If you have trained a large model, curated a multimodal dataset, or run a reinforcement learning loop across more than a handful of GPUs in the past five years, there is a decent chance Ray was somewhere underneath it. By its creators' own account, Ray has been used to build the GLM, Nemotron, Composer and MAI model families.
On Thursday, July 30, the company those creators built announced it is being acquired.
Nscale, a London-headquartered AI cloud provider, entered a definitive agreement to buy Anyscale. Nscale did not disclose financial terms. Bloomberg reported the price at $1.65 billion, citing an anonymous source, a figure TechCrunch and SiliconANGLE both carried. Roughly 200 Anyscale employees across the United States, Europe and India are joining Nscale. The deal is subject to closing conditions and regulatory approvals, and is expected to close in the second half of 2026.
Then there is the number that makes practitioners stop and reread the press release.
Anyscale's last publicly known private valuation was $1.38 billion, set in a Series C round in 2022, according to TechCrunch. That was before ChatGPT became a household product, before the GPU shortage, before every enterprise on earth started fine-tuning something. Four years and one generational technology boom later, the company that gave the field its most widely adopted distributed-compute framework is changing hands for about 20% more than it was worth going in.
The Software Layer Did Not Capture the Value
The gap between those two numbers is the story, and it says something uncomfortable about where money accrues in AI infrastructure.
Anyscale sits at the layer machine learning engineers actually touch: orchestration, observability, developer tooling, the machinery that turns a pile of GPUs into a job you can submit. Its platform is built on Ray, and Anyscale's own materials list Coinbase, Bedrock Robotics, and Runway among the companies running on it. The company told TechCrunch its revenue grew 70% in its most recent quarter compared with the previous one, which is not the profile of a business in trouble.
The buyer's trajectory is the contrast. Nscale raised $2 billion in a Series C in March 2026 at a valuation of 14.6 billion dollars, with Nvidia, Nokia, Blue Owl, Dell and the Norwegian industrial group Aker on the cap table. It has been stacking debt facilities and compute partnerships with Microsoft, British Telecom and Nordcraft on top of that.
Nscale owns power, data centers, and GPUs. Anyscale owns the software those GPUs run. One of them is worth roughly nine times the other.
Software-layer companies can still command large rounds, as the inference startup founded by the engineer behind PyTorch showed this month. What is harder is staying independent long enough to collect on it.
That asymmetry is the same one showing up across the sector, where the capital is flowing to physical capacity rather than to the tooling that sits above it. LDS has tracked the pattern in Big Tech's AI Spending Is About to Outrun Its Cash and in Google Serves 22 Billion Tokens a Minute. It Burned $5.9 Billion Doing It.
Nscale Is Buying the Part of the Stack It Could Not Build
Josh Payne, CEO and founder of Nscale, framed the deal as the last missing floor of a building the company has been putting up for two years.
"Most infrastructure providers just buy GPUs and rent them. Nscale is doing something unique. We build and own every layer ourselves: the power, the data centers, the compute, and the software that turns them into an AI cloud." — Josh Payne, CEO and Founder, Nscale (Nscale press release, July 30, 2026)
Keerti Melkote, CEO of Anyscale, made the case from the other direction, arguing that co-design beats optimizing one layer in isolation.
"Companies are moving beyond simply using AI to actually building their own. Doing that well requires the software and the infrastructure it runs on to be designed together." — Keerti Melkote, CEO, Anyscale (Nscale press release, July 30, 2026)
Melkote went on to describe the combined entity as "the first full-stack AI hyperscaler," which is a claim the market will test rather than a fact. AWS, Google Cloud and Microsoft all sell power, silicon and orchestration under one roof already. What Nscale is arguing is that it owns those layers rather than assembling them, and that a smaller vertically integrated stack can undercut the incumbents on cost.
The financial advisor list suggests both sides took the deal seriously. Goldman Sachs International served as lead financial advisor to Nscale with Morgan Stanley alongside, and Latham & Watkins as legal counsel. Qatalyst Partners was exclusive financial advisor to Anyscale, with Fenwick & West as counsel.
What Happens to Ray Is the Question That Matters
For anyone with Ray in production, the acquisition raises the obvious governance worry: does the open-source project get pulled into a GPU vendor's commercial strategy?
The structural answer is that it largely cannot, and the reason predates this deal. Ray was donated to the PyTorch Foundation in 2025. It sits there alongside PyTorch and vLLM, open source and community governed, outside Anyscale's control and therefore outside Nscale's. Nscale plans to join the Foundation as a platinum member.
