Nebius has purchased Eigen AI, a company with 20 employees, for $643 million, as inference optimization emerges as the most crucial component of AI infrastructure.

      TL;DRNebius, the Dutch neocloud that separated from Yandex in 2024, has agreed to purchase Eigen AI for $643 million, valuing the 20-person MIT-alumni startup at around $32 million per employee. Eigen’s technology optimizes inference, enhancing tokens produced per Nvidia GPU, a crucial capacity in AI infrastructure. This acquisition enhances Nebius's Token Factory inference platform as the neocloud market experiences rapid growth with CoreWeave and FluidStack securing billions in funding.

      Nebius Group, a Dutch cloud computing firm that split from the Russian internet provider Yandex in 2024, has reached an agreement to acquire Eigen AI for about $643 million in stock and cash. Announced on May 1, the deal targets a 20-person startup founded by alumni from MIT’s HAN Lab. In a landscape where leading AI firms are valued in the hundreds of billions and major acquisitions often involve vast teams of engineers, the $643 million for a small group prompts questions. The answer lies in inference. Eigen AI’s technology optimizes the number of tokens, the basic data units in large language models, generated by each Nvidia chip when processing AI models. “This is akin to the Olympic sport of today's market: who can maximize token output for the same cost?” remarked Roman Chernin, Nebius co-founder and chief business officer. He described the Eigen team as “like Olympic sprinters in this field.” This discipline, it turns out, is valued at $32 million per individual.

      The economics

      The priciest challenge facing the AI industry is not model training but model execution. Training a cutting-edge model incurs a one-time capital investment, in the hundreds of millions, to establish a set of weights. Inference, the process of utilizing those weights to generate user responses, constitutes a recurring operational expense that scales with every query, API call, and token generated. For companies providing AI as a service, inference is the primary cost driver. Each percentage point of efficiency improvement in inference, every extra token obtained from the same Nvidia GPU, leads to reduced expenses or enhanced profit margins. Eigen AI specializes in precisely this: enhancing the performance of open-source models from OpenAI, Alibaba, Meta, and Nvidia, allowing each chip to produce more output for the same input of electricity and silicon.

      Eigen AI’s founders gained recognition in the field through activation-aware weight quantization, a technique for compressing AI models from high-precision to lower-precision numerical formats without substantial loss of output quality. Co-founder Wei-Chen Wang was awarded the MLSys 2024 Best Paper Award for this achievement. In practical terms, quantization enables a model that typically requires four GPUs to function on two, or allows a single GPU to generate tokens twice as quickly. For a cloud provider like Nebius, which secured $700 million from Nvidia and Accel to expand its GPU fleet, deriving greater value from each chip alters the unit economics of the entire operation.

      The buyer

      The 💜 of EU tech: Recent updates from the EU tech arena, an account by our esteemed founder Boris, and some debatable AI artwork. It's complimentary, delivered weekly to your inbox. Subscribe today!Nebius holds a unique position in the AI infrastructure sector, being part of a category known as “neoclouds,” which provide AI computing power to enterprises without creating consumer products. While established hyperscalers like AWS, Microsoft Azure, and Google Cloud dominate the general cloud market, neoclouds have found a niche by offering AI-optimized infrastructure, with lower overhead costs and quicker deployment. Nebius has been tripling its Nvidia GPU capacity at its data center in Finland, deploying Nvidia’s H200 chips, and has launched a data center in Paris as part of a $1 billion investment plan for Europe. In November, it introduced Token Factory, a managed inference product that competes with startups like Fireworks and Baseten, as well as the inference solutions offered by hyperscalers.

      The acquisition of Eigen AI aims to position Token Factory as the most efficient inference platform on the market. With Eigen’s optimization integrated into Token Factory, Nebius can provide customers with lower per-token costs or increased output from the same hardware, giving it a competitive edge in a transparent pricing market with low switching costs. The neocloud sector is growing rapidly, as evidenced by companies like CoreWeave signing infrastructure contracts valued in the tens of billions. FluidStack, another neocloud, is seeking to raise $1 billion at an $18 billion valuation. The competitive landscape is clear: the goal is to provide the most tokens per dollar per GPU.

      The strategy

      The Eigen acquisition marks Nebius’s second purchase in three months, following its February acquisition of Tavily, an AI agent search company, for $275 million. Chernin mentioned that the firm is exploring additional acquisition opportunities. This trend indicates a strategy focused on acquiring small, highly skilled teams whose expertise would take years