The Next Chapter of Artificial Efficient Intelligence

Eigen AI has agreed to join Nebius.

When we founded Eigen AI, we named our mission Artificial Efficient Intelligence — the pursuit of the world's most efficient engines for generating intelligence. We believed then, and we believe now, that the next era of AI will be defined not only by how capable models become, but by how efficiently we can train, serve, and run them. Frontier intelligence is only as valuable as it is accessible, and accessibility is an engineering problem that spans every layer of the stack — from the model down to the kernel.

Today, we're taking the biggest step yet toward making that mission real: Eigen AI has agreed to join Nebius (NASDAQ: NBIS). Together, we aim to build the world's most performant AI cloud — uniting Eigen's full-stack model and inference software with Nebius's global hardware and infrastructure footprint, so any developer or enterprise on the planet can run the best models at the best price, with no capacity ceiling.

The Thesis: Efficiency, End to End

Generating intelligence efficiently is not a single problem. It's three problems, stacked.

• At the model layer, we co-design the model itself for the way it will actually be served — quantization, pruning, MoE routing, KV-cache architecture, speculative decoding, and post-training recipes that preserve quality while reshaping the workload.
• At the system layer, we engineer the inference engine — schedulers, memory managers, continuous batching, prefill/decode disaggregation, and parallelism strategies (tensor, pipeline, expert) that turn a fleet of GPUs into a single coherent inference engine.
• At the kernel layer, we hand-tune the lowest-level GPU code — fused operators, custom attention kernels, and low-bit matmuls — extracting every cycle the hardware can give.

Pulling on all three at once is what produces results that pulling on any one of them alone cannot. And the receipts are in:

• #1 on Artificial Analysis for inference speed across multiple pure-text, reasoning, and multimodal models — including DeepSeek, GPT-OSS, and Qwen.
• State-of-the-art RL post-training, with our customized open models surpassing leading closed systems including GPT-5 and Claude Opus 4.5 on agentic benchmarks like τ-Bench, WorkArena, and CureBench.
• A team that lives at this frontier: PhDs from MIT, Stanford, Purdue, University of Toronto, Duke, Virginia Tech, and so on, researchers and engineers from Google DeepMind, Meta TBD Labs/MSL, OpenAI, and Amazon AGI. Our roots trace back to MIT HAN Lab — home to research the entire industry runs on, including techniques like AWQ that have become the standard for low-bit model serving.

Software-only optimization, however, runs into a hard ceiling. To deliver Artificial Efficient Intelligence to the world — at the scale the world actually wants — you need the metal underneath.

Software, Meet Silicon

Nebius is one of the most thoughtfully engineered AI clouds on the planet. State-of-the-art GPU clusters. A growing global data-center footprint. Managed Kubernetes built specifically for AI workloads. They have spent years going deep on the hardware and infrastructure layers in exactly the disciplined, AI-native way the next generation of compute requires.

Where Eigen has gone deep on the software stack — model, system, and kernel — Nebius has gone deep on infrastructure and platform. The two skill sets reinforce rather than duplicate, forming the full vertical stack that powers Nebius Token Factory for deploying, customizing, and serving frontier open-source intelligence:

Infrastructure → Kernel → System → Model → Platform

This is the stack that lets us optimize across boundaries most companies can't even cross — co-tuning custom kernels with cluster scheduling, fusing model architecture with hardware topology, treating the GPU and the inference graph as one system instead of two.

Why this Works: a Partnership before It was a Deal

We didn't arrive at this conclusion in the abstract. Eigen and Nebius have been technical partners for several months, and our jointly optimized open-source endpoints have already ranked among the fastest in the world on Artificial Analysis. Every joint engineering sprint reinforced the same conclusion: 1 + 1 is meaningfully greater than 2.

The cultural fit is just as clean. In every conversation with Nebius team members, we found a shared conviction — that time is the scarcest resource, that software will be the defining differentiator of this datacenter build-out, and that the teams who own both the hardware and software layers will set the pace for the industry.

This is the right home for what we're building.

For the People Building with Us

If you're an Eigen AI customer or developer, the short version is: things get better, and nothing you depend on goes away (until closing, it is business as usual and Nebius and Eigen AI remain separate entities).

After the deal close, Eigen's optimization stack will be integrated directly into Nebius Token Factory — Nebius's managed inference platform, with autoscaling endpoints and fine-tuning pipelines across all major open-source models including Qwen, Llama, DeepSeek, GLM, Kimi, MiniMax, and Nemotron.

Concretely, this means:

• Continuity — your endpoints, fine-tuned models, and integrations all keep working. The same engineers and researchers building your stack today will keep building it tomorrow.
• No more capacity ceiling — run inference and post-training at any scale on Nebius's global compute footprint.
• More models, faster — new open-source releases will land on optimized endpoints sooner, with the day-one performance and unit economics you expect from us.
• Expanded post-training and fine-tuning — the same RL post-training capabilities that beat closed-source models on agentic benchmarks, now available directly on the platform.

The entire Eigen team — founding members, researchers, engineers — is joining Nebius in full, and we'll be establishing Nebius's engineering and research presence in San Francisco Bay Area.

What We Want to Build Next

There's a thing we've been talking about inside Eigen for a long time: an automatic optimization software layer that sits between applications and infrastructure, so that deploying world-class AI is as simple as writing an API call. Every enterprise, every developer — accessing the best possible model performance without owning the underlying complexity. With Nebius's compute behind us and Token Factory as the platform, we can actually build it.

That's where this is going. The mission doesn't change. The team doesn't change. What changes is the leverage behind both.

With Gratitude

Companies are built by people, so before anything else — thank you.

To our customers and developers — you trusted a young team with workloads that mattered to your business. That trust is the entire reason we exist, and we are not finished serving you. We are finally equipped to serve you the way we always wanted to.

To the Eigen AI team — every one of you could be working anywhere in the industry. You chose to build something hard, together, in a market that does not forgive shortcuts. I could not be more proud, or more grateful.

To the investors who backed Eigen when it was an idea on a whiteboard:

• Mark Weber and Matt Rhodes-Kropf at Tectonic Ventures
• Shin Chen, Habib Haddad, and Calvin Chin at E14 Fund
• Salil Deshpande at Uncorrelated Ventures
• Rocky Yu at AGI House Ventures
• And every angel investor who took an early bet on us — we are deeply thankful.

To our advisors, mentors, friends, and supporters — the calls taken at odd hours, the introductions made, the candid feedback, the quiet encouragement when it mattered most. So much of what we built rests on shoulders we don't always name in public. Thank you for being in our corner.

To the entire Nebius team — thank you for the conviction, the speed, and the partnership. We can't wait to build alongside you.

Onward.

— Ryan Hanrui Wang, Wei-Chen Wang, Di Jin on behalf of Eigen AI Team