Nebius Strengthens Its Token Factory with Eigen AI and Clarifai

In May 2026, Nebius made significant moves to strengthen the value proposition of Token Factory, its managed inference platform. The company acquired independent inference provider Eigen AI for approximately US$643 million, following a collaboration earlier in 2026 that produced leading Artificial Analysis benchmark results across multiple models.

Shortly afterward, Nebius announced that Clarifai’s core engineering and research team would join the company. Clarifai founder and Chief Executive Officer (CEO) Matthew Zeiler will join Nebius as Senior Vice President of Research, overseeing work in areas including multimodal agentic reasoning, world models, token efficiency, and long-term memory. Nebius has also agreed to license Clarifai’s inference and compute orchestration technology, although the commercial terms of the agreement were undisclosed at the time of writing.

Although both moves support inference optimization, they strengthen different parts of the Token Factory stack: Eigen AI adds model-level optimization expertise, while Clarifai contributes compute orchestration technology and research talent.

As production Artificial Intelligence (AI) workloads scale, and open-source models gain traction on cost advantages, inference optimization is quickly becoming a technical and strategic priority. Nebius’ moves reflect several market dynamics that are beginning to shape the competitive positioning across the AI cloud market:

• Inference talent is emerging as a strategic differentiator. The acquisition and retention of specialized inference talent is becoming a major competitive battleground in AI infrastructure, particularly as agentic systems increase the need for optimization across the full stack. Especially as open-source AI becomes adopted more widely within enterprises, the differentiation opportunity for cloud providers is how effectively the models run on top of their stack.
• Managed inference remains an important enabler of enterprise adoption. Large-scale inference deployments remain difficult to optimize and operate, creating demand for managed platforms. This is especially relevant as open-source models continue to narrow the performance gap with frontier models while offering lower-cost economics.
• Latency and efficiency are becoming central to competitive positioning. Agentic workloads place greater emphasis on low-latency inference, token efficiency, and orchestration. Providers that continuously optimize for these factors are better positioned to attract startups, AI-native developers, and eventually traditional enterprises. This requires a combination of both model-level and compute-level optimization, which is why the acquisitions of both Clarifai and Eigen AI make sense for Nebius.
• Utilization matters as much as capacity. Most Graphics Processing Unit (GPU) clusters are not fully utilized due to the spiky nature of most predominant workloads; for example, coding-associated token traffic. In a market still shaped by infrastructure constraints, improving GPU utilization at scale can help Nebius serve more demand through Token Factory and improve monetization. Bare-metal GPU access may still suit some training workloads, but inference consumption remains mostly via serverless Application Programming Interface (API)-based delivery.

Nebius’ strategy suggests it is moving closer to a more integrated, hyperscaler-like model as the inference market matures. While some neoclouds still focus on bare-metal infrastructure and allow third-party inference providers to resell underlying capacity, Nebius is seeking greater control over the inference stack. Although a leaner service approach has helped neoclouds move quickly, deeper platform integration will likely be necessary for them to become credible long-term challengers to incumbents such as Amazon Web Services (AWS) and Microsoft Azure.

Moving forward, Nebius should use its newly acquired inference talent to not only bolster its service offerings, but to continue strengthening its credibility in the developer community. Publishing research papers and demonstrating benchmark leadership can help convert engineering depth into ecosystem growth. Furthermore, Nebius should consider Forward Deployed Engineers (FDEs) as a core go-to-market capability to accelerate how quickly enterprises scale their production workload. The newly acquired engineering talent can also be leveraged here as part of the inference optimization work they will be undertaking.

For the wider AI cloud market, it is becoming clear that as open-source AI adoption scales, communicating clearly to developers and customers how inference is served within a specific Cloud Service Provider (CSP) is becoming more important. Communicating performance capabilities across latency, throughput, and reliability is critical for sustained customer traction. ABI Research anticipates that more activity will be seen from neoclouds in acquiring inference optimization players, while hyperscalers may prefer to simply provide access to these inference players optimized models via their existing channels. For example, Fireworks AI is available on Microsoft Foundry as a first-party inference provider.