The headline number is $643 million. The detail that explains it is $32 million per employee.

Nebius Group, the Dutch AI cloud company listed on Nasdaq under the ticker NBIS, announced on May 1, 2026 that it had agreed to acquire Eigen AI, a 20‑person startup founded by alumni of MIT's HAN Lab, in a deal valued at approximately $643 million. The deal includes about $98 million in cash and 3.8 million Nebius Class A shares, priced on the company's 30‑day weighted average stock price. Nebius stock rose 6.8 percent intraday on the announcement.

Thirty‑two million dollars per employee is not irrationality. It is the market price for a specific capability that has become the most commercially valuable in AI infrastructure: the ability to make Nvidia GPUs generate more tokens per unit of time at lower cost per token. Nebius co‑founder and chief business officer Roman Chernin said exactly this when explaining the acquisition rationale: Eigen's technology maximizes the number of tokens generated by each Nvidia chip Nebius uses for inference.

Why Inference Efficiency Is the New Competitive Moat

The AI infrastructure market is undergoing a structural shift that Nebius described explicitly in its acquisition announcement: inference, the process of running AI models to generate outputs for users, is forecast to account for two‑thirds of total AI compute demand by the end of 2026. That means the ability to generate more tokens per GPU hour, at lower latency and cost, is not an optimization detail. It is the primary competitive variable that determines which AI cloud providers win enterprise contracts.

Eigen AI was built specifically to address this problem. The company's technical contributions, each representing genuine research advances that have been adopted across the industry, span the full inference optimization stack:

• Activation‑Aware Weight Quantization (AWQ), developed by co‑founder Wei‑Chen Wang, is a technique for compressing AI models from high‑precision numerical formats to 4‑bit precision without significant degradation in output quality. A model that required four GPUs to run can be served from two using AWQ, or can generate tokens twice as fast on the same hardware. Wang received the MLSys 2024 Best Paper Award for this work, and AWQ is now standard in production model serving across the industry.
• Sparse Attention optimization, developed by co‑founder Ryan Hanrui Wang, is the most‑cited HPCA paper since 2020. Sparse attention techniques reduce the computational cost of the attention mechanism in transformer models, which scales quadratically with sequence length. For long‑context applications, sparse attention is not an optional optimization. It is what makes the application viable.
• KV‑cache optimization and custom CUDA kernels round out the stack, with Di Jin, a third co‑founder, having contributed directly to Meta's Llama 3 and Llama 4 post‑training processes.

Eigen AI's optimization services cover the open‑source model ecosystem comprehensively: GPT‑OSS, Gemma, Qwen, Llama, Nemotron, DeepSeek, GLM, Kimi, and MiniMax are all supported. The company's technology specifically addresses production challenges in complex architectures including Mixture‑of‑Experts models and Compressed Sparse Attention variants, the architectures that are becoming increasingly dominant in frontier open‑source model releases.

What This Does for Nebius Token Factory

Following the acquisition, Eigen AI's optimization layers will be integrated directly into Nebius Token Factory, the company's managed inference platform that provides enterprise‑grade autoscaling endpoints and fine‑tuning pipelines across all major open‑source models. The integration means Token Factory customers will gain access to inference efficiency that Nebius has previously had to source from third‑party optimization libraries, and that other neocloud providers are still building internally.

Nebius's chief revenue officer Dimitry Shevelenko described the intended outcome in competitive terms: the combination will make Token Factory the most efficient inference platform on the market. That claim will be tested by the market, not by the press release. But the technical foundation Eigen AI brings, three co‑founders whose individual research contributions are already standards in production inference, provides the credibility to make the claim.

The acquisition also establishes Nebius's engineering and research presence in the San Francisco Bay Area for the first time, through Eigen AI's existing team. For a company headquartered in Amsterdam and listed on Nasdaq, having a Silicon Valley technical hub matters both for hiring and for customer proximity.

This is Nebius's second acquisition in three months, following the February 2026 deal to buy Tavily. Chernin confirmed the company is reviewing other acquisition opportunities, with a stated goal of becoming one of the key players in inference over the next 18 months. At a pace of one acquisition per quarter and a willingness to pay $32 million per engineer for the right talent, Nebius's infrastructure buildout is moving faster than its Nasdaq listing might suggest.

More at nebius.com