Nebius buys Eigen AI for $643M
- Nebius said on May 1 it will acquire Eigen AI for about $643 million, adding the startup’s inference software to its Token Factory platform. - The mix is unusually specific: about $98 million in cash plus 3.8 million Nebius shares, with Eigen’s founders joining to build a Bay Area hub. - It matters because AI cloud economics are shifting from raw GPU access toward software that squeezes more useful output from each chip.
AI cloud is turning into a software game, not just a GPU rental game. That is the real point of Nebius buying Eigen AI for about $643 million on May 1. Nebius already sells the hardware layer — big clusters of Nvidia chips, networking, capacity. What it was missing was a sharper answer to the question every AI customer asks next: how do I get more output, lower latency, and lower cost from the same machines? Eigen is basically that answer. (nebius.com) ### What does Eigen actually do? Eigen works on inference — the phase where a trained model is serving real users, generating tokens, answering prompts, and handling production traffic. Training gets the headlines, but inference is where many AI companies actually bleed money. Every extra milliseco(nebius.com) handles more work, faster. (nebius.com) ### Why is that worth so much? Because inference is becoming the recurring cost center. If you can make one GPU behave like 1.2 or 1.5 GPUs in real workloads, you do not just save money — you change pricing power. That is why a roughly 20-person startup could command a $643 million deal. Nebius sai(nebius.com)e, not a small acqui-hire. (sec.gov) ### Why Nebius in particular? Nebius is trying to become a full-stack AI cloud — not just a place to rent compute, but a place to build, tune, and serve models. Its Token Factory product is the managed inference layer in that stack. Buying Eigen plugs a missing piece directly into that product. Nebius had already partnered with Eigen last month(sec.gov)tory. The acquisition turns that partnership into vertical integration. (nebius.com) ### Why not just partner and stop there? Because the optimization layer is becoming too important to leave outside the company. If Nebius wants to compete with hyperscalers and other AI clouds, it needs control over the serving stack — scheduler behavior, kernel-level tweaks, model-specific optimiz(nebius.com)any boundaries. That usually matters most when customers start demanding predictable production performance, not just access to chips. (nebius.com) ### What else comes with the deal? The team. Nebius said Eigen’s founding group, including researchers from MIT HAN Lab, will establish Nebius’s Bay Area engineering and research presence. That matters more than it sounds. Nebius is based in Amsterdam, but AI infrastructure buying decisions and mod(nebius.com)e center of inference tooling. (nebius.com) ### Did investors like it? Yes — immediately. Nebius shares jumped after the announcement, with market coverage showing a strong positive reaction on May 1. Investors seem to read this as Nebius spending aggressively, but in the right place: software that can improve utilization of very expensive hardware. That is a cleaner story than just buying more GPUs and hoping demand stays hot. (insidermonkey.com) ### So what is the bigger shift here? The market is starting to treat inference optimization as a core layer of AI infrastructure. First came the scramble for chips. Now comes the scramble for efficiency — better throughput, lower latency, higher utilization, lower cost per token. In plain English, the gold rush is moving from owning shovels to making each shovel dig more dirt per hour. (nebius.com) The bottom line is simple. Nebius did not just buy a startup. It bought a way to make expensive AI hardware earn more. In this market, that may be one of the most valuable tricks there is.
