Baseten raises $1.5B Series F as open-source inference demand surges across every sector

Baseten has closed a $1.5B Series F, one of the larger infrastructure rounds in the current AI cycle. Tuhin Srivastava, the company's co-founder and CEO, frames the raise as a straightforward response to demand: inference is the cost of goods for AI, and that demand is growing across every sector with no sign of slowing.

What Baseten does

Baseten operates as an inference cloud, renting compute from 20 cloud providers across 90 regions and layering software primitives on top for inference, reinforcement learning, routing, and eval sandboxes. The model is aggregation rather than ownership — Baseten sits between customers and heterogeneous compute, handling the infrastructure complexity so customers don't have to.

Srivastava says going down the stack, including owning compute directly, is likely inevitable as the business scales, but the near-term product focus stays on inference software.

“We raised our Series F — a billion and a half dollars. Open source models are getting very good. Inference is the COGS of AI. It's gonna grow indefinitely. We rent compute from 20 different clouds, 90 different regions, and we have all these different software primitives around RL and inference. We call it an inference cloud.”

The open-source thesis

The round is built on a bet that open-source models will remain a durable, competitive alternative to frontier closed models. Srivastava's argument is that as open-source quality improves, any company that combines its own data, workflows, and user signal with a strong open-source base can match or beat frontier performance on specialized tasks, at lower cost. He points to the Thinking Machines and Bridgewater collaboration as a visible example of that logic playing out. Enterprises, he argues, are now recognizing they need to build this capability now or risk being economically exposed later.

Open-source FUD

Srivastava addresses several common objections directly:

• Benchmark gaming / distillation concerns — he dismisses the framing as focusing on the wrong problem. The models exist and are improving; the question is how to build on them, not whether they're legitimate.
• China closing its models at larger training scales — he points to US-funded alternatives like NVIDIA's Nematron and Microsoft's MAI models as evidence that domestic open-source investment is deep enough to sustain the ecosystem, and argues the incentive to keep models open is structural.
• Chinese backdoors in model weights — he says there is no evidence for this, and that in practice these models run in isolated, air-gapped environments with security controls that make network exfiltration implausible regardless.
• Government banning open-source inference — he doesn't claim to know how regulators will move, but believes a ban would run against US economic and strategic interests.

The through-line across all four is the same: the real priority is building infrastructure for AI sovereignty, where companies own their intelligence stack rather than depending entirely on a small number of closed providers.