Summary: The Wall Street Journal reported on June 18 that Baseten was nearing a $1.5 billion round at up to a $13 billion valuation, five months after a $300 million Series E at $5 billion. The company rents compute from about 20 cloud providers and routes requests toward cheaper, sufficient models, with open-source options preferred over defaulting to frontier models. The article says a $13 billion valuation depends on surviving margin pressure from an inference price war.
The Wall Street Journal reported on June 18 that Baseten, a startup building AI inference infrastructure, is nearing the close of a $1.5 billion round at a valuation of up to $13 billion. Its $300 million Series E, announced in January at a $5 billion valuation, came only five months earlier — a paper climb of about 160% in that span.
The capital is chasing the hottest layer of the AI stack right now: inference. With training increasingly held by a handful of labs, the everyday question for every product team has become “how do we run models fast and cheap in production” — and the venture money racing after that problem is what The Next Web dubbed the “inference gold rush.”
Five Months, a 160% Jump: The Timeline
Put Baseten’s recent funding history in sequence and the pace speaks for itself:
- Mid-2025: the $150 million Series D, nine months ahead of the Series E
- September 2025: a valuation of about $2.15 billion (cited by TNW)
- January 2026: the $300 million Series E at a $5 billion valuation, with IVP, CapitalG, and Nvidia among the participants
- June 2026: a reported $1.5 billion raise at up to a $13 billion valuation
Inside a single year, the valuation climbed from just over $2 billion to $13 billion — about a six-fold increase.
One Round, Two Price Tags
Per the WSJ, this round carries a split-priced structure: one group of investors enters at a $13 billion valuation, the other at $11 billion. In a March explainer, TechCrunch observed that this device is often deployed to pump up headline valuations and flatter the lead investors on paper — the genuinely blended valuation sits between the two figures. Co-leading the round are Spark Capital, Sands Capital, Altimeter Capital, and Wellington Management; Conviction also appears on the TNW roster.
The Business Baseten Is In: Selling Inference
Baseten was founded in 2019 and works out of San Francisco, and what it sells is inference infrastructure: the company rents compute from about 20 cloud providers, builds its own inference software layer on top of that capacity, and lets customers deploy and fine-tune models on data they own. Its edge is routing — each request gets steered toward a model that is cheap and good enough, with open-source options deliberately preferred over defaulting everything to OpenAI or Anthropic frontier models. The pitch is straightforward: once inference becomes a commodity input, the account goes to whoever cuts the most cost out of every token — by choosing the right model, the right hardware, and the right cloud for each request. Its customers include Cursor, Mercor, and OpenEvidence, and one of them reportedly finished a task at roughly 30% of the cost the proprietary-model alternative would have required. Speaking to the WSJ, CEO Tuhin Srivastava said “open-source models are getting very, very good,” adding that customers increasingly mix open and proprietary models depending on task difficulty.
The Risk Side of the Gold Rush
Beneath the excitement runs a vicious price war. Open-source models out of China — DeepSeek and Moonshot AI among them — together with Nvidia’s open Nemotron family keep pushing the per-token cost of inference downward. In that same market, Cerebras is now a public company and Fireworks AI competes head-on, while the hyperscalers whose capacity Baseten rents also run managed inference businesses of their own. TNW identifies the central tension: valuing Baseten at $13 billion presumes the company can ride out the margin squeeze the price war creates — it gains as open models get better and stands just as exposed as they get cheaper. A broker earning a spread between expensive and cheap models hits a hard ceiling once everything converges toward cheap.
Three Takeaways for Engineering and Product Teams
Three signals stand out. One: multi-model routing has become the default architecture — choosing whether a request goes to a frontier model or an open one now operates as a cost lever at the product level. Two: open-source models have cleared their production trial with the heaviest users — an inference-hungry product such as Cursor putting its traffic on them is the stress test that counts. Three: nothing about the open-source supply side is guaranteed — Meta’s Avocado frontier model has been reported to possibly go closed-source. Anchoring a cost structure to a single open-source lineage now carries no less risk than betting it all on one proprietary model.