Ray's five creators, Robert Nishihara, Philipp Moritz, Ion Stoica, Richard Liaw and Edward Oakes, published their own note the same day. Their argument is that the bottleneck has moved: reinforcement learning now blends training, inference and simulation into a single workload, data processing has gone multimodal and GPU-bound, and long-context inference demands memory management and routing that cannot be tuned without knowing the rack topology underneath. Optimizing one layer at a time, they wrote, is no longer enough.
They also listed who has been improving Ray over the past year: engineers from Google, NVIDIA, Microsoft, Red Hat and Alibaba, plus the wider community, working on latest-generation GPU and TPU support, topology-aware scheduling, GPU-native data processing and Kubernetes integration. That is a governance structure with too many stakeholders for one acquirer to quietly redirect.
Nscale also committed to keeping Anyscale operating under its own brand and serving existing customers as it does today. Anyscale says its platform will keep running across all major cloud providers after closing, with Nscale's capacity added as an option rather than a requirement.
Those are the right commitments to make in a press release. Whether they hold through an integration is a different question, and the answer usually arrives twelve to eighteen months later in the form of pricing changes and quiet feature gating. The first real read comes at Ray Summit in San Francisco in August.
| Anyscale | Nscale | |
|---|---|---|
| What it sells | Ray-based platform for data processing, training, inference, reinforcement learning | Power, data centers, GPUs, cloud platform services |
| Headquarters | San Francisco | London |
| Team size | About 200 people (US, Europe, India) | Not disclosed in the announcement |
| Last known valuation | 1.38 billion dollars (Series C, 2022) | 14.6 billion dollars (Series C, March 2026) |
| Named customers or partners | Coinbase, Bedrock Robotics, Runway | Microsoft, British Telecom, Nordcraft |
The Counterargument Is That the Comparison Is Unfair
The valuation gap makes a tidy narrative, and it deserves pushback.
Private valuations from 2022 were set in a market that no longer exists. A great many companies priced in that window are worth less today than they were then, and an exit at any premium is a better outcome than most of that cohort will see. Anyscale's 70% quarter-over-quarter revenue growth argues the business was compounding, not stalling, and founders sometimes sell into strength because integration unlocks something they cannot build alone.
The price is also not confirmed. Nscale explicitly declined to disclose terms. The $1.65 billion figure comes from Bloomberg's reporting on an anonymous source, and the 2022 valuation was a post-money private mark, not a cash valuation of the business. Comparing the two is directionally useful and precisely wrong.
What is not in dispute: the team that built the most widely used framework for scaling Python and AI workloads concluded that its future runs through a company that owns power plants and data centers.
The Bottom Line
For ML engineers, the immediate practical impact is close to zero. Ray is governed by the PyTorch Foundation. Anyscale keeps its brand, its customers, and its multi-cloud posture. Nothing in your Ray cluster changes on Monday.
The strategic impact is larger and slower. Every serious neocloud is now buying its way up the stack, because renting GPUs is a commodity business with commodity margins and orchestration software is where the switching costs live. Anyscale is the second orchestration-layer company this year to be absorbed by a capacity owner rather than to grow into an independent platform, and the pattern points toward an infrastructure market with fewer neutral layers in it.
The teams that will feel this first are the ones who assumed their tooling vendor would stay independent long enough to matter. Ray's donation to a foundation looks, in retrospect, less like a philosophical gesture and more like insurance.
Anyscale's own framing of the deal was that neither company could optimize its layer alone. That is probably true. It is also the argument every acquirer makes right before the layers stop being separable.
Sources
- Nscale Acquires Anyscale, Enhancing its Full Stack AI Cloud Platform (Nscale press release, Jul 30, 2026)
- Anyscale signs definitive agreement to join Nscale (Anyscale, Jul 30, 2026)
- Nscale buys Anyscale as it seeks to own more of the AI compute stack (TechCrunch, Jul 30, 2026)
- Nscale to Buy AI Software Startup Anyscale for 1.65 Billion (Bloomberg, Jul 30, 2026)
- Nscale buys AI infrastructure optimization startup Anyscale for reported 1.65B (SiliconANGLE, Jul 30, 2026)
- Sandberg, Clegg join Nscale board as this Stargate Norway startup hits 14.6B valuation (TechCrunch, Mar 9, 2026)
- Anyscale unveils Ray 2.0 and Anyscale innovations at Ray Summit 2022 (Anyscale, 2022)