Baseten has acquired Blaxel, bringing together Baseten’s AI model inference and training infrastructure with Blaxel’s execution, storage and networking technology for autonomous agents. Financial terms of the transaction were not disclosed.
The companies plan to build an integrated infrastructure platform where developers can train and serve AI models while also operating long-running agents within the same system.
Baseten has focused on infrastructure for training and serving AI models with an emphasis on performance, reliability and efficiency. Blaxel, meanwhile, has developed a stateful execution layer specifically designed for AI agents, including isolated computing environments, persistent storage and production-grade connectivity.
The combination is intended to address a growing infrastructure challenge created by agentic AI applications. Unlike traditional model inference workloads, autonomous agents can execute code, interact with tools and APIs, maintain state over extended periods and repeatedly communicate with AI models while completing complex tasks.
Blaxel has spent more than 18 months developing infrastructure intended specifically for those workloads.
Its technology includes Sandboxes, isolated environments based on individual micro virtual machines that allow agents to safely write and execute code while maintaining separation from other workloads.
Blaxel said its sandboxes can suspend and resume in approximately 25 milliseconds, which the company says is as much as five times faster than competing sandbox products. The environments can remain idle for extended periods at close to zero computing cost while preserving their state for future use.
The company’s infrastructure also includes Agent Drive, a distributed filesystem designed to persist the files, code and working context generated by AI agents across different sessions and execution environments.
Blaxel has additionally developed networking infrastructure enabling agents to securely communicate with tools, APIs, Model Context Protocol servers and other agents while maintaining access controls and workload isolation.
One component Blaxel had not developed internally was model inference infrastructure. The company said open-weight and custom models are becoming increasingly important to production agent deployments, making it important for inference to operate close to agent compute, storage and networking rather than across distant network boundaries.
That requirement was a significant factor behind the combination with Baseten.
Baseten operates an inference platform across dozens of regions and has raised more than $2 billion to build its AI infrastructure capabilities. By combining the two platforms, the companies plan to colocate model inference with the compute, storage and networking resources agents need to perform tasks.
The longer-term objective is to create an infrastructure layer capable of supporting millions of autonomous agents while optimizing performance, reliability, security and economics.
The combined platform is expected to support agent execution, model inference and training within one system. Agents could operate directly alongside the models they call while maintaining persistent environments across turns and sessions.
The companies also envision connecting post-training workflows with the same storage layer used by agent sandboxes, potentially allowing models to be improved based on the activities and outputs generated by agents operating on the platform.
Blaxel already supports large-scale agent workloads. Sapiom, for example, runs hundreds of millions of agent loops using Blaxel infrastructure, according to the announcement.
For existing Blaxel customers, the company said its current product and support operations will continue without immediate changes. The existing team will remain in place and continue developing and shipping features.
Over time, Baseten plans to introduce additional products based on Blaxel’s infrastructure primitives, beginning with Sandboxes.
Those capabilities will be combined with Baseten’s inference and training products as the company expands from primarily model infrastructure into a broader agentic infrastructure platform.
The acquisition reflects a broader shift in AI infrastructure requirements as developers move beyond individual model calls toward autonomous systems that can execute tasks over longer periods, retain memory and context and continuously interact with external systems.
For Baseten, adding Blaxel expands its position across more of the underlying technology stack required to operate those applications, from model training and inference to agent execution, persistent storage and secure networking.
