How LLM Companies Compute ARR — Run-Rate Math, Revenue Recognition, and What the Headlines Actually Mean
Research page: 2026-08-30. All figures are as reported by the cited sources; most are annualized run rates or preliminary disclosures, not audited GAAP revenue. Treat every number as a snapshot — these companies change disclosures monthly and both OpenAI and Anthropic have confidential S-1 filings pending.
TL;DR
When the news says an AI lab has "$40B ARR" or "$65B ARR", it is usually not annual recurring revenue in the classic SaaS sense. It is a run rate: take a short recent revenue period and multiply it up to a year.
- OpenAI reportedly multiplies total revenue from a recent four-week period by 13 (52 weeks ÷ 4 weeks).
- Anthropic reportedly splits the math: last 28 days of consumption-billed (API) revenue × 13, plus monthly recurring subscription revenue × 12, summed.
- Run rate is not booked revenue. OpenAI booked $13.07B of audited revenue in 2025 but exited the year at a ~$21.4B run rate; Anthropic reported ~$11.5B of Q2 2026 revenue while its July ARR was reported near $65B.
- Two recognition choices can inflate or deflate the number: gross vs net accounting for API revenue sold through cloud channels (AWS, Azure, GCP), and whether prepaid tokens and one-off spikes are included.
- The useful questions are always: run-rate or contracted ARR? Gross or net? What period was annualized? And what does booked revenue look like?
1. Classic ARR vs the AI-lab run rate
In classic SaaS, ARR is a fairly clean number:
ARR = active subscribers × annual subscription price
It is derived from contracts, it is roughly stable month to month, and it is a reasonable predictor of next year's revenue (subject to churn and expansion). Public SaaS companies disclose ARR alongside GAAP revenue, noting that ARR "generally aligns to billings" rather than recognized revenue (e.g. Progress Software's definition).
AI labs have a different revenue shape. A large share comes from consumption (API) usage — tokens burned, not seats held — plus a mix of consumer and enterprise subscriptions, prepaid credits, and newer lines such as ads and licensing. So "ARR" in AI headlines is almost always an annualized run rate: the current pace of revenue, projected forward. The Drops/MTS compilation puts it bluntly: "ARR is a peak-month or peak-quarter figure multiplied by twelve. It is not realized GAAP revenue."
2. The actual formulas
The most detailed descriptions were reported in March 2026 (Reuters Breakingviews, The Information), when Anthropic's growth first drew scrutiny:
| Lab | Reported formula | Multiplier logic |
|---|---|---|
| OpenAI | Total revenue for a recent 4-week period × 13 | 52 weeks ÷ 4 weeks = 13 |
| Anthropic | (Last 28 days of consumption-billed / API revenue × 13) + (monthly recurring subscription revenue × 12) | 28 days × 13 ≈ 52 weeks; subscriptions annualized at 12 months |
Details worth noting:
- Anthropic's subscription side is based on active subscriptions as of the calculation date, not average subscriptions during the period — so a subscription wave right before the snapshot boosts the number.
- The Information reported both companies take "a similar approach": OpenAI multiplies a recent four-week period by 13. Simpler descriptions (e.g. The Elec, Bloomberg coverage) say ARR is "the latest monthly revenue level annualized over 12 months" — same idea, cruder window.
- Simon Willison, who has tracked Anthropic's disclosures, describes the common shorthand as "most recent month × 12" and flagged a leaked description of the formula above.
Why ×13 instead of ×12 for the API leg? A 28-day window is exactly four weeks, so ×13 reaches 52 weeks without calendar-month artifacts. The catch: if the chosen 28-day window is a peak period (a product launch, an enterprise deployment wave, a $500M month), ×13 amplifies that spike into the headline.
3. What goes into the number
For a lab like OpenAI, run-rate ARR aggregates very different revenue types:
| Revenue line | What it is | Annualization treatment |
|---|---|---|
| Consumer subscriptions (Plus, Pro, Team) | Monthly or annual seats | ×12 from MRR (or recognized from annual contracts) |
| Enterprise subscriptions / commitments | Seats or annual minimum commitments | Contract-based or MRR ×12 |
| API consumption (direct) | Pay-per-token usage billed by the lab | Trailing 28 days × 13 |
| API consumption (indirect / cloud channel) | Tokens sold through AWS Bedrock, Azure, GCP and billed by the cloud | Trailing period × 13 — and see gross vs net below |
| Prepaid credits / committed spend | Customer prepays for future tokens | Recognized over consumption, but prepaid inflow can be annualized early |
| Ads, licensing, media (Sora) | Non-recurring-ish lines | Included in the total-revenue window even though not "recurring" |
OpenAI's August 2026 breakdown, per Value Add VC's compilation of disclosures: roughly $22B annualized from ChatGPT subscriptions, ~$12B from API consumption, and ~$6B from Sora, ads, and licensing — against the ~$40B total. The point is the shape: subscription, consumption, and "other" all flow into one headline number, and the composition of that number changes month to month.
