Pomegra
Nebius Group N.V. (NASDAQ: NBIS), the Amsterdam-headquartered AI cloud company, announced on May 1, 2026 an agreement to acquire Eigen AI, a leading inference and model optimization startup, in a deal valued at approximately $643 million. The consideration will be paid through a combination of cash and Nebius Class A shares, calculated using the company's 30-day weighted average stock price at signing. The transaction is expected to close within weeks, pending standard antitrust clearance.
- Nebius (NBIS) will pay approximately $643 million in cash and Class A shares to acquire inference startup Eigen AI.
- NBIS stock surged over 11% on the announcement, reflecting strong investor conviction in the inference infrastructure play.
- Eigen AI's MIT HAN Lab founders — behind the industry-standard AWQ quantization and SpAtten sparse attention techniques — will anchor a new Nebius engineering hub in the San Francisco Bay Area.
A Strategic Bet on the Inference Layer
The acquisition centers on Nebius's flagship managed inference platform, Token Factory, which provides enterprise-grade autoscaling endpoints and fine-tuning pipelines across all major open-source models. By integrating Eigen AI's full-stack optimization technology directly into Token Factory, Nebius removes the single most complex bottleneck in production AI deployment — the gap between raw compute capacity and efficient real-world inference performance.
Eigen AI has developed a comprehensive model optimization stack spanning the entire AI lifecycle: from post-training quantization and fine-tuning to GPU-level kernel execution and real-time workload scheduling. Its technology addresses the hardest challenges in modern inference architectures, including Mixture-of-Experts (MoE) routing, Compressed Sparse Attention (CSA), long-context memory management, and compute efficiency across models including DeepSeek, Llama, Qwen, Gemma, Nemotron, and Kimi.
Prior to the acquisition, the two companies already collaborated to deliver jointly optimized model implementations that ranked among the fastest on Artificial Analysis benchmarks, with throughput reaching up to 911 tokens per second on select models — a preview of the combined platform's capabilities at scale.
MIT Pedigree Behind the Technology
Eigen AI was founded in 2025 by three researchers from MIT's HAN Lab, one of the world's foremost centers for AI computing and model efficiency. Co-founder Ryan Hanrui Wang authored the SpAtten (Sparse Attention) paper — the most-cited HPCA publication since 2020 — a foundational technique underpinning how modern large language models operate. Co-founder Wei-Chen Wang received the MLSys 2024 Best Paper Award for Activation-aware Weight Quantization (AWQ), now the de facto standard for 4-bit model serving in production deployments globally.
The third co-founder, Di Jin, holds a PhD from MIT CSAIL and contributed directly to the post-training of Meta's Llama 3 and Llama 4, while co-authoring the CGPO reinforcement learning from human feedback (RLHF) framework. Together, the founding team represents a concentration of elite inference research talent rarely found in a company of just 20 people — making the $643 million price tag, at roughly $32 million per employee, a deliberate acqui-hire at the frontier of the field.
Following close, the Eigen AI team will establish Nebius's first Bay Area engineering and research presence, anchoring the company's growing footprint in the United States.
Inference: The Fastest-Growing AI Segment
The deal arrives as AI inference overtakes training as the primary driver of compute demand. Inference — the process of running a trained model to generate outputs — is now forecast to account for approximately two-thirds of total AI compute consumption in 2026, driven by the explosion of production deployments across enterprise, developer, and consumer applications.
Roman Chernin, co-founder and Chief Business Officer of Nebius, framed the rationale sharply: "We are operating in a capacity-scarcity world where AI builders need optimized inference and infrastructure scale. The integration of Eigen AI's optimization capabilities and founding team will establish Nebius Token Factory at the frontier of inference, offering customers market-leading model performance and unit economics with massive compute capacity to back it at scale."
Ryan Hanrui Wang, co-founder and CEO of Eigen AI, added: "Together, we are removing the friction of AI model customization and deployment so developers can run models reliably in production without managing the underlying infrastructure."
NBIS Shares Jump 11% on Deal News
Markets responded with conviction. NBIS shares surged more than 11% on the announcement day, building on a broader rally that had already lifted the stock to $41.69 — a gain of approximately 7.89% in recent sessions. The deal is Nebius's second acquisition in three months, following its $275 million purchase of Tavily in February 2026, which added agentic search capabilities to the platform.
The Eigen acquisition fits within a dramatically larger strategic buildout. Nebius entered 2026 with $3.68 billion in cash and cash equivalents as of December 31, 2025, backed by a $2 billion investment from Nvidia in March 2026, a $27 billion multi-year cloud capacity agreement with Meta, and a $17.4 billion GPU infrastructure deal with Microsoft. The company is targeting over 5 gigawatts of hyperscale AI cloud capacity by 2030.
Outlook: Separating Scale from Optimization
The Nebius-Eigen AI combination represents an increasingly important competitive axis in AI infrastructure — the convergence of raw GPU scale with deep software-level optimization. While competitors such as CoreWeave have focused primarily on expanding compute capacity, Nebius is positioning Token Factory as a full-stack inference solution where hardware efficiency is built into the platform, not engineered separately by customers.
Existing Eigen AI customers will gain immediate access to Nebius's global infrastructure upon close. For enterprise operators deploying open-source models at scale, the integrated platform promises faster time-to-production, materially improved unit economics, and accelerated adoption of new model architectures — without the substantial in-house engineering overhead that currently limits most organizations' ability to run frontier models efficiently.
Mentioned Tickers: NBIS
