The economics of rented GPUs pushed Nscale to buy the software that controls them. On July 30 the London AI cloud operator, which builds its own power generation and data centers, announced it will acquire Anyscale, creator of the Ray distributed-computing framework. About 200 employees across the US, Europe and India transfer to Nscale, and the transaction is slated to finish before the end of the year.
Anyscale’s commercial product is the tooling ML teams rely on to spread one workload across thousands of accelerators. Data preparation, pre-training, fine-tuning, reinforcement learning and serving all flow through it. The Anyscale name survives, and the existing client list – Coinbase, Runway and Bedrock Robotics – stays on whatever infrastructure it already uses.
Nscale’s founder sees most rivals as simple GPU landlords. Josh Payne argues that renting identical Nvidia hardware by the hour leaves the real margin on the table, because profitability comes from how much productive work each hour yields. That is decided by scheduling, by how data preparation hands off to training, and by how fast accelerators return to service between reinforcement-learning cycles.
The acquisition rests on a large physical base. Nscale already operates beyond 1 GW, holds a Microsoft letter of intent for 1.35 GW at its Monarch campus in West Virginia centered on Vera Rubin NVL72 racks, and has earmarked $2.5B for UK data centers. A $900M credit facility from a dozen banks, including J.P. Morgan and Goldman Sachs, closed on July 7.
Ray sits under the PyTorch Foundation within the Linux Foundation, so the open-source project stays neutral while Nscale takes the commercial platform and engineering team. Anyscale’s roster spans Cursor, Physical Intelligence and xAI, and its CEO framed the union as the first hyperscaler to own the entire AI stack.
