# Nscale Buys Anyscale to Move Up the AI Compute Stack > Source: > Published: 2026-07-30 12:13:02+00:00 ### [ acquisitions ](https://www.unite.ai/series/acquisitions/) # Nscale Buys Anyscale to Move Up the AI Compute Stack [Add Unite.AI to your preferred sources on Google](https://www.google.com/preferences/source?q=unite.ai) Nscale, the AI cloud operator that builds its own data centers and generates its own power, has agreed to buy Anyscale, the company behind the Ray distributed-computing framework. The acquisition, [announced on July 30, 2026](https://www.nscale.com/press-releases/nscale-acquires-anyscale), puts a workload-orchestration layer on top of Nscale’s GPU fleet and moves Anyscale’s roughly 200 staff across the United States, Europe and India into the London-based company. It is expected to close in the second half of 2026, subject to closing conditions and regulatory approvals. Anyscale sells the managed platform that machine-learning teams use to spread one job over thousands of accelerators at once: multimodal data preparation, pre-training, fine-tuning, reinforcement learning and inference serving. It keeps its brand and its existing customers, which Nscale’s announcement lists as including Coinbase, Runway and Bedrock Robotics. Those customers stay free to run on whatever infrastructure they use now, with Nscale’s own platform becoming an additional option over time. ## Why a power company buys a scheduler Nscale’s pitch until now has rested on owning the layers below the accelerator: the generation, the campus, the racks. This purchase moves it up the stack, and the reason sits in how an AI cloud makes money. Raw GPU capacity rents by the hour against competitors buying the same Nvidia systems, so the spread comes from how much useful work each GPU hour produces. That is decided above the metal, in how a job is partitioned, how CPU-bound data preparation hands off to GPU training, and how accelerators get reclaimed between reinforcement-learning rollouts instead of sitting warm and idle. “Most infrastructure providers just buy GPUs and rent them,” said Josh Payne, CEO and founder of Nscale, who described the company’s approach as owning the power, the data centers, the compute and the software above them. That cost line is also Anyscale’s commercial argument. In its [June 2, 2026 launch of a native Azure integration](https://www.anyscale.com/press/anyscale-on-azure-cost-efficient-and-sovereign-ai-microsoft-build), the company said customers report up to four times faster experimentation and up to 90% lower total cost of ownership than fragmented stacks that pair cloud-native data-processing engines with hosted model APIs. Those are Anyscale’s figures, and they describe the layer a compute seller would rather own than route around. The same logic is showing up elsewhere in the compute business. Qualcomm ([QCOM](https://www.securities.io/nasdaq/QCOM/) ) [closed its all-stock acquisition of compiler startup Modular](https://www.unite.ai/qualcomm-closes-all-stock-acquisition-of-compiler-startup-modular/) in July 2026, on the reasoning that the software deciding how hardware gets used is where performance and switching costs both live. ## The capacity underneath the deal Nscale is adding platform revenue to a build-out whose largest tranches land later. The company put its operating capacity at over 1 GW in a [March 16, 2026 announcement](https://www.nscale.com/press-releases/nscale-west-virginia-ai-factory) covering a letter of intent with Microsoft ([MSFT](https://www.securities.io/nasdaq/MSFT/) ) for 1.35 GW of AI compute at its Monarch campus in Mason County, West Virginia, built around Nvidia’s ([NVDA](https://www.securities.io/nasdaq/NVDA/) ) Vera Rubin NVL72 systems. That deployment arrives in tranches starting in late 2027, powered by Caterpillar natural-gas generator sets scaling toward 2 GW of on-site generation by the first half of 2028. Nscale has also committed to [$2.5 billion of UK data-center investment](https://www.unite.ai/nscale-to-invest-2-5-billion-in-uk-data-centres-powering-generative-ai-and-government-ambitions/). Financing has kept pace with that pipeline. Nscale [closed a $900 million revolving credit facility](https://www.nscale.com/press-releases/revolving-credit-facility) on July 7, 2026, syndicated across a dozen banks including J.P. Morgan, Goldman Sachs, Morgan Stanley ([MS](https://www.securities.io/nyse/MS/) ) and MUFG. Software sells against capacity that is already running, and against capacity on other providers’ clouds, which is a different revenue shape from a lease that starts delivering in 2027. ## Where Ray sits now The open-source project at the center of the deal now sits under neutral governance. Ray moved to the PyTorch Foundation on [October 22, 2025](https://www.linuxfoundation.org/press/pytorch-foundation-welcomes-ray-to-deliver-a-unified-open-source-ai-compute-stack), joining PyTorch and the vLLM inference engine under the Linux Foundation, with 237 million downloads recorded at the time. Nscale says it will join the foundation as part of the acquisition. That structure defines what Nscale is buying: the commercial platform, the engineering team and the customer relationships, on top of a framework that stays community-governed and runs anywhere. Anyscale says Ray already carries workloads at Cursor, Physical Intelligence and xAI, and its Azure product runs inside a customer’s own Azure tenancy rather than on Anyscale-operated hardware. Anyscale CEO Keerti Melkote described the combined company as the first full-stack AI hyperscaler. For enterprise AI teams, the near-term change is optionality. Anyscale field CTO Christian Stano set out in [a Unite.AI interview](https://www.unite.ai/christian-stano-field-cto-at-anyscale-interview-series/) how platform teams are structuring Ray-based systems for production, and that architecture is what the deal touches: the same runtime, priced and operated by a company that also owns the megawatts. Once the transaction closes, Nscale’s own capacity becomes one of the places that runtime can run.