Baseten is raising about $2 billion at a valuation of as much as $30 billion, according to people familiar with the matter, a price that would more than double what the AI software company was worth four months ago. The round is being offered in two tiers, with investors able to buy at $26 billion and a later allocation at $30 billion.
Sands Capital is expected to lead the financing, the people said. Nvidia, which has backed the company, remains an investor. The people asked not to be identified because the terms are private.
The June round valued Baseten at $13 billion. A doubling in that window, before a new product generation has shipped at scale, is the clearest signal yet of where the money in AI has moved.
Training a frontier model is a decision made by a handful of laboratories with a handful of budgets. Running one is a cost paid by everyone who uses it, on every request, for as long as the product exists.
That shift in the bill has pulled capital toward the companies that sit between a model and the customer.
Baseten sells the machinery for that middle layer. A developer hands it a model, and the company puts it on graphics processors, scales the number of machines up and down with traffic, and charges for the compute plus a margin.
The pitch is that a team can ship a working product without hiring the engineers who know how to keep GPUs busy.
Keeping GPUs busy is the whole business. A chip that is idle is a cost with no revenue against it, and the difference between a well-scheduled workload and a badly scheduled one can be a factor of several on the unit economics.
Batching requests, starting containers before traffic arrives, choosing which processor runs which model: these are unglamorous problems with large amounts of money attached.
The demand side has become more expensive per user. Reasoning models generate many more tokens to answer a question than earlier systems did, and each extra token is a billable event. Agentic products, which plan a task and call tools repeatedly, multiply the number of model calls behind a single user action.
A customer that cost pennies a year ago can cost dollars now, and the platform that absorbs that traffic is where the margin accumulates.
Competition is crowded and well funded. Fireworks and Together AI sell similar infrastructure. Modal and Replicate target developers who want to run a model without managing servers. The largest threat is not a startup: Amazon, Microsoft and Google all sell inference inside their cloud platforms and can price it as a feature of a broader contract.
That is the risk in the valuation. Inference prices per token have fallen for three years running, and they will keep falling, which means a platform has to grow volume faster than price declines to keep revenue rising.
The companies that survive that arithmetic are the ones with a software advantage in scheduling and a customer base that does not leave when a cheaper option appears.
Customers have their own reasons to buy the layer rather than build it. A bank or a hospital that runs a model on patient records cannot send that traffic to a public cloud endpoint without a legal review, and a platform that can deploy into the customer’s own machines, or into a region the customer chooses, removes an objection that sales teams hear every week. Data residency has become a purchase criterion rather than a footnote.
The spending pattern has shifted too. Enterprises that spent 2024 and 2025 on pilots are now being asked to justify the bills, and the justification is easier when the cost attaches to a product feature rather than to an experiment.
Inference revenue has the property that investors like: it recurs, and it grows with the customer’s success.
The two-tier structure of the round says something about the market as well. Pricing investors at $26 billion and later ones at $30 billion lets a company reward early commitments and still test how much appetite exists at the higher number.
It is a technique used when demand is strong enough to ration but not strong enough to set a single price.
Nvidia’s position as an investor is worth noting for a different reason. The chipmaker has put money into a long list of companies that buy its processors, and those investments help secure demand for capacity that is sold out quarters in advance.
For Baseten, the relationship is useful and binding: a supplier that is also a shareholder has an interest in allocation, and a customer that depends on one supplier has less room to negotiate price.
The company has not disclosed revenue, and the valuation implies expectations that will be tested by the next two quarters of customer growth. For now the market has made its judgment: the layer that runs models is worth more than it was in June, and the investors writing the checks expect the bill for AI to keep arriving on every request rather than in a single training run.
