Nebius launches Token Factory to challenge Microsoft, AWS, and Google in the AI cloud race

Artificial Intelligence

AI cloud provider Nebius has unveiled a new platform called Token Factory, designed to run and manage the world’s leading open-source AI models at scale. Positioned as a direct alternative to Microsoft Azure, Amazon Web Services (AWS), and Google Cloud Platform (GCP), Token Factory aims to give enterprises more flexibility, performance, and control when deploying open or custom large language models.

The new platform supports more than 60 open-source models, including DeepSeek, OpenAI’s GPT-OSS, Meta’s Llama, Nvidia’s Nemotron, and Alibaba’s Qwen. Nebius says enterprises can use Token Factory to host existing open models or deploy their own proprietary AI models on its infrastructure. Existing Nebius AI customers will be automatically upgraded to the new platform.

According to the company, Token Factory is built for efficiency and reliability, offering sub-second latency, autoscaling throughput, and 99.9% uptime, even for workloads that exceed hundreds of millions of requests per minute. This level of scalability puts Nebius in direct competition with established hyperscalers while also challenging new entrants such as Fireworks and Baseten, which provide similar AI hosting and orchestration solutions.

Nebius was originally part of Yandex but became an independent company in 2024 following international sanctions on the Russian search giant. Since then, it has expanded rapidly, operating data centres in the United States, Europe, and Israel, as reported by Bloomberg.

CEO and Chief Business Officer Roman Chernin says Nebius is not simply chasing higher margins by offering software-based AI services but is focused on becoming a broader enterprise provider. “Simply having infrastructure is far from enough,” Chernin said as per the report. “We want to become a large enterprise, but not merely a utility company. What we built is a scalable, reliable platform that allows clients to seamlessly switch between models and scale as they need.”