To read an AI lab's numbers, ask five questions. Is this booked revenue or an annualized run rate? Is it gross or net of partner payments? How is gross margin defined? Is the loss accounting or cash? Is a commitment cancellable? The same company can look very different depending on the answers.
AI lab finances are covered like sports scores, with big numbers and little context. This guide is a reading manual, built on real examples from reporting on OpenAI and Anthropic. It explains each term the first time it appears and shows how the same fact can produce two honest but very different figures. It is not investment advice, and every figure is as of 19 October 2026 unless dated otherwise. Remember the basic limit: neither company has public audited accounts yet, so almost everything below is leaked, reported to investors, or a projection.
Key numbers at a glance
- OpenAI run rate, end of September 2026: about $50 billion (Financial Times and CNBC, 8 October 2026), after an earlier figure near $70 billion was corrected.
- OpenAI 2025 booked revenue: $13.07 billion versus an exit 2025 run rate of about $21.4 billion (leaked audited accounts and company disclosures, as relayed by a tracker).
- Anthropic 2025 booked revenue: nearly $4.6 billion (Reuters, citing a draft prospectus, reported 28 September 2026).
- Anthropic run rate, late July 2026: about $65 billion (Epoch AI data, as cited by a tracker).
- Anthropic Q3 2026 revenue: $16.8 billion (Financial Times, 14 October 2026), a booked quarterly figure.
- Accounting gap on partner sales: up to $8 billion of annualized revenue (a person familiar, via Semafor, 10 April 2026).
- Anthropic margin range for 2026: an estimated 44% to 60% depending on how training cost is treated (tracker estimate), and above 80% before partner revenue share and training (Financial Times, September 2026).
- OpenAI 2025 losses: operating loss $20.92 billion, net loss about $38.5 billion, adjusted loss about $8 billion (leaked accounts and a Financial Times source, via secondary coverage).
- Anthropic 2025 losses: net loss $42 billion including a roughly $34 billion non-cash charge; operating loss above $8 billion on an adjusted basis (Reuters, draft prospectus).
- Anthropic obligations: $518 billion of cloud, computing and infrastructure obligations (Reuters, draft prospectus).
- OpenAI compute plan: about $856 billion for 2026 to 2030, against about $600 billion told to investors in February (Financial Times and CNBC, via TheNextWeb).
- OpenAI 2026 to 2030 negative free cash flow plan: $278 billion, versus about $305 billion in May (Financial Times, via TheNextWeb).
- Customer concentration, Anthropic 2025: nearly a quarter of revenue from two customers, and 47% of sales routed through Amazon and Google (draft prospectus, via a tracker).
Question 1: is it revenue or a run rate?
Revenue is what a company has recognized in its accounts for a past period. A run rate is a forecast dressed as a fact: take a recent period, usually a month, and multiply to get a yearly pace. Channel Insider notes that startups like run rates because multiplying one strong month by twelve creates large figures, and warns that they can mask underlying performance.
Neither is wrong, and for a fast grower the run rate is the more current number. But they answer different questions. OpenAI's booked revenue for 2025 was $13.07 billion, while its run rate at the end of that year was above $20 billion. Both are true. The difference is the growth during the year. If you compare one company's run rate with another's booked revenue, you are comparing a speedometer with an odometer.
What to do: write down the period. "$30 billion annualized" means roughly a month at $2.5 billion. "$16.8 billion in the quarter" is a result. When growth is as fast as these labs report, with Anthropic's quarterly revenue reportedly up about 46% from Q2 to Q3, the two can be miles apart.
Question 2: is it gross or net?
This is the biggest trap in the current reporting. When a customer buys a lab's model through a cloud provider, the money can be recorded two ways. Gross means the lab books the full price as revenue and records the cloud provider's cut as a cost. Net means the lab books only its own share.
Reporting from The Information, relayed by Semafor in April 2026, says Anthropic books the full dollar while OpenAI books its share, with an example of 20 cents. Channel Insider says both approaches fit US accounting rules, so this is presentation, not wrongdoing. The size matters: Semafor's source said the difference could be up to $8 billion of annualized revenue, and in October, the Financial Times and CNBC reported OpenAI's run rate at about $50 billion after investors' attempts to put the two companies on the same footing had produced an earlier figure near $70 billion.
