BREAKING September 16, 2026 4 min read

Baseten Acquires Blaxel for 25-Millisecond Agent Sandboxes

Official Baseten x Blaxel acquisition image: the two founders with the Baseten and Blaxel lockup

Baseten has acquired Blaxel. The joint post, dated September 10, 2026, puts the product numbers in the first technical paragraph: each agent gets its own microVM; suspend and resume in 25 milliseconds; a claim of up to 5x faster than other sandbox products. Deal terms were not disclosed.

The Next Web says Baseten co-founder and CTO Amir Haghighat told Axios Pro about the deal, and that Axios reported it first. Baseten posted the announcement later the same day. The byline on the post is Haghighat, Baseten co-founder Tuhin Srivastava, and Blaxel's Paul Sinai.

The thesis is colocation, not a sandbox bolt-on

Baseten's side of the stack is model inference and training. Blaxel's side is the stateful execution layer: isolated sandboxes, persistent storage, and production connectivity. The post's line is that joining the two is how they "colocate all four, bringing the brain next to the muscles." Inference called from another data center adds latency on every agent loop.

That is the deal, as the companies wrote it. Agents write and run code, call tools, and carry state across long tasks. A shared process is the unsafe version of that. A microVM per agent is the version they shipped. The missing primitive on Blaxel's side, the post says, was inference sitting next to compute, storage, and network.

25ms, idle months, more than $2B

The acquisition post's sandbox claims, quoted as company figures:

  • Each agent gets its own microVM.
  • Suspend and resume in 25 milliseconds.
  • Up to 5x faster than other sandbox products.
  • A sandbox can sit idle for months at close to zero cost and come back with state persisted, "before the model finishes its next sentence."
  • Baseten inference "running at scale across dozens of regions, with more than $2B raised to build it."

Blaxel's sandbox docs match the speed claim as product spec: sandboxes resume from standby in under 25 milliseconds, and the overview leads with "sub-25ms cold starts." The docs also describe scale-to-zero after a few seconds of inactivity, with filesystem and process state snapshotted in standby. That is the idle-cost story in engineering language.

The same post names the other Blaxel primitives. Agent Drive is a distributed filesystem for files, code, and working context that outlives a single run. A rebuilt networking layer connects tools, MCP servers, APIs, and other agents with isolation and access controls. Blaxel says it spent more than 18 months building those pieces. It cites Sapiom running hundreds of millions of agent loops on the platform.

"Blaxel handled that kind of growth in a way I'd normally expect to take a year or more to reach."

David Zhang, founding engineer at Sapiom — Baseten / Blaxel acquisition post, September 10, 2026

Blaxel keeps shipping. Baseten starts with Sandboxes.

The customer note is explicit. Blaxel is continuing. The support team stays. "The product doesn't change, and features will keep shipping." Over time, Baseten "will add new products based on Blaxel's primitives, starting with Sandboxes." Charles, Mathis, Chris, Nico, and the rest of the Blaxel customer team "remain the same people, doing the same work, with more behind them."

The Next Web, citing the release, adds a near-term read: Blaxel expands Baseten's support for code execution, tool use, and session persistence, with later integration into inference, storage, observability, and enterprise infrastructure. Those integration steps are TNW's summary of the release, not a separate product schedule from Baseten.

$1.5B Series F at $13B, then an undisclosed buy

Baseten's own June 22, 2026 Series F post is the primary for the valuation: a $1.5 billion Series F that "achieved a $13B valuation," led by Altimeter Capital, Conviction Partners, and Spark Capital, co-led by Sands Capital and Wellington Management. The acquisition post's "more than $2B raised" sits on top of that. TNW repeated the June $13 billion figure and noted the acquisition terms were not disclosed.

The hole in the announcement is the price. The architecture is not: keep the model call and the sandbox in one system, charge near zero when the agent is idle, and resume in 25 milliseconds when it is not.

This article was ultrathought.

Stay ahead of AI

Get breaking news, funding rounds, and analysis delivered to your inbox. Free forever.

Related stories