Mistral AI SAS today announced that it has acquired Koyeb SAS, a startup with a cloud platform built to run artificial intelligence workloads.

The deal’s financial terms were not disclosed.

Paris-based Mistral develops an eponymous series of open-source large language models. It also has several paid offerings, including a cloud platform called Mistral Compute. The company received a 11.7 billion-euro valuation last September after raising a 1.7 billion-euro round led by ASML Holding NV.

Koyeb, for its part, raised $7 million in 2023 from a consortium that included Samsung Next and other prominent backers. Its namesake cloud platform enables developers to run AI workloads without having to maintain the underlying hardware.

The company says that deploying an application on its infrastructure takes a few minutes. From there, an autoscaling feature automatically adjusts the amount of hardware allocated to the application as demand changes. If traffic goes down to zero, Koyeb deactivates the workload until it’s needed again.

Completely shutting down idle applications is often technically challenging for developers. One of the main reasons is that restarting an inactive workload can be fairly time-consuming, which creates delays for users. According to Koyeb, its platform addresses the challenge with mechanisms that enable it to relaunch workloads in under 200 milliseconds.

The company runs customer workloads in instances that each feature up to eight graphics processing units. Those instances are spread across 10 data centers located in different parts of the world. If one of the facilities goes offline, Koyeb can reroute traffic to another site.

One of the latest additions to the platform’s feature set is a capability called Koyeb Sandboxes. It creates isolated environments in which AI agents can run without creating cybersecurity risks. According to Koyeb, the capability enables applications to launch up to thousands of agents per day for tasks such as generating code.

Mistral will use the company’s technology to enhance a similar sandboxing feature it provides to users of its code generation models. Koyeb will also help improve Mistral Compute, a cloud platform designed to run AI models. The platform is powered by water-cooled servers equipped with Nvidia Corp. graphics cards. Alongside infrastructure, it provides access to software tools that ease tasks such as managing AI workloads’ encryption keys.

Mistral also plans to put Koyeb’s technology to use in other areas. According to the company, the acquisition will help it use GPUs more efficiently and improve its LLMs’ ability to run on customers’ on-premises infrastructure. The 16-person Koyeb team will join Mistral in March to support the engineering effort. 

Image: Mistral