Nebius Acquires Eigen AI for Over $1B; Wang Hanrui to Lead Token Factory as SVP

The two numbers most easily remembered about this deal are the consideration of over $1 billion and the team size of about 20 people before the acquisition. Nebius announced the acquisition of Eigen AI for over $1 billion, and when Eigen AI had been formally established for about 9 months, it had about 20 employees. Exchanging a very small team for a very high consideration is the part of this news that has been repeatedly cited in circulation.

But it is necessary to separate "confirmed" from "the interviewee's own account." The acquisition itself is a confirmed fact; the employee size of about 20 is a claim provided by the interviewee, and the report gives no third-party verification. The consideration of "over $1 billion" also does not disclose the specific amount, payment method, or closing conditions, and the report likewise does not state whether the deal has been formally completed or whether regulatory approval is required.

Therefore, a more prudent reading of this deal is: it confirms that the capital market has placed a high price on the link of "inference efficiency," but the specific basis behind that pricing—how deep the technical moat is and how strong customer stickiness is—remains vague in the available information.

What brought Eigen AI into public view was Nvidia's GTC conference in March 2026. In a slide during his keynote, Jensen Huang placed Eigen AI first in the inference speed ranking for the Chinese open-source model Kimi K2.5. Wang Hanrui recalls that the company did not know in advance and only found out at the moment it was presented on stage.

This exposure occurred when Eigen AI had been formally established for about 9 months and less than two months before the acquisition. According to the interviewee, the exposure brought more customers who proactively reached out and boosted team morale.

But the evaluation criteria of the ranking are not clear. The report does not give the evaluation method, time window, or complete leaderboard for that ranking, so what can be confirmed about "ranked first" is that "it was presented that way on the slide," and it cannot be equated with a reproducible independent conclusion.

Wang Hanrui's path was to study as an undergraduate at Fudan University, enter MIT after graduating in 2018, become Professor Song Han's first doctoral student, and start a business after receiving his PhD in 2024. The sparse attention system SpAtten, which he led the development of, is called one of the most-cited HPCA papers since 2020, and its accompanying accelerator completed tape-out on TSMC's 28nm process. He himself was selected for the Asia-Pacific edition of MIT Technology Review's "35 Innovators Under 35."

After joining Nebius, he became a senior vice president, fully in charge of the Token Factory business, covering three product lines: inference, post-training, and agent systems. From scholar to entrepreneur to manager at a large company, he changed identities three times in just a few years.

His judgment about organizational form is pragmatic: do not start a business for the sake of starting a business; academia and industry are not a single-choice question, and the key is which organizational form can best bring technology into practice. His reasons for choosing Nebius include GPU resources, its emphasis on software, and the existing foundation of cooperation.

What needs to be flagged here is: the citation status of SpAtten and the tape-out details come from the interviewee's own account and are source claims; while being selected for "35 Innovators Under 35" is presented in the report as confirmed information.

Eigen AI's business can be summarized as: helping enterprises optimize and deploy open-source models, using enterprise data for post-training on one hand, and deploying models to GPU clusters and selling them externally by token on the other. The key to this model is that it bets on two curves at the same time: "model capability" and "unit cost."

Wang Hanrui's estimate of the weight in customer decisions is: cost accounts for about half, accuracy improvement accounts for 30%, and controllability accounts for 10% to 20%. He also mentions that customers do not care whether a model is open-source or closed-source, nor do they need the strongest, all-encompassing model; doing post-training for the customer's task so that the open-source model outperforms the closed-source model on the issues the customer cares about more is enough.

These proportions are personal estimates, not quantitative survey data. They can show the interviewee's judgment about where the market's center of gravity lies, but they cannot be treated as an industry benchmark.

Wang Hanrui's judgment is: the gap between open-source models and the most cutting-edge closed-source models has narrowed from 9 months to a year a year ago, and 6–9 months half a year ago, to possibly less than 6 months now. This "less than 6 months" is one of the most widely circulated judgments in the entire conversation.

Its value lies in providing a coordinate that can be tested later: if the gap continues to narrow, the business space for inference service providers will change accordingly; if the narrowing stalls, then the narrative of "using open source to replace closed source" needs to be repriced.

But it must be made clear that this is the interviewee's personal estimate; the report provides no quantitative basis and no third-party verification. It can be referenced as a trend direction, but treating it as an established fact goes beyond the scope of the evidence.

At the end of July, Nebius Token Factory became the Day 0 launch partner for the 2.8-trillion-parameter mixture-of-experts model Kimi K3. This is the most concrete product-side progress after the deal, and it also places Token Factory at the front of the release chain for cutting-edge open-source models.

But Wang Hanrui has a clear-eyed description of where this business sits: inference services are squeezed between GPU vendors and customers, in the middle of the smile curve, and are more vulnerable to pressure. This contrasts with his long-term goal of "making intelligence as cheap as water and air"—the cheaper it becomes, the more the profit space of the middle link must be supported by scale and efficiency.

Regarding the Kimi K3 cooperation, the report does not explain the actual scope of cooperation or exclusivity, so "Day 0 launch partner" can be confirmed, but the boundaries of its commercial meaning remain unclear.

Amid the wave of self-developed inference chips, Wang Hanrui says he is in no hurry to follow suit, because the underlying operators of models are still changing rapidly, and chips built now may no longer fit three years from now. This is a judgment based on the speed of technological evolution, not a rejection of the chip route.

On the direction of scale, he believes model scale growth is far from over, and another 1,000-fold increase is even possible; the endpoint of Token Factory is to make intelligence as cheap as water and air. He mentions that in the post-training and reinforcement learning stages, models can freely explore and self-evolve in environments, no longer constrained by data.

Both judgments are personal predictions and carry high uncertainty. They explain why Eigen AI chose "software + cloud" rather than "chips," but they cannot be read as a definitive conclusion about the industry's direction.