Nscale agreed on July 30 to acquire Anyscale, the company founded by Ray's creators. But Ray itself was donated to the PyTorch Foundation in 2025 and remains community-governed — so Nscale is buying roughly 200 engineers and a commercial product, not the open-source standard. That distinction defines what vertical integration in AI infrastructure can and cannot capture.
On July 30, Nscale — a London-headquartered AI cloud company that builds and operates its own data centres — announced a definitive agreement to acquire Anyscale. Most coverage led with a number Nscale did not publish: $1.65 billion, reported by Bloomberg citing a source familiar with the deal. Nscale's own release states only that financial terms were not disclosed, and it is worth holding those two facts apart rather than merging them into a headline.
The more useful detail sits four paragraphs into Nscale's own announcement, and almost nobody quoted it. Anyscale was founded by the creators of Ray, the open-source framework for distributing AI workloads across large GPU fleets. But Ray was donated to the PyTorch Foundation in 2025 and remains open source and community-governed. Nscale is acquiring the company. It is not acquiring the standard.
The structure is straightforward. Nscale supplies the physical layer — power, land, data centres, GPUs. Anyscale supplies the layer machine-learning engineers actually touch: a platform for scaling data processing, training, inference and reinforcement learning across thousands of GPUs. Anyscale's entire team, roughly 200 people across the United States, Europe and India, joins Nscale. The Anyscale brand continues and existing customers are served as before. Close is expected in the second half of 2026, subject to closing conditions and regulatory approvals.
This is the neocloud playbook in its clearest form. A generation of GPU-first infrastructure companies built businesses renting compute, discovered that renting compute is a commodity with commodity margins, and are now buying their way up the stack toward the software where switching costs live.
Ray's governance status is the load-bearing fact in this deal, and it cuts in a direction most acquisition coverage got backwards.
When a company donates a project to a foundation, it gives up unilateral control of the roadmap, the licence and the trademark. Ray sitting under the PyTorch Foundation means Nscale cannot make Ray run better on Nscale hardware than on anyone else's, cannot relicense it, and cannot fork the community's direction by fiat. Nscale has said it will join the foundation, which is the correct move and also an acknowledgement that it is joining as one participant among many rather than arriving as an owner.
What Nscale bought: roughly 200 engineers, a commercial product, a customer base, and the institutional knowledge of the people who invented Ray. What it did not buy: the ability to control Ray.
That is not nothing — the people who created a standard are genuinely the best people to build the managed product on top of it, and deep expertise in a widely-adopted framework is a real asset. But it is a different asset from owning the framework, and the difference matters when you are the customer.
The pitch behind deals like this is that owning every layer produces cost advantages a rental business cannot match. Some of that is straightforwardly true. Owning land, securing power interconnects and operating your own data centres removes a margin stack and — more importantly in 2026 — removes a scheduling dependency. Power availability, not GPU availability, is the binding constraint on AI infrastructure right now, and companies that control their own sites control their own timelines.
Where the thesis is weaker is the claim that adding a software layer deepens the moat. The hyperscalers are the natural comparison and the comparison is unflattering. AWS, Azure and Google Cloud own their physical layers too, at vastly greater scale, with lower capital costs and decades of operational depth. A neocloud competing on integration alone is competing on the axis where incumbents are strongest.
The credible version of the neocloud argument is narrower and better: specialisation. A company that does nothing but AI training and inference can make architectural choices — network topology, cooling, cluster scheduling, contract structure — that a general-purpose cloud serving a thousand workload types cannot. That is a real advantage. It is just a smaller and more defensible claim than owning every layer.
For anyone signing a multi-year GPU reservation, this deal changes the questions worth asking more than it changes the answers.
The reassuring part is genuine. Because Ray is foundation-governed, a team building on open-source Ray today has not acquired a new dependency on Nscale. Ray code is portable to any infrastructure that runs Ray, which is most of it. Nscale has also said customers remain free to choose their infrastructure. Take that commitment at face value while noting it is a statement of intent, not a contractual term you can enforce.
The part that deserves scrutiny is the gap between open-source Ray and Anyscale's commercial platform. Managed platforms built on open standards routinely add proprietary conveniences — schedulers, observability, autoscaling policies, managed state — that are not part of the upstream project. Those additions are usually the reason to pay for the managed product, and they are also exactly what does not port. The open core being neutral tells you nothing about how portable the commercial layer is.
“Owning the layer your customers build on is a moat. Owning the company that used to own the layer your customers build on is a talent acquisition with good branding.”
This is a sensible deal that has been described in slightly grander terms than it merits. Nscale gets a strong engineering team, a commercial product with existing revenue, and credibility with the developer community that uses Ray — which is a legitimately valuable thing for an infrastructure company that has until now sold to procurement rather than to engineers.
What it does not get is the structural lock-in that the phrase full-stack AI cloud is meant to imply. Ray stays neutral, and that is the outcome the ecosystem should want. The interesting question for 2027 is not whether Nscale can own the stack — the governance structure has settled that — but whether specialisation and power access are enough to sustain margins against hyperscalers who have both, at scale, and are not standing still.
On the price: until either party confirms it, $1.65 billion is one outlet's sourcing, not a fact. It has already been repeated widely enough to harden into one, which is its own small lesson about how deal coverage works.
Not directly. Ray was donated to the PyTorch Foundation in 2025 and remains open source and community-governed, so Nscale does not control its licence, roadmap or trademark. Nscale has said it will join the foundation. Teams running open-source Ray on their own infrastructure acquire no dependency on Nscale from this deal.
DrafterDaily covers what acquisitions actually transfer — and what they only appear to. One briefing on business, AI and technology, every weekday.
The headline was the 94% drawdown from a $3.44 billion peak. The more useful arithmetic is the cash sitting inside the thing being bought — and the fact that everyone except the angels was paid in another private company's paper.
The biography is the least interesting part. What the board actually did was unbundle the CEO job — keeping Tim Cook on with a defined policymaker-engagement remit — and that separation says more about Apple's next decade than the appointment does.
Every outlet reported the size of Meta's teen-safety settlement and the screen-time limits. Almost nobody read the payment structure — where roughly $5.3bn is released only if two competitors act.