Ray was never for sale. Anyscale donated it to the PyTorch Foundation in October 2025, and it stays open and community-governed regardless of who owns the company that wrote it. What Nscale is buying — in a deal Bloomberg valued at roughly $1.65 billion, a number neither company has confirmed — is the commercial layer above Ray: the hosted control plane that creates your clusters, schedules your jobs, manages their lifecycle and aggregates your metrics.
If you run Anyscale's managed platform, that control plane is your dependency, and its owner is changing from a venture-funded software company into a debt-financed GPU landlord with a gigawatt-scale data centre buildout to pay for. The transaction is subject to closing conditions and regulatory approvals and expected to close in the second half of 2026. The window in which your renewal terms, exit rights and architecture commitments still carry leverage is open right now. It shuts when the deal does.
What Nscale Actually Bought
Nscale bought a company, a team and a customer base — not a framework. Anyscale's roughly 200 employees across the US, Europe and India join Nscale, the brand continues, and named customers including Coinbase, Runway and Bedrock Robotics come with it. Anyscale was founded in 2019 by the Berkeley researchers behind Ray — among them Robert Nishihara, Philipp Moritz and Ion Stoica — and is run today by CEO Keerti Melkote.
Be careful with the price, because the public record is thinner than it looks. TechCrunch puts Anyscale's last mark at $1.38 billion in a "2022 Series C", but no 2022 round appears in the funding record. The round's own investor announced a $100 million Series C at a $1 billion valuation in December 2021, and $1 billion is the last valuation funding trackers carry. Against that mark, and reported 70% revenue growth in its most recent quarter, an unconfirmed $1.65 billion is a real step up across four and a half years — not a distressed outcome. Nobody has disclosed the sale process, the cash-versus-stock mix or whether rival bidders existed, so do not read the headline number as evidence of how much autonomy the acquired team keeps after close. Underwrite the acquirer instead.
Nscale's own framing is vertical integration. CEO Josh Payne describes a company that builds and owns every layer itself — the power, the data centres, the compute and the software, and Melkote calls the combination "the first full-stack AI hyperscaler." That is the whole thesis. It is also the whole risk.
Ray Is Open Source. The Control Plane Is Not.
The distinction that matters is architectural, and it is documented in public. Anyscale splits its platform into two halves: a control plane it operates, and a data plane you operate. Microsoft's own documentation for the Anyscale on Azure preview spells out which is which — "The control plane is managed by Anyscale and hosted in Azure. The data plane runs entirely within your Azure subscription."
A control plane is the service that decides what runs; a data plane is where it actually runs. In Anyscale's case the control plane owns the console, the scheduling and job management APIs, monitoring and log aggregation, and cloud and cluster lifecycle management. Your data plane holds the Kubernetes cluster, the Ray clusters themselves, your storage and your load balancer. Microsoft's component table is unambiguous about ownership: the console and the scheduling APIs sit in an "Anyscale-hosted Azure tenant," owned by Anyscale. Your compute and your data are yours.
The bridge between them is an operator running inside your cluster. The Anyscale Kubernetes operator is installed into your AKS cluster automatically during cloud creation and "polls the control plane endpoint (<cloud-id>.anyscale-cloud.dev) for pending operations." All connections originate outbound from your cluster, so there are no inbound firewall rules to review — which is good security design and precisely why this dependency is easy to forget you have.
This is the honest version of the lock-in question. Your data does not move. Your GPUs do not move. Your models do not move. But the thing that starts, scales and stops your training and inference jobs is a hosted SaaS whose corporate parent, after close, sells GPUs in competition with the cloud you are currently running on.
The Counterparty You Did Not Underwrite
Nscale is a two-year-old balance sheet, and that is the change most Anyscale customers have not priced. The company grew out of a cryptocurrency-mining business in early 2024, raised a $2 billion Series C in March 2026 at a $14.6 billion valuation from investors including Nvidia, Nokia, Dell, Blue Owl and Aker, and has layered on debt: a $790 million Norwegian facility, a $1.4 billion delayed-draw term loan and a revolving credit facility, with a $900 million revolver closed in July 2026. It has committed $2.5 billion to UK data centres and is building a 1.35GW site for Microsoft in West Virginia with deployment starting in 2027. An IPO is reportedly possible later this year. The board includes Sheryl Sandberg and Nick Clegg.
None of that is a red flag on its own. Capital-intensive infrastructure is financed with debt; that is what debt is for. But it changes the shape of your vendor risk in three specific ways, and each one has a contractual answer:
- Utilisation pressure is structural. Interest payments and data centre capital commitments do not care whether your workload runs on Nscale or on AWS. The incentive to migrate the acquired customer base onto owned capacity is not a conspiracy — it is the business model, stated plainly.
- A pre-IPO owner optimises for reported metrics. Pricing changes, packaging changes and support-tier changes are the cheapest levers on a platform with an existing installed base.
- Roadmap priority follows the fleet. Feature work that makes Anyscale run better on Nscale hardware is cheap to justify. Feature work that makes it run better on a competitor's is not.
Nscale has addressed the first of these in writing. The press release commits that "Customers will remain free to choose the infrastructure on which they run their AI workloads" and that Anyscale "will continue to operate under its brand and serve its existing customers as it does today," with Nscale's platform arriving "over time" as an additional option. Anyscale's own announcement goes further on both counts: "Post-closing, the Anyscale Platform will continue to run across all major cloud providers. Portability remains core to our roadmap" — and Nscale "plans to join the Foundation as a platinum member," putting its own money behind the governance that keeps Ray neutral. Take that seriously — and then get it into your contract, because a press release is not an obligation and the people who wrote it are not the people who will own the P&L in 2028.
