Five months. That is how long Baseten waited between its last giant fundraising round and its next one. The AI inference company, which raised $300 million at a $5 billion valuation in January, is now close to finalizing a $1.5 billion round at a $13 billion valuation, according to TechCrunch and the Wall Street Journal, citing people familiar with the matter. If it closes, the round would be one of the largest ever for a company whose product is not a model but the plumbing that runs models.
The deal has a wrinkle that is becoming common in AI financing: a split price. Some investors are coming in at the $13 billion valuation, while others are entering at $11 billion, according to the Journal, a structure that lets the company claim a headline number while giving later-stage investors a discount. The round is co-led by Spark Capital, Sands Capital, Altimeter Capital and Wellington Management, the reports said.
Baseten’s business sits at a specific point in the AI stack. The company builds software that lets customers deploy machine-learning models — including open-source models that are far cheaper than the frontier systems from OpenAI and Anthropic — and serve them to users at scale. Its pitch to customers is simple: don’t pay premium prices for a model that a cheaper one can handle; Baseten will route each request to the model that fits the task, and bill for the compute it actually uses.
The inference market has become the industry’s new center of gravity. For the first years of the AI boom, money and attention flowed to training — the process of building models — but the industry has shifted to serving them, and the numbers are enormous. Every chatbot answer, every code completion and every AI feature in a product runs on inference infrastructure, and the companies that operate it are absorbing spending that was once reserved for chip design and cloud data centers.
Baseten’s growth numbers, reported in the funding coverage, tell the story: revenue run-rate near $600 million as of early 2026, up sharply from the prior year, with a customer list that includes well-known AI-native companies such as Abridge, Clay, Cursor, OpenEvidence, Mercor and Notion. The company’s bet is that the winners in AI will be the tools that make models cheap and fast to run, and that the developers who build on top of them will stay for the same reason.
The January round was notable for its investors: NVIDIA participated, reportedly with $150 million, a sign that the chip maker sees inference infrastructure as a strategic layer worth owning a piece of. NVIDIA’s presence also illustrates the ecosystem’s circularity — Baseten buys NVIDIA’s GPUs, and NVIDIA invests in the companies that buy them, an arrangement that has drawn scrutiny but no action from regulators.
The rapid escalation in valuations is the defining feature of the current AI funding cycle. Baseten was valued at $5 billion in January and $13 billion in June, a 160% increase in less than six months, and its competitors are raising at similar clips. Venture funds, worried about missing the next platform company, are paying prices that would have been unthinkable for infrastructure software two years ago, and the split-price structure is their hedge.
The risks are as large as the valuations. Inference demand is real, but it is concentrated among a small number of customers whose own futures are uncertain; a slowdown in AI adoption would hit infrastructure providers first. Competition is intense, with cloud giants offering their own inference services, model makers building vertically, and a wave of startups chasing the same workloads. Baseten’s answer, its founders have said, is specialization — serving the long tail of models and workloads that the giants handle poorly.
The round’s structure also tests the market’s tolerance for creative pricing. Split-priced rounds, in which different investors pay different prices for the same shares, have drawn criticism from governance experts who say they obscure the true valuation, and regulators have begun asking questions about the practice. Supporters counter that the structure lets companies keep growing without forcing every investor to pay the top price.
The inference boom that Baseten is riding has been building for two years. Training a frontier model is a one-time cost, however enormous; serving that model to millions of users is a recurring one, and as adoption spreads the recurring bills come to dwarf the training budgets. Analysts who track the sector estimate that inference spending is growing faster than any other part of the AI stack, and the companies that control it are being valued accordingly. Baseten has also moved up the stack, helping model developers distribute their work and manage access to it, a service that puts it in direct contact with the labs whose models its customers run, and that strengthens its position as a neutral layer between the model makers and the applications.
For the industry, the significance is simple. Capital is flowing to the inference layer at a scale that matches the earlier flow to model training, and the companies building that layer are consolidating. Baseten, with a $13 billion price tag and a $1.5 billion war chest, is now among the best-funded players in that contest — and the bet, shared by its investors and by NVIDIA, is that running AI will be as valuable as building it.