4. Gross vs net: the recognition trap
The sharpest critique of AI-lab ARR is not the multiplication — it is revenue recognition for API sold through cloud partners. When a customer buys Claude on AWS Bedrock or GPT via Azure, the billing relationship is between the customer and the cloud. The lab can recognize its share of that revenue either gross (cloud price shown as lab revenue, cost of cloud shown as COGS) or net (only the lab's margin is revenue). Gross recognition makes revenue look larger without making the business larger.
Sacra has explicitly cautioned that Anthropic's revenue "from cloud channels such as AWS, Google, Microsoft may be recognized on a gross basis, making the revenue scale appear larger", and Ed Zitron has pointed at prepaid tokens and cloud-channel revenue as ARR inflators.
Barclays' August 2026 unit-economics work (reported by CLS / China Star Market) quantifies how much this matters:
- A hypothetical Lab A (70% API / 30% subscriptions, gross recognition for indirect API) shows an adjusted gross margin of ~55%.
- A hypothetical Lab B (80% subscriptions / 20% API, net recognition, or no recognition of partner-operated indirect API) shows ~38%.
- That is a 17-percentage-point gross-margin gap driven almost entirely by business mix and accounting choice — Barclays explicitly compares it to the Uber vs Lyft reporting difference.
- Paid-inference margins have climbed from low double digits in 2025 to 50–65%+ in 2026; direct API is the richest line (>80% inference margin), while subscriptions such as Claude Code and Codex run around ~70% because labs subsidize tokens to keep retention.
The revenue-quality question is therefore not "how much did they sell" but "how much of it did they recognize, and on what basis".
5. Why run-rate ≠ booked revenue
Run rate annualizes the latest period, so during hypergrowth it always sits well above trailing booked revenue — often far above 4× the latest quarter.
OpenAI:
| Metric | Value | Source |
|---|---|---|
| Booked 2024 revenue | ~$3.7B | The Information |
| End-2024 ARR | ~$6B | CFO Sarah Friar, Jan 2026 |
| Booked 2025 revenue (leaked audited) | $13.07B | FT/Fortune-verified leak |
| End-2025 ARR | $20B+ | Friar / Bloomberg |
| Mid-2026 run rate | ~$25B (flat Feb–Jul) | The Information / Value Add VC |
| August 2026 run rate | $40B+, +20% MoM in July; enterprise +32% MoM | CNBC / Bloomberg |
The $25B run rate in spring 2026 was a projection off roughly $2B/month — while the audited 2025 number behind it was $13.07B. Both can be true; they answer different questions.
Anthropic:
| Date | Disclosed run-rate revenue |
|---|---|
| Jan 2024 | ~$87M |
| Dec 2024 | ~$1B |
| End-2025 | ~$9B |
| Feb 12, 2026 (Series G) | $14B, "growing over 10x annually in each of the past three years" |
| Mar 2026 | ~$20B (Bloomberg) |
| Apr 6, 2026 (Google/Broadcom deal) | >$30B |
| May 2026 (Series H, $965B valuation) | >$47B |
| Jul 2026 | ~$65B (Bloomberg; Sacra May $47B → July $65B; some trackers $69–74B) |
And the booked numbers underneath: Anthropic reported preliminary Q2 2026 revenue of >$11.5B (vs $787M in Q2 2025, ~14× YoY; Q1 2026 was $4.73B) with its first positive adjusted operating income. A $65B July ARR is ~5.6× the Q2 quarterly number — plausible during acceleration, but only if the growth rate, the recognition basis, and the period chosen all hold up.
6. The peak-month trap, with numbers
Worked example of the Anthropic-style formula. Suppose:
- Last-28-days API revenue: $2.5B
- Subscription MRR: $300M
ARR = (2.5 × 13) + (0.3 × 12) = 32.5 + 3.6 = \$36.1B
Now suppose the snapshot lands in a peak 28-day window ($3.0B API):
ARR = (3.0 × 13) + (0.3 × 12) = 39.0 + 3.6 = \$42.6B
One $500M spike in the window adds $6.5B to headline ARR. Simon Willison passed along exactly this kind of anecdote from Axios: a consultant said a client spent half a billion dollars in a single month after failing to put usage limits on Claude licenses — "times that by 12 and you get an extra $6 billion in annualized run-rate". Whether that usage repeats is the whole game.
