Nebius (NBIS) strikes $643m Eigen AI acquisition to anchor inference at Token Factory

Nebius Group N.V. (NASDAQ: NBIS) has agreed to acquire Eigen AI, a specialist inference and model optimization company whose founding team is rooted in MIT’s HAN Lab, in a cash-and-stock transaction valued at approximately $643 million. The acquisition is structured around Nebius’s 30-day weighted average share price at signing and is expected to close within weeks, subject to antitrust clearance and customary conditions. The deal lands as Nebius trades around $155, near a 52-week high of $168.71 and roughly six times higher than its 52-week low of $23.25, giving Nebius an unusually strong equity currency to fund growth. For Nebius Token Factory, the company’s managed inference platform, the acquisition supplies the missing optimization layer that determines whether enterprise customers stay or migrate to a cheaper alternative.

What does the Eigen AI acquisition mean for Nebius Token Factory and its managed inference roadmap?

Nebius Token Factory has so far been positioned as an enterprise endpoint service that runs autoscaling open-source models with fine-tuning pipelines. The platform competes against Together AI, Fireworks AI, Anyscale, Groq, and the inference offerings of AWS Bedrock, Microsoft Azure AI Foundry, and Google Vertex AI. In that field, raw compute is rapidly becoming undifferentiated. The differentiator is how many tokens per second a customer extracts from a given GPU, and at what cost. That is precisely the layer Eigen AI specialises in. By integrating Eigen AI’s post-training and inference optimization stack directly into Token Factory, Nebius is moving from a capacity-led pitch to a unit-economics pitch.

The strategic intent is clear. Nebius already has the contracted compute. The company’s backlog is anchored by a multi-year revenue agreement with Microsoft worth approximately $17 billion and a separate Meta deal that has been reported in the $27 billion range. What it lacks today is a defensible reason for a developer choosing between three or four neoclouds to pick Nebius over Together AI or Fireworks. A quantization and serving stack from a team that wrote the standard 4-bit quantization technique used in production deployments is exactly that reason. The execution risk sits in integration speed. Optimization libraries are notoriously sensitive to hardware generations, and Nebius is fielding NVIDIA HGX B300 clusters alongside earlier Hopper inventory. Migrating Eigen AI’s stack to run cleanly across that mix without breaking customer workloads is the first real test.

Why is the Eigen AI founding team a more important asset than the technology itself?

Acquisitions in this corner of AI infrastructure are increasingly acqui-hires dressed as technology deals, and the pattern fits here. Eigen AI’s three named co-founders carry research credentials that matter for hiring downstream. Ryan Hanrui Wang authored the Sparse Attention work that has become the most-cited HPCA paper since 2020. Wei-Chen Wang received the MLSys 2024 Best Paper Award for Activation-aware Weight Quantization, the technique now serving as the default 4-bit production quantization standard. Di Jin, an MIT CSAIL graduate, contributed to Meta’s Llama 3 and Llama 4 post-training pipelines and co-authored the CGPO RLHF framework. That trio gives Nebius a credible recruiting magnet in the San Francisco Bay Area, where the company will now establish a formal engineering and research presence.

The hiring leverage matters because the inference talent pool is genuinely thin. The cohort that understands MoE routing, CSA memory layout, kernel-level scheduling, and reasoning-model serving across long contexts numbers in the low hundreds globally. Most of that cohort is concentrated at OpenAI, Anthropic, Google DeepMind, NVIDIA, and a small set of well-funded startups. Nebius arriving in the Bay Area without an anchor team would have meant another two years of ramp. With the Eigen AI core in place, Nebius can credibly recruit against Together AI and Fireworks for the same engineers who would otherwise never consider an Amsterdam-headquartered employer.

How does this acquisition reshape competition with Together AI, Fireworks, and the hyperscaler inference platforms?

Together AI and Fireworks AI have built their commercial proposition around exactly this combination of optimized open-source serving and developer-friendly endpoints. Both have raised aggressively, and both have recruited inference researchers as a core differentiation strategy. Nebius has historically competed on capacity rather than software depth. The Eigen AI integration narrows that software gap meaningfully. For customers running GPT-OSS, Gemma, Qwen, Llama, Nemotron, DeepSeek, GLM, Kimi, and MiniMax in production, the question increasingly becomes which provider extracts the most tokens per dollar from a given model. Eigen AI’s stack already covers all of those families, which is unusual breadth for a startup of its size.

The hyperscalers face a different threat. AWS Bedrock and Azure AI Foundry rely on a mix of in-house and partner-supplied inference engines. Their advantage has always been distribution and bundling rather than raw inference efficiency. If Nebius can deliver materially lower cost per token on the same models the hyperscalers offer, customers running large-scale inference workloads have a credible reason to peel those workloads off the hyperscalers entirely. That is the same playbook CoreWeave executed for training, and it is the reason CoreWeave and Nebius are increasingly grouped together as the neocloud category that institutional investors now treat as a structural threat to hyperscaler margins.

What does the $643 million price tag signal about valuation discipline at Nebius?