The rule behind the choice is the US revenue standard, ASC 606, and its "principal versus agent" test. Deloitte's guidance explains that a company is a principal, and reports revenue gross, if it controls the good or service before it is transferred to the customer. If it only arranges for another party to provide it, it is an agent and reports net. Control means the ability to direct the use of the service and obtain substantially all of its remaining benefits. The standard lists indicators such as being primarily responsible for fulfilling the promise, carrying inventory risk and having discretion over price, but Deloitte stresses that no single indicator decides the answer. For a model sold through a cloud marketplace, reasonable people can read those indicators differently, which may explain how two labs end up with different presentations of similar deals. The choice moves revenue and gross margin percentages, though it does not by itself change the profit made on the sale.
What to do: before comparing, ask whether partner sales are included gross. Reporting on Anthropic's draft prospectus says nearly half of Anthropic's 2025 sales went through Amazon and Google, so it is not a rounding detail. See OpenAI revenue, losses and cash burn and Anthropic's revenue, losses and margins for the figures on each side.
Question 3: how is gross margin defined?
Gross margin is revenue minus the direct cost of delivering the product, divided by revenue. It is not profit, because it ignores research, sales and overhead. The argument is about what counts as a direct cost.
The Anthropic example is a clean illustration. One tracker estimates Anthropic's 2026 gross margin at 44% to 60% depending on whether the amortized cost of training is netted against quarterly compute spend. The Financial Times reported a figure above 80% in September 2026, but only before partner revenue share payments and training costs. That is a narrower definition. A third measure, reported by Yahoo Finance via the same tracker, is compute cost per revenue dollar: $0.71 in Q1 2026.
For OpenAI, reporting traced to The Information says gross margin was 33% in 2025 and 40% in 2024. A deeper look is in OpenAI gross margin and inference spend.
What to do: ask four things. Is training included? Is the cloud partner's share deducted from revenue or counted as a cost? Are free users' costs included? Is it before or after stock based pay? Any margin that does not say is a headline, not a measurement.
Question 4: is the loss accounting or cash?
A loss in accounts is not the same as money leaving the bank. Three different numbers show up for OpenAI's 2025: an operating loss of $20.92 billion, a net loss of about $38.5 billion that includes a large non-cash charge tied to its corporate restructuring, and an adjusted loss near $8 billion that excludes non-cash items. Anthropic's draft prospectus shows a similar pattern: a $42 billion net loss, of which roughly $34 billion was a non-cash charge tied to financing instruments that could convert into shares, and an operating loss above $8 billion excluding certain liability writedowns, according to Reuters.
Non-cash charges can be enormous because they reflect changes in the estimated value of the company or of instruments that convert into shares. They matter to investors, but they do not tell you how many months of cash remain.
Cash burn, usually measured as free cash flow, is the better guide to survival. OpenAI's plan, reported by the Financial Times on 18 September, shows negative free cash flow of $278 billion from 2026 to 2030, against a $122 billion round raised in March that the Financial Times says is on track to be spent by 2028. Anthropic held $20.28 billion in cash and short term investments at the end of 2025 according to the New York Post, as relayed by a tracker.
Also check whether stock based pay is in the operating loss. The Wall Street Journal's $9.3 billion and $12.3 billion quarterly operating losses for OpenAI include it. Stock based pay is a real cost to existing owners, but it is paid in shares, not cash.
What to do: for any loss, ask "is this the number that leaves the bank?" If not, find the cash flow or burn figure.
Question 5: is the commitment cancellable?
Large AI labs sign contracts to buy computing power years in advance. Reported totals are huge: $518 billion of obligations in Anthropic's draft prospectus, and about $856 billion of computing and infrastructure spending planned by OpenAI through 2030.
The key phrase is non-cancellable. In plain terms, a non-cancellable commitment is a promise you cannot walk away from without paying, even if you need less than you expected. A cancellable one lets you cut or exit on notice, maybe with a fee. A "take or pay" contract is a common version: you pay for a fixed amount whether or not you use it. A commitment is not the same as money spent or owed today. It is a future obligation that may or may not appear on the balance sheet as a liability, and companies usually disclose it in a notes section.
Why does this matter? If revenue grows as planned, large commitments are just capacity you needed. If revenue stalls, the same commitments become fixed costs. That is the difference between a bet and a debt, and only the contract terms tell you which.
