Baseten, which runs infrastructure for training and serving AI models, said on September 13, 2026 that it had acquired Blaxel, a San Francisco startup that builds computing infrastructure for autonomous AI agents. Financial terms were not disclosed.
The companies plan to build a single platform on which developers can train and serve models and also run long-lived agents in the same system, with inference placed close to the compute, storage and networking that agents use.
Blaxel's current product and support will continue without immediate changes, and its team will stay in place and keep shipping features. Over time Baseten plans to launch new products based on Blaxel's technology, starting with sandboxes.
What Blaxel built
AI agents behave differently from ordinary model calls. They write and run code, call tools and APIs, keep state for long periods and talk to models repeatedly while completing a task. Blaxel spent more than 18 months building infrastructure for those workloads.
Its Sandboxes are isolated environments based on individual micro virtual machines, which let an agent execute code safely apart from other workloads. Blaxel says they can suspend and resume in about 25 milliseconds, as much as five times faster than competing products, and sit idle for long stretches at close to zero cost while keeping their state.
It also built Agent Drive, a distributed filesystem that keeps the files, code and context an agent produces across sessions, and networking that lets agents reach tools, APIs, Model Context Protocol servers and other agents under access controls. Customers include Sapiom, which runs hundreds of millions of agent loops on Blaxel, according to the announcement.
Why Baseten is buying
The one piece Blaxel had not built was model inference. The company said open-weight and custom models are becoming more important in production agent systems, which makes it useful to run inference next to the agents rather than across distant network boundaries.
Baseten operates an inference platform across dozens of regions and has raised more than $2 billion to build its infrastructure. With Blaxel it moves from serving models to hosting the agents that call them. The companies also want to connect post-training to the same storage layer that agent sandboxes use, so that models could be improved using what agents do on the platform.
Their longer-term goal, they said, is an infrastructure layer that can support millions of autonomous agents.
The founders and the funding
Blaxel has six co-founders. Chief executive Paul Sinai was co-founder and CEO of ForePaaS, a data platform later acquired by OVHcloud, and he recruited former colleagues from ForePaaS and OVH: Christophe Ploujoux, Nicolas Lecomte, Thomas Crochet, Charles Drappier and Mathis Joffre.
The company went through Y Combinator and in 2025 raised a $7.3 million seed round led by First Round, with Y Combinator, Liquid 2, Multimodal and others taking part. By the end of its YC batch it had deployed infrastructure across 16 regions and was handling millions of agent requests a day, it said.
Blaxel charged customers on a pay-as-you-go basis, only when agents were actively processing tasks, and offered SOC 2 and HIPAA compliance and data residency controls. One customer was running more than a billion seconds of agent runtime a month on the platform at half the cost of typical serverless infrastructure, the company said when it raised its seed round.
What the deal means for Baseten
Blaxel pitched itself as an "AWS for AI agent companies" and targeted smaller AI-first startups, arguing that existing clouds were designed for the software-as-a-service era of human users. Joining Baseten gives it the inference capacity and capital of a much larger company. For Baseten, which until now focused on training and serving models, the deal extends its platform to the agents that call those models, from execution and persistent storage to secure networking.
