Nebius acquires 20-person Eigen AI for $643 million to dominate AI inference optimization

Nebius Acquires Eigen AI to Accelerate AI Inference Capabilities

Nebius Group NV, the Dutch cloud computing company that split from Yandex in 2024, has agreed to acquire Eigen AI for approximately $643 million in stock and cash 2

2

3

3

. The deal, announced on May 1, targets a 20-person California startup cofounded by alumni of MIT's prominent HAN Lab 1

1

2

2

. The transaction represents Nebius's second acquisition in three months, following its February purchase of Tavily for $275 million 2

2

. Nebius expects to close the deal in a few weeks 3

3

.

Why AI Inference Solutions Command Premium Valuations

The $643 million price tag for a 20-person team reveals a fundamental shift in AI infrastructure economics. "This is like the Olympic sport of the current market: who can extract more tokens for the same price?" said Roman Chernin, Nebius co-founder and chief business officer 1

1

.

Eigen AI specializes in AI model optimization that maximizes the number of tokens—the basic units of data in large language models—generated by each Nvidia chip used for AI inference 1

1

. While training a frontier model is a one-time capital expenditure measured in hundreds of millions of dollars, inference represents a recurring operational cost that scales with every query and API call 2

2

. For companies selling AI as a service, every percentage point of efficiency gained in inference translates directly into lower costs or higher margins.

Optimizing Performance of Chips Through Advanced Quantisation Techniques

Eigen AI's core technology centers on activation-aware weight quantisation, a method for compressing AI models from higher-precision to lower-precision numerical formats without significant loss in output quality 2

2

. Co-founder Wei-Chen Wang received the MLSys 2024 Best Paper Award for this work 2

2

. The platform works on optimizing performance of some of the leading open-source models from OpenAI, Alibaba, Meta, and Nvidia 1

1

. In practice, quantisation allows a model that would normally require four GPUs to run on two, or enables a model running on one GPU to generate tokens twice as fast 2

2

. The software also optimizes other components by compressing weights to lower memory requirements and enhancing the KV cache where language models store information used to answer prompts 3

3

.

Strengthening Nebius Token Factory as a Frontier Inference Platform

Nebius unveiled Token Factory in November as a managed inference product competing with startups like Fireworks and Baseten, as well as cloud giants 1

1

. The acquisition of Eigen AI is intended to make Nebius Token Factory the most efficient frontier inference platform on the market 2

2

. With Eigen's optimization layer integrated, Nebius can offer customers lower per-token prices or higher throughput from the same hardware—a competitive advantage in a market where pricing is transparent and switching costs are low 2

2

. Token Factory enables customers to perform inference using more than a dozen open-source AI models 3

3

. The platform will place particular emphasis on Eigen AI's post-training features, which use LoRA technology to extend neural networks with a small number of external parameters, making the process faster than reconfiguring a large subset of existing settings 3

3

.

Neocloud Strategy and AI Infrastructure Competition

Nebius occupies a specific position as one of a group of companies called "neoclouds"—cloud providers that rent AI computing capacity to enterprises rather than building consumer products 2

2

. While established hyperscalers like AWS, Microsoft Azure, and Google Cloud dominate the overall cloud market, neoclouds have carved out a niche by offering AI-optimized infrastructure with lower overhead and faster deployment 2

2

. Nebius has been tripling its Nvidia GPU capacity at its data center in Finland, deploying Nvidia's H200 chips, and launched a data center in Paris as part of a $1 billion European investment plan 2

2

. The company raised $700 million from Nvidia and Accel to build out its GPU fleet 2

2

. Right now, with data center capacity in short supply, Nebius is reserving some computing power for Token Factory rather than selling it ahead of time to large clients in multiyear deals, allowing it to charge higher prices for short-notice contracts 1

1

.

Moving Up the Stack to Capture Customer Relationships

"We don't want to be the infrastructure and someone above us works with the real customers," Chernin explained 1

1

. This statement captures the neocloud dilemma: renting GPU capacity is profitable but commoditized, while margins improve closer to the application layer 2

2

. The pattern of acquisitions suggests a strategy of acquiring small, technically excellent teams whose capabilities would take years to build internally 2

2

. The company's goal is to become one of the key players in the inference market in the next 18 months 1

1

. Chernin said the company is looking at other deal opportunities, seeking to purchase companies with teams or capabilities that speed its planned strategy or add products and features closer to direct customer usage 1

1

. The acquisition also marks Nebius's strategic push into the US market, where demand for AI infrastructure continues to grow 4

4

. As the neocloud market expands rapidly—with competitors like CoreWeave signing infrastructure deals worth tens of billions and FluidStack in talks to raise $1 billion at an $18 billion valuation—the competitive dynamics are clear: whoever can offer the most tokens per dollar per GPU wins 2

2

.