A public filing is where this becomes visible. The SEC's rules for the management discussion section of a US filing say "The discussion should analyze material cash requirements from known contractual and other obligations" (Regulation S-K, Item 303(b)(1)). Once a lab files publicly, the size and timing of its compute obligations should appear in that discussion and in the notes to the accounts, rather than in leaks. Look for a schedule by year, because a commitment that falls mostly in 2029 and 2030 carries different risk from one due in the next twelve months.
Deeper explainers: Anthropic's compute commitments and OpenAI's compute commitments and Stargate.
Three further warnings. Headline numbers such as "up to $1.4 trillion" over eight years, which Fortune reported as an OpenAI announcement via TechCrunch, are ceilings, not signed bills. Reporting also notes that OpenAI's compute estimate moved by about 43% between February and July, though the two figures may count different categories. And the July presentation behind the newest numbers was reportedly prepared to win a compute deal, which is a reason for caution about how neutral it is.
Reading a projection
Plans are not results. A good habit is to look at how often the plan has moved. TheNextWeb notes the OpenAI projection has been revised twice this year, from about $305 billion of negative free cash flow in May to $278 billion in July. Fortune reported in November 2025 that OpenAI expected a $9 billion net loss for 2025, and leaked audited accounts later showed a $20.92 billion operating loss. The measures differ, but the gap is instructive.
A good projection tells you the assumptions. Ask what revenue growth it needs, what price per unit of computing it assumes, and who prepared it and why.
Concentration and where the money comes from
Check who pays. According to a tracker summarizing Anthropic's draft prospectus, nearly a quarter of 2025 revenue came from two customers, and 47% of sales were routed through Amazon and Google. That is a business risk and an accounting question at once. It affects how much revenue is gross or net, and how exposed the company is to a few buyers. Fortune also reported in November 2025 that Anthropic's corporate customers made up about 80% of its revenue, which is a different kind of concentration: business customers versus consumers.
What is firm and what is uncertain
Firm: the structure of the questions. Run rate versus booked, gross versus net, adjusted versus accounting, committed versus spent. Those hold for any company.
Uncertain: nearly every number. Almost all of it is leaked or unaudited. Semafor's author warns that leaked revenue numbers rarely match what companies report once they face regulators and investors. Several trackers repeat figures without naming a source; those are left out here.
What to watch next
- Public prospectuses from either lab. Anthropic confidentially submitted a draft on 1 June 2026, and OpenAI filed confidentially on 8 June according to one tracker. When these become public, they must define revenue, cost of revenue and obligations in audited form.
- A consistent gross or net presentation of partner sales.
- Notes on commitments that say how much is non-cancellable and by when.
- Whether either lab reports free cash flow, not just operating loss.
Safeguard offers SBOM and dependency scanning for the software you depend on, with a free scan at https://gold.safeguard.sh/scan.
Quick check
Which number should I use to compare two AI labs?
Use booked revenue for the same period, on the same gross or net basis, if you can find it. If you only have run rates, check that both count partner sales the same way before comparing.
Does a bigger net loss mean a lab is running out of cash faster?
Not necessarily. Net loss can include large non-cash charges. Look at free cash flow and cash on hand, which show how fast money is actually leaving.
Is a reported compute commitment money already spent?
No. It is a future obligation. Whether it is binding depends on whether it is non-cancellable, which only the contract terms and the notes in audited accounts can show.
Sources
- OpenAI's September annualized revenue run rate, Channel Insider, 9 October 2026.
- Anthropic is gaining on OpenAI's revenue but has not yet eclipsed it, Semafor, 10 April 2026.
- Is Anthropic profitable in 2026?, Value Add VC, updated 15 October 2026, relaying Reuters, Financial Times and others.
- OpenAI revenue 2026, Value Add VC, updated 15 October 2026.
- OpenAI revenue and valuation statistics (October 2026), Exploding Topics, updated 12 October 2026.
- OpenAI's $856bn compute plan versus the $600bn reset, TheNextWeb, 19 September 2026.
- OpenAI forecasts cash burn near $280 billion by 2030, FT reports, Reuters via The Star, 19 September 2026.
- OpenAI and Anthropic miss gross margin targets as inference costs rise, BigGo Finance, 25 February 2026.
- Revenue recognition: evaluating whether an entity is acting as a principal or as an agent, Deloitte Accounting Spotlight, 28 May 2019.
- 17 CFR 229.303, Management's discussion and analysis, Regulation S-K Item 303, Legal Information Institute.
- OpenAI says it plans to report stunning annual losses through 2028, Fortune, 12 November 2025.
Last reviewed: 19 October 2026.
