Nine billion at the end of 2025. Forty-seven billion in May. More than sixty-five billion by late July.
That is Anthropic’s annualized revenue, as reported by TechCrunch, and the interesting number is not the total. It is the interval.
The company added roughly eighteen billion in annualized revenue in about ten weeks. Whatever else that curve describes, it is not an experiment budget.
This is procurement, not curiosity
Consumer products do not grow like that at that size. Consumer growth flattens, because there is a ceiling on how many humans will pay a subscription.
Enterprise contracts grow like that. They grow when a pilot converts into a platform commitment, and when that commitment then expands across departments in the same fiscal year. That is what enterprises actually buy when models commoditize.
Investors reportedly expect the year to close somewhere between 100 and 120 billion, against a May valuation of 965 billion and public market ambitions above two trillion, possibly as soon as this autumn.
Compare that with OpenAI, which doubled to roughly 40 billion from 20 billion. Both curves are extraordinary. Only one of them is accelerating into a category the other one defined.
I have no interest in the horse race. I care about what it implies for the people reading this, which is that a very large amount of money has already been moved into this line item by companies that look like your customers, your competitors, or you.
That money came from somewhere
Here is the part that should show up in your planning.
Enterprise software budgets do not expand by sixty billion dollars because everyone got a raise. That money was reallocated, and the sources are predictable: agency retainers, outsourced production, headcount in support functions, tooling that AI now absorbs, and the classic first casualty, discretionary marketing spend.
If you sell services to mid-market and enterprise buyers, you have already felt this. The budget conversation now includes a comparison you were never part of before, where your retainer is weighed against an internal capability the client believes they can build.
Often they are wrong about that. It does not matter. The comparison is happening regardless, and “we are better than the model” is not an argument that survives a CFO review.
What survives is a specific claim about commercial outcomes that the alternative cannot produce, ideally with your own numbers attached.
The agencies I see holding their pricing are the ones that changed what they sell. Not hours, not deliverables, but the judgment about what should be built and the accountability when it does not work.
The ones under pressure are still selling production volume, which is precisely the thing the client just bought a subscription to.
The pricing risk nobody has modelled
There is a second implication, and it points the other way.
If you have built anything on top of these APIs, your unit economics currently depend on a vendor in the steepest part of a growth curve, carrying enormous capital costs, heading toward a public listing.
Every one of those conditions creates pressure on price, and not downward pressure forever. Cheap inference has been subsidised by a land grab. Land grabs end, and the model layer getting cheaper was never a guarantee that it stays that way.
I am not predicting a price shock. I am pointing out that most product margins I see modelled assume today’s token costs continue indefinitely, which is a forecast, not a fact, and nobody labels it as such in the spreadsheet.
The sensible architecture response is unexciting and worth doing anyway. Keep your prompts, evaluation sets, and orchestration logic in your own layer rather than welded to one provider’s interface. Know what your product costs to run per customer, per month, at current prices, and what happens to that number if inference doubles.
Most teams cannot answer the second question. That is the actual risk, not the vendor.
There is a cheaper version of this discipline for anyone who thinks it sounds like enterprise overhead. Run your three most expensive workflows against a smaller model once a quarter and record what breaks.
Half the time nothing does, and you have found margin. The other half you learn exactly which part of your product depends on the frontier, which is worth knowing before somebody else prices it for you.
What the curve is really telling you
Strip out the valuations and the IPO talk and one fact remains.
Companies with real procurement processes, legal review, security questionnaires and budget holders are signing large contracts at a rate that has doubled twice in seven months. That is the least hype-driven signal available in this market, because those buyers are structurally slow and deeply conservative.
The strategic conclusion is not that you should spend more on AI. It is that the spending decision has already been made around you, by your buyers, and it changed what they expect a supplier to look like.
They now expect you to know this stuff cold, use it in your own delivery, and price accordingly.
The budget moved. The question is whether your offer moved with it.