7. Why it matters now: IPO math and valuation multiples
Both labs have confidential S-1s filed (OpenAI in June 2026, Anthropic reported to be preparing one), so run-rate figures are being priced as if they are revenue:
- OpenAI: ~$852B valuation vs ~$25B ARR in March–July 2026 ≈ ~34× ARR; ~$40B ARR by August compresses that multiple automatically.
- Anthropic: ~$965B Series H (May 2026) vs ~$47B run rate ≈ ~21× ARR; at the reported $65B July run rate that falls toward ~15×.
The S-1 will replace run rates with audited revenue, gross margin, customer concentration, prepaid/cloud-channel treatment, compute-purchase obligations, and cash flow — the items Ken Koo/ThinkFast and others say are the real test. Simon Willison's counterpoint is fair too: these numbers are disclosed to investors in fundraising rounds, and materially lying there would be securities fraud; the S-1 will be the referee.
8. Checklist for reading an ARR headline
- Run-rate or contracted? Most AI-lab headlines are run rates. Ask if any part is contractual (annual enterprise commitments) vs extrapolated (×12/×13 of a short window).
- What period was annualized? Calendar month ×12, trailing 28 days ×13, or a peak quarter ×4? Peak periods inflate.
- Gross or net? How is indirect API through AWS/Azure/GCP recognized? Same business, very different top lines (Uber vs Lyft).
- What's inside? Subscriptions, API, prepaid credits, ads, licensing — and is any of it one-time?
- Compare to booked revenue. Trailing booked revenue and the latest quarter are the sanity check (OpenAI $13.07B booked vs $21.4B exit ARR; Anthropic $11.5B Q2 vs $65B July ARR).
- Unit economics. Mix matters: API >80% inference margin, subscriptions ~70%, cloud-channel revenue recognition varies. A growth multiple on gross run rate ignores the ~$35–40 of every $100 that flows to the big three clouds for inference.
- Growth and retention. What is MoM/quarterly growth, what is net revenue retention, and does the run rate assume the spike repeats?
9. Related lab numbers for context
Smaller labs use the same language, which makes comparability hard:
| Lab | ARR / run-rate trajectory |
|---|---|
| xAI | ~$1B run rate disclosed Apr 2025; SuperGrok ~$30/mo, SuperGrok Heavy ~$300/mo |
| Mistral | ~$16M end-2024 → ~$312M Dec 2025 → ~$400M Jan 2026 (Sacra); targeting $1B by end-2026 |
| Cohere | ~$35M early 2025 → ~$240M end-2025 (beat $200M target) |
Barclays' aggregate framing: global AI-lab revenue was ~$7B in 2024, projected ~$137B in 2026 and ~$690B in 2028, with year-end ARR running ahead of booked revenue (~$200B by end-2026, ~$782B by end-2028, per the same model) — which is exactly why the gap between run-rate language and recognized revenue is becoming a mainstream accounting discussion.
Sources
- The Information — The Math Behind Anthropic's Mad Revenue Growth (2026-03-24)
- Reuters Breakingviews — Anthropic gives lesson in AI revenue hallucination (2026-03-10)
- Simon Willison — Anthropic's run-rate revenue hits $47 billion (2026-05-29)
- FourWeekMBA — Anthropic's Run-Rate Revenue Went From $1B to $47B in 17 Months (2026-07-18)
- CNBC — OpenAI CFO Friar tells investors that enterprise bigger than consumer by revenue (2026-08-14)
- Bloomberg — Anthropic Revenue Surges to Over $11.5 Billion in Second Quarter (2026-08-14); OpenAI ARR >$40B reporting (2026-08-13)
- The Elec — OpenAI Annualized Revenue Run Rate Estimated Above $40 Billion (2026-08-17)
- 36Kr — Two Sets of OpenAI Pre-IPO Figures (2026-08-18)
- BlockBeats — Anthropic ARR Triggers Valuation Debate (2026-08-18)
- Value Add VC — OpenAI Revenue 2026: $25B ARR and a $14B Loss (updated 2026-08-20)
- CLS via China Star Market — Barclays unit-economics research on AI labs vs cloud take (2026-08-29/30)
- Drops/MTS — Demand and the Revenue Gap (compilation of lab disclosures)
Related pages on this site: AI subscriptions & pricing, model intelligence & cost per task, and the LLM landscape snapshot.