Baseten buys Blaxel to put the agent sandbox on the same network side as the model call

An agent step is several calls. The model returns something, code runs somewhere with a filesystem, a tool or an MCP server answers, then the agent waits. When the serving layer and the execution layer come from different vendors, each of those handoffs crosses the boundary between them. Blaxel's argument in its announcement thread is that inference has to sit beside compute, storage and networking because that latency compounds every time an agent crosses a network boundary ^[11]. Baseten, whose business is serving and training models ^[24], says the combination lets it optimize the model call, the execution environment, state and the network path as one system ^[22].

The figure under that argument is 25 milliseconds. Blaxel says its per-agent micro-virtual-machine sandboxes suspend and resume in roughly that time, preserving memory state and scaling down while idle ^[8]. It is a measurement of Blaxel's own environment. It transfers to your agent if the resume lands inside a step a user is already waiting through, and if the agent idles in blocks long enough that suspending is worth the bookkeeping. Where it pays is idle compute and network distance, which accumulate across a task as the agent executes code, reads files and contacts tools ^[21]. A sandbox placed far from the serving endpoint pays the boundary cost on every one of those steps ^[11].

Baseten was founded in 2019 by Tuhin Srivastava, Amir Haghighat, Phil Howes and Pankaj Gupta after they worked together at Gumroad ^[16], on the thesis that model serving would become the bottleneck as machine learning moved into customer-facing products ^[17]. Haghighat said people initially called the inference focus "stupid," because a model had to be trained before it could be served ^[18].

Blaxel's six founders came out of ForePaaS, the analytics platform OVHcloud acquired in 2022 ^[12][13]. Paul Sinai, Blaxel's co-founder and CEO, told VentureBeat in a July 17th, 2025 interview that conventional cloud providers were designed for the software-as-a-service era, while agents required their own type of infrastructure ^[19]. The seed round closed in July and the acquisition was announced on September 10th, about two months apart ^[23].

The near-term product is narrower than the rationale. Baseten says it will add products built on Blaxel's technology, beginning with its Sandboxes ^[5]. For a team running an agent across two vendors today, the question to put to that product is whether a sandbox and a serving endpoint can be pinned to the same region, and what the resume path costs when they cannot. Sinai's stated interest in the deal is having the model-serving layer inside the same organization as the sandboxes, storage and networking his six-founder team has built since 2024 ^[20].