Nebius acquires AI model optimization startup Eigen AI for $643M

Nebius Group NV, a Dutch operator of artificial intelligence data centers, today announced plans to buy software maker Eigen AI Inc. for $643 million.

The company will finance the acquisition with cash and stock. It expects to close the deal in a few weeks.

Nebius provides access to graphics card unit clusters that developers can use to train AI models and run inference workloads. It also offers a managed inference service, Token Factory, that removes the need to manage the underlying GPUs. Nebius will use Eigen AI’s technology to enhance the service.

Token Factory enables customers to perform inference using more than a dozen open-source AI models. Eigen AI’s software, in turn, optimizes open-source models to improve their speed and hardware efficiency.

Neural networks comprise code snippets called kernels. Eigen AI’s platform can replace some of the default kernels in an open-source model with custom modules that provide better performance. According to the company, those custom modules are implemented in CUDA and Triton. CUDA is the interface through which applications interact with Nvidia Corp. chips, while Triton is a Python-like programming language optimized for AI projects.

Eigen AI says that its software also optimizes other components of open-source AI models. It compresses their weights, the settings that determine how input data is processed, to lower memory requirements. Additionally, the platform enhances the KV cache in which language models store the information they use to answer prompts.

Nebius’ push to integrate Eigen AI into Token Factory will place particular emphasis on the former platform’s post-training features. Post-training is the process of fine-tuning a model that has already been trained to boost output quality. In many cases, developers go about the task by providing the model with examples of how to perform actions that it doesn’t support well out of the box. 

Post-training a model usually requires software teams to adjust many of its parameters, which requires a significant amount of infrastructure. Eigen AI’s platform uses a more efficient approach called LoRA. The technology works by extending neural networks with a small number of external parameters. During post-training, LoRA only recalibrates the new parameters, which is faster than reconfiguring a large subset of the model’s existing settings.

Eigen AI offers its core model optimization features alongside several complementary products. One offering, Eigen Data, speeds up the process of assembling training datasets. A cloud service called Eigen Inference helps developers lower their AI models’ latency.

Eigen AI’s employees will join Nebius after the deal closes. According to the data center operator, the team will help it establish a new engineering hub in the Bay Area.

Photot: Nebius