Baseten announces Blaxel acquisition to put agent runtimes beside model inference

Why it matters

The proposed combination would let Baseten optimize the model call, execution environment, state and network path as one system, extending its reach beyond inference into the infrastructure that keeps long-running agents working.

Baseten, led by co-founder and CEO Tuhin Srivastava (@tuhinone), announced on September 10th that it is acquiring Blaxel, bringing Blaxel's execution infrastructure for AI agents into an inference business valued at $13 billion in June.

The proposed combination connects two technical layers that developers often buy separately. Baseten serves and trains models. Blaxel gives agents isolated computers, persistent files and networking so they can execute code, call tools, wait and resume work without losing state. Baseten plans to add products built on Blaxel's technology, beginning with its Sandboxes, according to the acquisition announcement.

Financial terms were not disclosed. Baseten also did not specify whether the acquisition involved cash, stock or a combination of the two. Blaxel will continue operating, and its existing product will remain available while Baseten begins the integration.

The acquisition would turn inference from Baseten's destination into one component of a broader agent cloud. For Blaxel co-founder and CEO Paul Sinai, it would place the missing model-serving layer inside the same organization as the sandboxes, storage and networking that his six-founder team has built since founding Blaxel in 2024.

Two founding teams, one recurring infrastructure problem

Baseten and Blaxel arrived at the deal after solving different versions of the same operational headache.

Srivastava, Amir Haghighat, Phil Howes and Pankaj Gupta founded Baseten in 2019 after working together at Gumroad, according to Haghighat's account of the company's origins and TechCrunch's account of Baseten's early funding and founding story. Srivastava and Howes had trained a fraud classifier there, while Haghighat ran engineering. Training the model was manageable. Keeping it fast and reliable in production required a much larger set of systems.

That experience became Baseten's original thesis: model serving would become a bottleneck as machine learning moved into customer-facing products. In a Q&A about Baseten's founding, Haghighat recalled that people initially called the inference focus "stupid," because a model had to be trained before it could be served. The bet improved with age as pretrained and open-weight models allowed developers to begin with deployment and customize later.

RuntimeWire examined Baseten's inference economics earlier in September, including the trade-offs among latency, throughput, model quality and the infrastructure bill. Blaxel extends those economics beyond each model call. An agent may make many calls while executing code, reading files and contacting tools, so network distance and idle compute costs accumulate across an entire task.

Sinai, Christophe Ploujoux, Charles Drappier, Nicolas Lecomte, Thomas Crochet and Mathis Joffre formed Blaxel after working together at ForePaaS, the analytics platform acquired by OVHcloud in 2022. The unusually large founding group had already built and sold infrastructure together. In a July 17th, 2025 interview with VentureBeat, Sinai said conventional cloud providers were designed for the software-as-a-service era, while agents required their own type of infrastructure.

Blaxel entered Y Combinator's Spring 2025 batch and raised a $7.3 million seed round led by First Round Capital that July. Y Combinator, Liquid2 and Transpose participated. At the time, Blaxel said it was processing millions of daily agent requests across 16 regions. Those usage figures were self-reported, although they gave investors an early indication that the runtime problem existed outside demos.

Baseten moves to buy the parts around the model call

Blaxel's core product is a micro-virtual-machine sandbox for each agent. Blaxel says those environments can suspend and resume in roughly 25 milliseconds, preserve memory state and scale down during inactivity. Agent Drive stores files, code and working context across sessions, while Blaxel's networking layer connects agents to APIs, tools, Model Context Protocol servers and other agents.

In its announcement thread, Blaxel argued that inference must sit beside compute, storage and networking because latency compounds each time an agent crosses a network boundary. That is the technical rationale for ownership rather than a loose partnership: the combined company could optimize the path between an agent deciding what to do, receiving model output and taking the next action.

Baseten identified Sandboxes as the first Blaxel primitive it will bring onto its platform and gave no schedule for Agent Drive or the networking layer.

Blaxel's product will continue separately in the meantime. That leaves existing users on Blaxel's current product while Baseten begins the integration.

A $13 billion inference provider moves outward

Baseten raised a $1.5 billion Series F at a $13 billion valuation on June 22nd. The company's Series F announcement says Altimeter Capital, Conviction Partners and Spark Capital led the round, with Sands Capital and Wellington Management as co-leads. Baseten said at the time that revenue had increased 20-fold and inference volume 40-fold over the prior year, without publishing revenue or annual recurring revenue.

The financing gave Baseten substantial room to expand around inference. The proposed Blaxel acquisition shows where Srivastava and his co-founders are spending some of that strategic capacity: on controlling a larger share of the infrastructure used during every agent loop.

The combination would put Baseten into closer competition with agent-compute providers including Modal, E2B and Daytona, while retaining its existing fight with inference platforms such as Fireworks AI, Together AI and Replicate. The cloud vendors remain the larger opponent. AWS, Microsoft and Google already sell compute, storage, networking and model access under one account, although their services were assembled for a wider range of workloads.

Sinai built Blaxel around the view that agents deserved a cloud designed around their stop-and-start behavior. Srivastava would gain a team that has built cloud infrastructure together through two companies and one previous acquisition.

The undisclosed purchase price would show how Baseten valued Blaxel's technology, customers and six-person founding group after a $7.3 million seed round. The product plan offers a clearer answer about the strategy. Baseten intends to own the model call and the machine doing the work around it.