Baseten is in talks to raise US$1 billion at an US$11 billion valuation, according to reports on 26 May. The round is not closed, so the figure is a target, not a done deal. If it lands, it more than doubles the US$5 billion valuation Baseten set in January.
What Baseten does
Inference is the part of AI that happens after training — the model running in production, answering a query, generating a response. Baseten rents Nvidia servers to application developers and helps them deploy, customise and run mostly open-source models on that hardware. It sells the unglamorous layer between a model and a working product, which is exactly the layer demand has rushed into this year.
Why investors are circling
The revenue is the draw. Baseten's annualised revenue reached roughly US$600 million by the end of the first quarter, up from about US$200 million at its start, per PYMNTS. A tripling in three months is the kind of curve that pulls a valuation up with it. Its January round was led by IVP and CapitalG, with Nvidia putting in US$150 million — a sign the chipmaker wants its hardware booked through fast-growing inference resellers.
The read
Two AI-infrastructure rounds in a week — Baseten here, the data layer repricing at Snowflake — point at the same thing. The market seems to have decided that the money in this cycle is in infrastructure — the compute, serving and plumbing that every AI product needs. Whether US$11 billion is the right price depends on margins holding as Baseten rents hardware it does not own and resells at a markup the cloud giants would happily undercut. The revenue is solid, but the durability of Baseten's margins is the open question.
The round closed four weeks later, half as large again
The deal that eventually landed was bigger than the one being reported.
Baseten closed a Series F of about US$1.5bn on 22 June, not US$1bn, structured across two tiers so that some investors came in at the US$11bn valuation reported here and others at US$13bn. Altimeter Capital, Conviction, Spark Capital, Sands Capital and Wellington Management co-led. The company had been valued at US$5bn in January, which puts the move at roughly 2.6 times in five months rather than the doubling anticipated above.
The revenue figure held up as the reason. Annualised revenue of about US$600m at the end of the first quarter, up from roughly US$200m at its start, is the curve the round was priced against, and no larger figure was needed to justify the higher tier.
The question about the spread is now the whole sector's question
The doubt raised here was whether margins hold for a company renting hardware it does not own and reselling at a markup the cloud giants could undercut. Three months on, that has stopped being a Baseten question.
Five hyperscalers have disclosed roughly US$1.2 trillion of lease obligations, of which about US$725bn has not yet commenced and therefore sits in the notes rather than on the balance sheet, with Goldman expecting around 35 per cent of 2027 AI capital expenditure to be debt-funded. The companies that could undercut an inference reseller are themselves financing the capacity to do it, on terms that are only partly visible.
This creates a two-sided risk for a business like Baseten. A competitor carrying committed lease obligations has a strong incentive to fill capacity at whatever price keeps it booked, which compresses the reseller's spread. It also means the reseller is not the party holding the duration risk on a data centre.
The layer kept attracting money, and it kept moving
The reading above was that the money in this cycle is in the picks and shovels. That held, and the definition of which shovel expanded.
Together AI booked up to 10,000 Nvidia B300s at a Chennai campus in August, its second buildout in a week, which is an inference provider taking on exactly the physical commitment Baseten's model avoids. River AI raised US$1.1bn at two months old with an API still in preview, and both Nvidia and AMD Ventures were in the same round. Stripe agreed to buy OpenRouter, which routes across roughly 400 models and meters the spend.
These are three different bets on the same layer, owning the hardware, the models or the routing. Baseten sits between them, renting hardware to serve other companies' models, which is the position with the lowest capital requirement and the least defensible moat. The Nvidia investment noted above is worth rereading in that light. A chipmaker putting US$150m into a fast-growing reseller is booking demand for its hardware, and it made a comparable bet on River AI alongside its closest competitor two months later.
What is still unsettled
Revenue tripling in a quarter was real and the valuation followed it. Neither then nor now has anyone published the margin.
Margin is the figure that would settle the question, and inference resellers do not disclose it. Until one does, a valuation in this layer is a bet on revenue growth outrunning a spread nobody outside the company can measure.
View author profile → · Editorial policy
