Nscale, the British AI cloud company that owns the power and data centers underneath its chips, is buying its way up the stack. The London firm said Thursday it has signed a definitive agreement to acquire Anyscale, the San Francisco software startup built by the creators of the open-source Ray framework. Bloomberg put the price at about $1.65 billion, citing a person familiar with the deal. Neither company disclosed terms publicly.

The purchase hands Nscale the software layer that decides how efficiently each expensive GPU actually works. It slots into a strategy the company has pushed since it emerged from a crypto-mining business two years ago: own every layer of the AI compute chain rather than rent chips out by the hour.

The deal is expected to close in the second half of 2026, pending regulatory approval. Anyscale's staff of about 200 will join Nscale, though the startup keeps its own brand and its existing customers.

What Nscale is actually buying

Anyscale's story starts in a Berkeley research lab. Its founders built Ray, an open-source framework for spreading Python and machine-learning jobs across large fleets of machines. The commercial company grew up around that project, selling the orchestration and observability tools that let teams run heavy workloads without wiring everything together by hand.

When GPT-3 pulled AI into the mainstream, Anyscale pointed Ray at the new demand. Its platform now handles data preparation, model training, fine-tuning, inference, and reinforcement learning across thousands of GPUs at once. The company says stitching those jobs into one system can cut the total cost of running them by as much as 90% against a patchwork of separate tools.

Ray itself is no longer Anyscale's to sell. Governance of the open-source project moved to the PyTorch Foundation under the Linux Foundation in late 2025, and it stays community-run after the deal closes. Nscale said it will join the PyTorch Foundation as part of the acquisition.

The logic is about margins

Renting raw GPUs by the hour is a commodity business.

Every neocloud is buying the same Nvidia systems and fighting the same price war to rent them out, and the hardware depreciates fast enough to eat roughly half of revenue at the largest players. There is not much room to stand apart when the product is an interchangeable chip billed by the hour.

The software sitting on top of those chips is where the pricing power lives. It is stickier than raw capacity, and it carries the fatter margins. It also determines how much useful work a customer pulls from every GPU-hour they buy.

That is the pitch Nscale's leadership has been making in public. "Most infrastructure providers just buy GPUs and rent them," CEO Josh Payne said, arguing that Nscale instead builds and owns the power, the buildings, the compute, and the software itself. Product chief Dan Bathurst framed Anyscale as the piece that turns Nscale into a single home for model training and the inference that comes after.

By folding Anyscale in, Nscale can tune the software and the hardware beneath it together, a coordination that Anyscale said neither company could pull off alone while optimizing its own layer in isolation.

An IPO story that needed a better middle

The timing is not an accident. Nscale is reportedly preparing a public listing for the second half of 2026, with Goldman Sachs and JPMorgan among the banks lined up behind it. Goldman advised Nscale on the Anyscale purchase, while Morgan Stanley advised Anyscale.

A pure GPU-rental business is a hard sell to public investors who have watched that market turn into a race to the bottom on price. A full-stack story reads better. Owning the software that manages workloads lets Nscale tell would-be shareholders it captures value the commodity players cannot touch.

The company has spent 2026 building toward that pitch. In March it raised $2 billion in a Series C that valued it at $14.6 billion, with Nvidia, Nokia, Dell, Blue Owl, and Norwegian industrial group Aker among its backers. It signed a letter of intent with Microsoft covering more than a gigawatt of power and committed $2.5 billion to UK data centers. In July it secured a $900 million revolving credit line. Anyscale is the software capstone on top of all that spending.

Anyscale brings its own momentum to the table. The startup was valued at about $1.38 billion in a 2022 funding round, and it said revenue grew 70% in its most recent quarter against the one before.

Not the first neocloud to reach for software

Nscale is walking a path its rivals are already on. Nebius, another AI-focused cloud, recently bought a startup called Eigen AI for $643 million in the same kind of move up the value chain. The neocloud category, which barely existed before 2024, is maturing past the phase where owning the most GPUs was the whole game.

The competitive backdrop is crowded. CoreWeave, the one large neocloud already trading publicly, sits on a revenue backlog reported in the tens of billions of dollars. Nebius, Lambda, Crusoe, and IREN are all chasing the same customers and the same Nvidia allocations.

Differentiation now comes from the layers around the chip. Cheaper power is one lever. Data residency inside a customer's own borders is another. Software that wrings more output from each GPU is a third, and it is the one Nscale just paid $1.65 billion to add.

The risk sitting under the whole trade

For all the vertical-integration logic, the neocloud model carries a structural worry that a software acquisition does not erase.

The biggest customers can become the biggest competitors. In early July, reporting that Meta was exploring its own cloud offering under the working name Meta Compute knocked around 15% off the share prices of CoreWeave and Nebius in a single stretch. The hyperscalers and large AI labs that rent neocloud capacity today are also capable of building their own, and some are signaling they intend to.

Profitability is the other overhang. None of the major neoclouds turns a GAAP profit, because the GPUs they buy lose accounting value so fast that depreciation swallows much of what they earn. Buying Anyscale gives Nscale a higher-margin software line, but the company has not disclosed Anyscale's absolute revenue or the integration costs. Nor has it said how long the software will take to lift Nscale's overall margins.

Those are the questions public-market investors will ask when Nscale files to go public. Owning the software layer strengthens the story it will tell them. Whether it strengthens the financials underneath is the part the $1.65 billion price tag has not yet answered.