Nscale's $1.65B Anyscale buy: keep your escape hatch

Nscale, the British AI neocloud, agreed on July 30 to buy Anyscale for roughly $1.65 billion. A company that sells you GPUs now owns the orchestration software that decides how your jobs run on them. That's the whole story, and there's a version of it sitting in your stack right now — the difference is whether you left yourself a way out.

What actually happened

Per TechCrunch and Bloomberg's reported price, Nscale is acquiring Anyscale's commercial platform, engineering team, and customers — Coinbase, Runway, and Bedrock Robotics among them — with the deal expected to close in the second half of 2026. Anyscale was last valued at $1.38 billion in a 2022 Series C and said revenue grew 70% in its most recent quarter. Nscale's business already spans power, data centers, and GPU clusters; buying the workload layer means it now owns training, fine-tuning, and inference end to end.

Here's the part worth stealing. Anyscale built its company on Ray, the distributed compute framework its founders wrote at Berkeley. In October 2025, Anyscale handed Ray to the PyTorch Foundation — neutral governance, alongside PyTorch and vLLM, with 237 million downloads at the time. So when the parent company got bought by its own infrastructure supplier nine months later, Ray didn't get bought with it. It's still community-governed and still runs anywhere.

Why the Anyscale acquisition matters for your business

You are not buying GPU clusters. You are buying five or six SaaS tools that will get acquired, repriced, or sunset in the next 36 months — and the question is always the same: when your vendor's incentives change, what do you actually still control?

Anyscale customers get to answer that with "the framework, because it lives at a foundation." Most of your vendors don't offer that. So build the equivalent yourself, deliberately:

• Your data lives in a database you can dump. Postgres, MySQL, a CSV export that actually round-trips. If the only copy is inside someone's UI, you don't own it.
• Your integrations run through code you can read. A workflow in your repo beats a workflow in a vendor's canvas, even when the canvas is nicer.
• Your model calls go through one adapter. Swapping providers should be a config change, not a rewrite.
• You've priced the exit before you sign. Ask every vendor what export looks like. Watch how fast they answer.

None of that is paranoia. It's the same move Anyscale made on purpose, and it's the reason this acquisition is a business event instead of a migration emergency for everyone downstream.

Sources: TechCrunch, PyTorch Foundation.