Nscale, the London-based AI cloud that raised $2 billion in March at a $14.6 billion valuation, announced Thursday it has signed a definitive agreement to acquire Anyscale, the San Francisco company built by the creators of Ray. Bloomberg reported the price at roughly $1.65 billion, citing an anonymous source. Nscale’s press release says financial terms were not disclosed. Anyscale’s entire team — about 200 people across the U.S., Europe and India — joins Nscale. The deal is expected to close in the second half of 2026, subject to regulatory approval.
What Nscale bought is not capacity. It already has plenty: data centers in Norway, the UK, Texas and West Virginia, partner sites in Portugal and Iceland, compute and data center deals with Microsoft, BT and Nordcraft, and an investor list that runs through Nvidia, Nokia, Dell, Blue Owl and Norwegian industrial group Aker. What it bought is the layer sitting on top — the orchestration software machine-learning teams use to spread a training or inference job across thousands of GPUs without hand-wiring it.
The price tells its own story. Anyscale was last valued at $1.38 billion in a 2022 Series C. If Bloomberg’s number is right, four years and one global AI boom later, the company changed hands for about 20% more than that — while the cost of the GPUs its software schedules repriced several times over. Anyscale says revenue grew 70% quarter over quarter in its most recent quarter, so this isn’t a distressed sale. It’s a repricing of what the market thinks the software layer is worth versus the metal.
Our take: Renting GPUs is a spread business. You borrow to buy chips, rent them out, and pray utilization holds until the depreciation schedule runs out. Anyone with capital and a power contract can do it, which is exactly why the price grinds down toward cost of capital plus electricity. The software layer is the only part a customer can’t re-tender in an afternoon. Nscale’s CEO said the quiet part out loud in the release: “Most infrastructure providers just buy GPUs and rent them.” He paid a modest multiple to stop being one of them. Watch whether the rest of the neocloud field starts buying software instead of land.
The part that matters if you actually buy compute
Three things hold, at least on paper. Ray stays open — it was donated to the PyTorch Foundation in 2025 and remains community-governed, and Nscale says it is joining the foundation. Anyscale keeps its own brand and its existing customers, which include Coinbase, Runway and Bedrock Robotics. And customers stay free to run their workloads on whatever infrastructure they choose.
The incentive still changes. Over time, the cheapest and best-tuned place to run Anyscale will be Nscale, because the same company will be tuning both ends. That is the entire pitch — Anyscale’s CEO called the combination “the first full-stack AI hyperscaler.” If your AI stack currently treats compute as a commodity you shop every renewal, the layer above the compute is where the switching cost is quietly being installed.
What to watch
- Regulatory clearance. The deal needs approvals on both sides of the Atlantic and is only targeted to close in the second half of this year. A UK neocloud buying a US software company at this moment is not a rubber stamp.
- Copycats. The other capacity-first neoclouds have the same margin problem. If one of them buys an orchestration or observability layer in the next two quarters, the pattern is confirmed.
- Ray’s governance. Community-governed open source with a vertically integrated commercial patron is a well-worn arrangement, and it works right up until it doesn’t. Watch the PyTorch Foundation, not the press release.
- Advisor tell. Goldman Sachs led for Nscale with Morgan Stanley alongside; Qatalyst ran the sell side. That is bank bench you assemble when you expect to keep buying.
The AI infrastructure trade spent two years being about who could pour the most concrete and secure the most megawatts. This is the first big deal of the next phase: the buildout is producing more raw capacity than anyone can differentiate on, and the winners will be whoever owns the thing customers are too embedded to leave.