said in April 2026 that its run rate had passed $30 billion, up from about $9 billion at the end of 2025.

The gap between run rate and booked revenue is easy to see here. Multiplying the first quarter of 2026 ($4.73 billion) by four gives about $19 billion a year. Yet the company said the run rate had passed $30 billion in early April. Both can be true, because a run rate usually reflects the latest month, and in a business growing this fast the last month is far bigger than the quarter average. The same logic explains why quarterly revenue of $16.8 billion in the third quarter (about $67 billion if multiplied by four) sits below the later, higher run rate figures. Neither is wrong. They measure different windows.

How the story built up: a timeline

  • Early 2026. The Series G announcement on 12 February 2026 says run rate was $14 billion, that revenue has grown more than ten times a year for each of the past three years, and that the company had been earning revenue for less than three years.
  • 6 April 2026. The Google and Broadcom announcement cites run rate above $30 billion.
  • May 2026. The cloud computing news site reports a run rate above $47 billion.
  • 1 June 2026. A confidential IPO filing, according to the tracking page.
  • Mid August 2026. Bloomberg, in a story republished by Fortune on 15 August, reports Q1 and Q2 revenue figures from investor documents, and that adjusted operating income for Q2 was positive though no amount was given. Anthropic declined to comment. The tracking page notes that the Wall Street Journal had projected $559 million in May.
  • Late September 2026. Reuters reports on 28 September on the draft S-1: the 2025 revenue, net loss, operating loss and commitments. The Financial Times reports the narrow gross margin above 80 percent.
  • 1 October 2026. Bloomberg reports that some prospective investors see fair value at $1.8 to $2 trillion, and that marketing could start as early as the week of 9 November.
  • 14 October 2026. The Financial Times reports Q3 revenue, an adjusted profit and positive cash flow for the quarter, and the 11 percent adjusted operating margin. The Wall Street Journal reports investors' projection of $110 billion by year end.
  • 15 October 2026. Reuters reports that some attendees discussed a valuation nearer $1.5 trillion.

What is driving the revenue

The company's own figures give some colour. At the Series G, Anthropic said customers spending over $100,000 a year had grown seven times in the past year, that customers spending over $1 million annualised had gone from about 12 two years earlier to over 500, and that eight of the Fortune 10 were customers. It said its coding product, Claude Code, had a run rate above $2.5 billion, more than doubled since the start of 2026, and that enterprise was over half of revenue. By April the number of customers above $1 million annualised was over 1,000.

The tracking page estimates the mix at roughly 80 to 85 percent enterprise and API, which is the page's own estimate and not a company figure. The draft prospectus reportedly shows concentration: nearly a quarter of 2025 revenue came from two customers, and 47 percent of sales were routed through Amazon and Google cloud platforms. That matters because a small number of decisions can move quarterly revenue a long way.

Margins: three different numbers for three different questions

The same page lists several margin figures, and they do not contradict each other as much as they appear to. They answer different questions.

  • Gross margin before partner revenue share and training: above 80 percent, reported by the Financial Times in September 2026. This looks only at the cost of serving requests and leaves out two big items.
  • Gross margin estimated by analysts for 2026: about 44 to 60 percent. This is an outside estimate, and the definition is not given. The tracking page also lists an estimated negative 94 percent for 2024, which shows how fast the picture has been described as changing.
  • Compute cost per revenue dollar: $0.71 in the first quarter of 2026, with $0.56 projected for the second quarter. Subtracting those from a dollar gives a rough compute margin of 29 percent and 44 percent. That is simple arithmetic, not a reported margin.
  • Revenue per dollar of compute: $1.70 in the third quarter, as summarised by the tracking page from Financial Times reporting. That is the same idea inverted: for every dollar spent on compute, $1.70 of revenue.
  • Adjusted operating margin, Q3 2026: 11 percent, reported by the Financial Times. This is non-GAAP, which means it adjusts out items that standard accounting rules would include.
  • GAAP operating result: not disclosed for the second or third quarter.
  • A late 2025 target: a 77 percent gross margin, reported by TechCrunch as part of the breakeven plan.

The practical lesson is that "gross margin" is not one number across companies. Whether it includes the revenue shared with cloud partners, the cost of free users, or training, changes the answer by tens of percentage points. The guide to how to read an AI lab's finances walks through the definitions, and the post on OpenAI's reported gross margin shows the same problem at a peer.

A plain way to hold these together: the narrow margin says serving a request is very profitable once the model exists. The wider estimates say that after you add partner shares and the very large cost of building the next model, much less is left. Both are statements about the same business at different cut lines.

Losses: the large accounting charge

Reuters, reporting on 28 September 2026 on the draft prospectus, put the 2025 net loss at about $42 billion, of which about $34 billion was an accounting charge tied to earlier financing that could convert into Anthropic shares. As the company's valuation climbed, those instruments became more valuable, and the rules require that rise to be booked as a loss. The same report puts the 2025 operating loss above $8 billion, a figure that excludes those write-downs. Coverage of the same report gives total 2025 operating expenses of roughly $12.7 billion, of which compute and infrastructure made up $7.33 billion. A non-cash charge does not mean money left the building that year. It means an accounting rule required a large write-down or expense to be recorded.

Two cash facts are reported by the New York Post, as cited on the tracking page: compute and infrastructure spend of $7.33 billion in 2025, about three times the 2024 figure, and cash of $20.28 billion at the end of 2025. The page says no cumulative loss total has been published since 2021, and its own "above $42 billion" figure is described as an inference. An earlier estimate of $10 to $15 billion of cumulative losses on that page has been corrected and should not be used.

Cash burn and the path to break even

The burn figures are projections from a late 2025 plan reported by the Wall Street Journal and TechCrunch, so they predate the revenue jumps above.