The Last Time a GPU Landlord Bought the Software Layer
This has run before, recently, and the early evidence cuts against the alarmist reading. CoreWeave announced its acquisition of Weights & Biases on 4 March 2025 — a GPU cloud buying the ML tooling layer sitting above it — and made almost the same promise, that customers "will also be able to continue deploying workloads wherever they like, be it on-premise or with the infrastructure provider of their choice." Over a year after that deal closed, Weights & Biases still runs on other people's infrastructure. IREN's purchase of Mirantis followed the same pattern, and we covered it as a bitcoin miner buying its way into the AI cloud.
The analyst case is stronger still. Futurum's Nick Patience points out the genuine paradox in the deal: Ray "is open source, was donated to the PyTorch Foundation in 2025, and runs perfectly well on AWS, Google Cloud, Azure, or any GPU cluster," so Nscale gains no exclusive technical moat at all. What it gains is a roughly 200-person team, engineers deep in Ray among them, who can co-design the software and the hardware together — which is a real advantage and not one that requires trapping anybody.
So steel-man it properly: the most likely outcome is that Anyscale keeps working on your cloud, gets better at GPU scheduling because the people writing it now own the racks, and adds a cheaper deployment target you are free to ignore. That is a good deal for you.
The point is not that it will go badly. The point is that you find out over three years, and the terms that protect you are only negotiable for the next few months.
What Portability Actually Costs You
Your fallback is real, free and boring: KubeRay, the open-source Kubernetes operator maintained under the broader Ray project and licensed Apache 2.0. It gives you RayCluster, RayJob and RayService as Kubernetes custom resources, plus autoscaling that sizes clusters to the workload. If your team already runs Kubernetes competently, this is the exit, and it is the reason your negotiating position is better than it feels.
Price the difference honestly, because it is not zero. Moving to KubeRay costs you the console, the hosted scheduling and job-management APIs, log aggregation and the metrics dashboard, and the vendor support contract behind all of it. Somebody on your team inherits cluster lifecycle management, upgrade testing and on-call. For a platform team of four running dozens of jobs, that is a serious tax. For a large ML platform group already operating GPU capacity at low utilisation, it is mostly work they are doing anyway.
Then price what you pay today. Anyscale's hosted tier bills in credits per instance-hour — AC 4.9591/hr for an NVIDIA A100 and AC 1.3635/hr for an A10G, with H-, B- and GB-series families requiring a sales conversation — on a pay-as-you-go basis with committed-contract discounts above that. In a bring-your-own-cloud deployment, that platform charge sits on top of the cloud bill you are already paying. The delta between that number and the loaded cost of running KubeRay yourself is your leverage. Calculate it before you walk into the renewal, not after.
Do This Before the Deal Closes
This Week: Pull your Anyscale agreement and find three clauses — assignment and change of control, price protection on renewal, and termination for convenience with a data and configuration export right. If the change-of-control clause is silent or permits assignment without consent, that is the one to fix. Separately, have your platform lead produce a one-page map of every place the Anyscale control plane sits in a production path: which jobs, which services, which on-call runbooks assume the console is up.
This Month: Run one real workload on KubeRay in a non-production cluster and time it end to end — provisioning, submitting a job, autoscaling under load, and failure recovery. You are not migrating. You are establishing a number, in engineer-days, that you can put on the table. Then re-baseline the platform charge against that number using current published rates.
Before Close: Ask for three things in writing, in an amendment rather than an email. First, a continued-support commitment naming AWS, GCP and Kubernetes explicitly for the length of your term. Second, price protection through at least one renewal, with a cap on increases. Third, a defined migration-assistance and data-export obligation triggered by any material change to supported deployment targets. Pre-close, when the acquirer wants a clean customer base and no churn story ahead of a possible IPO, is the cheapest time you will ever ask. We made the same argument about contract timing in the Tricentis–Tabnine deal and Okta's purchase of Permiso, and the pattern holds: leverage is highest in the gap between announcement and close.
The Bottom Line
We have twice written about running Anyscale on Azure to cut inference and training costs, including the own-models-versus-rent-APIs case. That advice still holds — the economics did not change on 30 July, and Ray is genuinely safe. What changed is who your counterparty will be, and the fact that Anyscale on Azure is still a preview service with no service level agreement makes the Azure path in particular worth a second look before you commit further.
This is the same lesson every consolidation cycle in enterprise infrastructure has taught, from virtualisation to observability to the wave of AI vendor acquisitions running through 2026. Open source protects the code. It does not protect the service. The framework you can fork; the console you cannot. When a company that sells compute buys the software that decides where your compute runs, nothing bad has to happen for you to end up with fewer options than you had.
Ray is fine. Go read your contract.
Continue Reading
- How Anyscale Cuts Enterprise AI Costs 90% on Azure
- Anyscale Azure Cuts AI Costs 90%: Own Models vs Rent APIs
- Tricentis Bought Tabnine's Context Engine, Not Your IDE
- Okta Bought Permiso. Your Leverage Expires Oct 31.
- $145B Cloud War: Meta's Move That Wiped $12B in One Day
- AI Vendor Bankruptcies 2026: 7 Warning Signs + CIO Checklist
- Qualcomm Spent $4B to Break Nvidia's Lock on Enterprise AI