The deal consideration is paid in a combination of cash and Class A shares, with the equity portion priced off Nebius’s 30-day weighted average. Given where NBIS shares have traded over the past month, between roughly $130 and $168, that mechanism gives Nebius the ability to fund a meaningful chunk of the transaction in stock that the market is currently valuing at approximately 64 times sales. Using expensive equity to acquire scarce talent and proprietary optimization techniques is a textbook capital allocation choice, and it is the same playbook that allowed Microsoft, Meta, and Google to absorb critical AI startups in 2014 to 2018 when their own equity was trading at premium multiples.

The capital allocation question is whether Eigen AI is genuinely worth $643 million on a fundamental basis or whether Nebius is paying a strategic premium to keep the team away from competitors. Both can be true. Inference efficiency improvements of even 15 to 25 percent across a contracted backlog approaching $50 billion translate into hundreds of millions of dollars in gross margin over the contract lives. Set against that arithmetic, $643 million for the team that authored AWQ and SpAtten is defensible. The risk is that competitors close the gap quickly, since open-source quantization research diffuses faster than almost any other branch of AI work.

How should investors read the market reaction and the broader Nebius capital strategy?

NBIS shares have run from a 52-week low of $23.25 to a recent high of $168.71, a move of more than 600 percent over twelve months, and currently trade around $155. That trajectory reflects the contracted revenue backlog, the NVIDIA equity stake, and the Microsoft and Meta agreements rather than near-term profitability. Nebius reported a Q4 2025 earnings miss and is forecast to remain unprofitable through the current capital expenditure cycle. The company closed a $4.34 billion convertible debt offering in April to fund 2026 capex of $16 to $20 billion, a figure that places Nebius among the largest non-hyperscaler infrastructure spenders globally.

The Eigen AI acquisition does not change that capital-intensive trajectory. What it does is improve the gross margin profile of the revenue Nebius is committing all that capex to capture. A capacity-only neocloud with no optimization moat earns the spread between GPU cost and lease price, which compresses as competition intensifies. A capacity-plus-optimization platform earns a software-style margin layer on top, which is far more defensible. Wolfe Research’s recent initiation with a $80 to $170 fair value range captures both ends of that distribution. The Eigen AI deal pushes Nebius toward the upper end if execution holds, and the Q1 2026 earnings release scheduled for May 13 will be the first opportunity for management to outline integration milestones publicly.

What execution risks could undermine the strategic thesis behind the Eigen AI acquisition?

The first risk is integration drift. Optimization libraries that perform brilliantly in research environments often degrade when ported into multi-tenant production with strict SLAs. Token Factory customers expect autoscaling endpoints with predictable latency, and Eigen AI’s stack will need hardening across the full Nebius hardware fleet before it can be exposed at scale. The second risk is talent retention. Acquisitions of small, research-heavy teams routinely lose half their key engineers within 18 months when they collide with larger organisational structures. Nebius’s decision to anchor a new Bay Area office around the Eigen AI founders is a sensible mitigation, but it is not a complete one.

The third risk is open-source erosion. Quantization and sparse attention techniques have a habit of becoming commodity within 12 to 18 months of publication. AWQ is already widely implemented across vLLM, TensorRT-LLM, and SGLang. The proprietary edge Eigen AI brings is not the published research itself but the engineering depth around model-specific tuning, kernel optimization, and post-training pipelines that have not yet been replicated in open source. Nebius will need to keep that gap open through continued research output, which is why the founding team’s research credentials carry as much weight in the deal rationale as the existing product.

Key takeaways on what the Nebius and Eigen AI deal means for the company, its competitors, and the AI infrastructure industry

• The $643 million Eigen AI acquisition shifts Nebius Token Factory from a capacity-led to a unit-economics-led inference platform, narrowing the software gap with Together AI and Fireworks AI.
• Nebius is using its 600 percent equity rerating over the past twelve months as transaction currency, a textbook capital allocation move when stock trades at approximately 64 times sales.
• The Eigen AI founding team carries research credentials including the AWQ quantization standard and Sparse Attention work that give Nebius credible Bay Area recruiting leverage against Together AI, Fireworks, and the hyperscalers.
• Inference is forecast to account for roughly two-thirds of compute demand this year, which makes the optimization layer strategically more valuable than incremental capacity in the current cycle.
• Integration risk is concentrated in porting Eigen AI’s stack across mixed Hopper and Blackwell inventory while preserving Token Factory’s autoscaling SLAs for enterprise customers.
• Nebius’s contracted backlog approaching $50 billion through Microsoft and Meta deals gives the optimization investment a clear monetisation surface, with even modest efficiency gains translating into hundreds of millions in gross margin.
• The acquisition is a direct competitive response to Together AI and Fireworks AI, both of which have built their commercial proposition around optimized open-source inference rather than raw compute.
• AWS Bedrock, Azure AI Foundry, and Google Vertex AI face a sharper neocloud threat if Nebius can deliver materially lower cost per token on the same open-source models they bundle.
• Open-source diffusion of quantization techniques means Nebius must continue investing in proprietary research depth to keep the inference moat from collapsing into commodity within 18 months.
• The May 13 Q1 2026 earnings release will be the first public test of how management plans to translate the Eigen AI integration into measurable improvements in gross margin and customer retention metrics.