  • 2026 burn was planned at about 33 percent of revenue.
  • 2027 burn was planned at about 9 percent of revenue.
  • The company was reported to be aiming for positive cash flow in 2028, with up to $70 billion of revenue and $17 billion of cash flow.

Revenue has since run well ahead of that plan, according to the reports above, so treat these as the original intent, not the current forecast. Payments to Amazon, Google and Microsoft are reported at at least $80 billion through 2029 (The Information), a figure that the larger $518 billion commitment reporting has since overtaken in scale. For a peer comparison, the tracking page lists OpenAI's planned burn at about 57 percent of revenue for 2026 and 2027 and a breakeven target of 2029 to 2030, which are late 2025 plans. See the post on OpenAI's reported revenue, losses and cash burn for the detail and the corrections to that company's run rate.

The compute bill behind the margins

Margins cannot be read without the compute side. Anthropic's reported commitments are at least $518 billion over roughly a decade, about 80 percent non-cancellable (Reuters, relayed 1 October 2026). Because most compute is contracted ahead of time, costs do not fall when growth slows. That is why analysts focus on compute cost per revenue dollar. The company's own April 2026 posts describe commitments to Google and Broadcom and to Amazon, and the guide to Anthropic's compute commitments sets out the provider by provider breakdown, the chips and the timeline.

The tracking page also cites an estimated peak training cost of about $30 billion (SaaStr, a comparison based figure). Treat that as a rough outside estimate, since its basis has not been published.

Valuation and listing reports

Valuation numbers swing widely. The company's own figure from February 2026 is a $380 billion post-money valuation in the $30 billion Series G. The tracking page lists a later Series H valuation of $965 billion without naming a source, and says it is unverified. Press reports around a possible listing give a target of about $2 trillion and a raise of up to $100 billion (Wall Street Journal, 18 September 2026), a range of $1.8 to $2 trillion that some prospective investors see as fair value (Bloomberg, as reported by The Next Web on 15 October 2026) and a discussion nearer $1.5 trillion after the third quarter (Reuters, 15 October 2026). Nothing here is a price you can trade at. They are reported targets and conversations.

On timing, Bloomberg reported that formal marketing could start as early as the week of 9 November 2026 with trading targeted before 26 November, and added that the plans remain fluid. The Financial Times has named Nasdaq as the venue, according to the tracking page. The Wall Street Journal had reported a November target that was pushed back from October. Such schedules move often. This is not investment advice.

What is firm and what is uncertain

  • Firm, as company statements: run rate of $14 billion in February 2026, above $30 billion in April 2026, and the Series G terms. These are the company's own words, though a run rate is not audited revenue.
  • Reasonably firm, as reported from a prospectus: 2025 revenue of nearly $4.6 billion, the $42 billion net loss with its $34 billion non-cash part, and the $518 billion commitments. No public filing was available as of 15 October 2026, so these rest on press accounts of a draft.
  • Preliminary: Q1, Q2 and Q3 2026 revenue, reported from investor communications and leaks. Even the Q2 figure differs slightly between outlets ($11.5 billion and $11.6 billion).
  • Estimated by outsiders: gross margin of 44 to 60 percent, the $74 billion annualised Q3 figure, the 80 to 85 percent enterprise mix.
  • Projected: the $110 billion year-end figure and the late 2025 burn and breakeven plan.
  • Not disclosed: GAAP operating results for any 2026 quarter, and cumulative losses.
  • Internally inconsistent in the source: the tracking page headline refers to a $2 trillion listing while its latest body figure is $1.5 trillion. The dated reports are used here instead.

What could change these numbers

  • A published prospectus. Audited figures would replace every second-hand number in this post.
  • A change in definitions. Gross margin and operating margin can be restated as the company chooses its reporting basis.
  • Revenue concentration. A small number of very large customers can move quarterly revenue sharply.
  • Compute cost. Because most compute is contracted in advance, a slowdown in demand does not reduce the bill quickly.

What to watch next

  • The public S-1. As of 15 October 2026 the tracking page says none had appeared on the SEC filing system. When it does, compare its definitions of gross margin and operating profit with the press versions above.
  • Whether GAAP operating results appear, and how the $34 billion non-cash item is described.
  • The fourth quarter. Does the year-end figure approach the $110 billion projection?
  • Concentration disclosure. The two large customers and the 47 percent channel share may be described in more detail.
  • Valuation talk. The move from about $2 trillion to about $1.5 trillion shows how quickly it changes.
  • Peer filings. If OpenAI or others publish audited numbers, comparisons become fairer. The guide to AI lab headcount and revenue per employee shows another way these figures get compared.

What it means if you buy or build on Claude

For a customer, the useful reading is not the valuation. It is whether the vendor is investing heavily and growing quickly, which supports continuity, and whether it is carrying large fixed costs, which can pressure pricing. For a practical view of what AI libraries and services your own code depends on, Safeguard offers a free scan at https://gold.safeguard.sh/scan.

Quick check

Is run rate the same as revenue?

No. Run rate multiplies a recent period out to a year. Revenue is what has actually been earned in a period.

Why do reported gross margins for the same company differ so much?

Because they define cost differently. One figure leaves out partner revenue share and training. Another includes more of them. The numbers answer different questions.

Does a $42 billion net loss mean $42 billion of cash was spent?

No. About $34 billion of the reported figure was a non-cash accounting charge.

Sources

Last reviewed: 19 October 2026.

Never miss an update

Weekly insights on software supply chain security, delivered to your inbox.

Self-healing security runs on Safeguard.

Your first fix PR is minutes away.

No sales call required, even your agent can complete the purchase over MCP.