Revenue figures for these companies get thrown around constantly — annualized to numbers that make great headlines. Whatever the exact figure, the reading trap is the same. This piece gives you the one distinction that fixes every AI revenue headline, plus the specific line items to demand before you believe any of them.
Ranjan Roy, on the Big Technology Podcast, has a version of this complaint. Multiplying by twelve "always," he says, is COVID-era extrapolation — like pricing Zoom off two pandemic months. He's right, as far as it goes. But it stops one step short. The multiplier isn't the deepest problem. The billing model is.
First, kill the word "ARR"
Classic ARR — the kind a SaaS company reports — has three properties you can check.
It comes from committed contract value. You take total contract value, divide by the term in years, and get annual contract value. There's a signed multi-year commitment underneath. It shows up as deferred revenue on the balance sheet, which an auditor watches burn down. And it carries a retention assumption, because the contract is what enforces recurrence.
Now look at what OpenAI and Anthropic actually sell. Two engines. First, subscriptions — ChatGPT Plus, the business tier, and Enterprise. Recurring-ish, but the consumer and mid-tier plans are month-to-month cancelable, no committed term. Second, API and consumption — billed per token consumed. No floor, no commitment, no minimum.
Neither produces the deferred-revenue balance a multi-year SaaS contract does. There's no ACV to divide, because there's no committed TCV. So when someone calls metered token revenue "ARR," they're borrowing the recurrence guarantee from an instrument that doesn't have one. If usage drops, the number drops the same month. That's the tell.
Next time you see an AI revenue figure, find out whether it's labeled contracted, committed, or run-rate. For these two companies it's essentially always run-rate. If the source can't say, discount it.
The meter runs backwards from SaaS
This is the part outsiders miss, and it inverts the usual intuition.
In SaaS, price per seat is sticky and the arrow points up-and-to-the-right — you add seats, you expand. In AI, two forces push the opposite way. The vendor keeps cutting per-token price to win volume and fend off cheaper rivals; per-token prices have fallen sharply over successive model releases. And the buyer actively optimizes consumption down. Prompt caching can materially cut the cost of repeated context — vendors publish their own cached-input discounts, so price it against their pricing pages rather than trusting a number. Swapping a frontier model for a mini or distilled version runs the same task for a fraction of the price. Trimming context and output does the rest.
Follow that one step. A customer flips on prompt caching and their bill drops meaningfully — with zero churn event. They didn't leave. They got happier and stickier. But the vendor's run-rate just fell.
So run-rate can decline while the business gets healthier. The headline number and the underlying health move in opposite directions. Token retention is not seat retention. Tokens are deflationary by design.
If you're a buyer, price out the same workload on a frontier model versus a mini model, and with caching on versus off, using the vendor's own published rates. The spread is often large. That spread is the vendor's "revenue" that can evaporate without anyone canceling.
The same workload can get counted more than once
Consider how a large enterprise might evaluate these tools. You take one workload and run it across OpenAI, Anthropic, and Bedrock at once to benchmark quality and price. Where that happens, the same task shows up in multiple vendors' revenue during evaluation.
This is a plausible accounting trap, not a law of nature — but it points somewhere real. To the extent parallel bake-offs are common, industry-level run-rate is structurally double- or triple-counting the same underlying demand. Add up every vendor's run-rate and you may be overstating real end-demand for AI.
The trap resolves itself, and the resolution is uncomfortable. When a workload clears budget review and goes to production, it tends to consolidate to one vendor. The losers drop that revenue. So as evaluations conclude, some vendor's run-rate should fall — not because AI failed, but because a multi-count collapses to a single-count.
Much of the recent AI spend has plausibly been evaluation and experimentation budget, not committed production P&L — though nobody discloses the split. The only dollar worth underwriting is the one that survived a budget review and went to production.
Separate the two kinds of cost, or you'll get the margin wrong
There's a lazy skeptic move worth avoiding: "the margins are deeply negative once you load in training." An auditor throws that out, and it's worth seeing why, mechanically.
Inference — the cost of actually serving a request, mostly GPU rental on Azure or AWS — is COGS. It hits gross margin. Training — building the next model — is R&D and capex. It does not hit gross margin. Mix them and your analysis is wrong.
On a per-request basis, at list price, inference gross margin can be positive. The bleeding is elsewhere: the free ChatGPT tier you serve at a loss, promotional credits, and below-cost frontier pricing to win logos. So the honest read is: per-request list-price margin, plausibly fine; blended margin including free-tier serving and loss-leader pricing, that's the trap.
Why does this matter for reading the top line? Because when revenue jumps fast, serving cost jumps too. Unlike Zoom — whose marginal user cost was near zero, so pandemic revenue was almost pure contribution — every incremental AI token carries real GPU cost. Revenue can go up while losses go up faster. That's exactly why Zoom is the wrong comp, even though the "don't times twelve" instinct is right.
"More in a month than all of last year" — true and nearly meaningless
The standard counterpoint: a fast-growing AI vendor now earns more in a single month than in all of a year ago, so the growth is real regardless of how you annualize it.
It's true. It's also nearly meaningless. Off a tiny base, any hypergrowth line clears that bar. It tells you the first derivative is positive. It tells you nothing about durability, revenue composition, or margin. It's a rhetorical move that makes anyone asking hard questions sound like they're denying growth. The growth is real. That was never the question.
The question nobody outside can answer yet
Strip away the arithmetic fights and one real disagreement remains. Is the second derivative positive because of durable adoption — workloads that go to production and compound — or because a capability step-change pulled a wave of demand forward, which then plateaus?
That's the crux. Not the multiplier. Composition is destiny: it decides compound versus collapse. And you can't settle it from a headline number, because the headline is built specifically to hide it.
What to hunt for instead
The clean figures, in order of usefulness.
RPO — remaining performance obligations — and deferred revenue on the balance sheet. This is contracted future revenue. Consumption vendors have little of it. Its absence is itself the signal.
NRR on production workloads — post-budget-review only, not blended with evaluations.
The monthly cohort retention curve — does a cohort's month-1 spend survive to month-12, net of the customer's own optimization?
Incremental gross margin per cohort — separating inference COGS from everything else.
None of these are in the headline. When a company won't disclose them, the non-disclosure is the answer.
FAQ
Is AI annualized revenue misleading?
The direction of growth is real; the annualization overstates the level. It multiplies one month of usage-metered billings by twelve with no committed contract underneath, so it's run-rate, not ARR.
What's the real difference between run-rate and ARR?
ARR comes from signed, committed contract value and shows up as deferred revenue. Run-rate is just last month × 12. For OpenAI and Anthropic, it's overwhelmingly run-rate.
Can AI revenue actually fall while the business improves?
Yes. A prompt-caching change or a model downgrade can cut a bill sharply with no cancellation. The customer gets stickier; the run-rate drops.
Do these three things now
The open question worth sitting with: how much of that curve is production adoption versus a demand pull-forward that plateaus? Nobody outside can prove it — so notice who claims they can.
Then go do three things. First, whenever you see an AI "ARR" figure, get it labeled committed or run-rate before you believe it. Second, look for RPO and deferred revenue — if they're absent, treat the top line as usage that can drop any month. Third, price the same workload frontier-vs-mini and cache-on-vs-off against the vendor's published rates; that spread is the revenue that can vanish without a single customer leaving.
So here's the question worth leaving you with: next time you see one of these numbers, what will you demand before you believe it?