Capital

The Scheduler and the Meter Now Have One Owner

Nscale agreed on 30 July to buy Anyscale for a reported 1.65 billion dollars. Ray stays open source. The control plane that places your jobs does not.

CapitalBy Servola Tech Desk2026-07-315 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

The Scheduler and the Meter Now Have One Owner

Key takeaways

On 30 July, Nscale agreed to buy Anyscale for a reported 1.65 billion dollars, which means the company selling you GPU hours will also own the software that decides which GPUs your job lands on. Ask for placement neutrality in writing before the deal closes in the second half of this year.
Ray itself is safe, and that is not the point. Anyscale donated Ray to the PyTorch Foundation in 2025 and Nscale says it will join the foundation, so the framework stays community governed. The managed control plane above it, where scheduling and placement actually happen, was never in the foundation and is the asset that changed hands.
Anyscale was last marked at 1.38 billion dollars in its 2022 Series C, so a reported 1.65 billion is an uplift of roughly 20 percent over four years. For an AI infrastructure asset in 2026 that is a restrained step, which tells you the price was set by the stack position rather than by the revenue line.
Nothing is contractually different yet. The transaction is subject to closing conditions and regulatory approvals, and Anyscale keeps its brand and its customers until completion. That window is the only leverage a customer is going to get.

A compute seller bought the thing that places the work

On 30 July, Nscale announced it had agreed to acquire Anyscale. Nscale is a British AI cloud operator that sells data centre capacity and GPU hours. Anyscale sells the managed platform that distributes machine learning workloads across those machines, and it was built by the team that created Ray, the open source Python framework for distributed computing. Neither company disclosed financial terms. Bloomberg reported a price of about 1.65 billion dollars citing a person familiar with the transaction, and that figure has not been confirmed on the record by either party.

The transaction is subject to closing conditions and regulatory approvals and is expected to complete in the second half of 2026. Anyscale's roughly 200 staff across the United States, Europe and India move to Nscale. Anyscale continues to operate under its own brand and to serve its existing customers, among them Coinbase, Runway and Bedrock Robotics, and Nscale says those customers remain free to choose their infrastructure.

What actually changed: one company will own both the layer that decides where a training job runs and the meter that charges for the machine it runs on.

The open source answer addresses a different question

Both companies moved quickly to reassure the Ray community, and the reassurance is real. Ray was donated to the PyTorch Foundation in 2025, remains open source under community governance, and Nscale says it will join the foundation to reinforce its commitment to the project. Keerti Melkote, Anyscale's chief executive, described the combination as creating the first full-stack AI hyperscaler. Josh Payne, Nscale's founder and chief executive, said Anyscale's managed services complete a vertically integrated AI cloud platform.

Read those two statements together. The reassurance is about Ray. The strategy is about vertical integration. Ray was never the part that could be used against a customer, because a framework you can fork does not lock anyone in. The managed control plane sitting above it, the part that schedules, autoscales and places workloads, is proprietary, was never inside the foundation, and is precisely the asset that changed hands.

The distinction worth holding on to: foundation governance protects the code you run. It says nothing about the commercial incentives of the company running that code for you.

What 1.65 billion dollars actually prices

Anyscale was last valued at 1.38 billion dollars in its 2022 Series C. If the reported figure is accurate, four years and a pivot into large model training, serving, data curation and reinforcement learning produced an uplift of roughly 20 percent. Measured against 2026 valuations for almost anything adjacent to AI infrastructure, that is restrained, and it is not a story about a business that stopped growing: Anyscale reported 70 percent revenue growth in its most recent quarter against the preceding one.

Now set it against the buyer. Nscale raised 2 billion dollars in a March 2026 Series C at a 14.6 billion dollar valuation, backed by Nvidia, Nokia, Blue Owl, Dell and the Norwegian industrial group Aker, and has arranged debt facilities of more than 2.1 billion dollars. A 1.65 billion dollar purchase is roughly eleven percent of its own equity value and slightly less than its last primary raise, and it is happening while the company is reported to be positioning for a public listing.

The inference for anyone buying AI capacity: the price reflects what owning the software layer does for the seller's margin and its listing story, not what that software layer earns today.

Where a European team actually gets exposed

Most European engineering teams using Ray run it themselves on their own Kubernetes, and for them the foundation guarantee is the whole answer. The exposure sits with teams on the managed product, or on the roadmap that feeds it. Three things are worth watching rather than assuming. Whether new platform capabilities arrive on Nscale capacity first and elsewhere later. Whether pricing for the managed platform stays independent of where the compute is bought. And whether support quality for a customer running on a competing cloud holds up after integration.

There is a sovereignty dimension, and for once it points the other way. Nscale is British and operates capacity in Europe, so for a German, Nordic or Dutch team this moves a piece of the AI stack from a San Francisco vendor to a European one. That is a real gain in jurisdiction. It does not offset the concentration point, and the two should be priced separately.

Do this before completion: the deal is not closed. Terms agreed with Anyscale now bind the acquirer afterwards.

Four questions to put to your provider this week

Ask whether the managed platform will guarantee workload placement on a cloud of your choosing after completion, in the contract rather than in a blog post. Ask what the notice period is for a change to platform pricing, and whether that period survives a change of control. Ask which platform features are being built against Nscale specific hardware, because divergence starts there rather than in a pricing announcement. Ask what happens to your data plane if you leave, and how long a full export takes at your volume.

If you self-host Ray, the exercise is different and much smaller. Confirm that nothing in your stack depends on an Anyscale hosted endpoint, a proprietary operator or a licence key, and write down every place it does. That list is your actual exposure, and for most teams it will be short.

Frequently asked questions

Does this acquisition put Ray at risk?

No. Anyscale donated Ray to the PyTorch Foundation in 2025 and it remains open source under community governance, and Nscale has said it will join the foundation. If you self-host Ray, your licence position is unchanged. What changed is the ownership of Anyscale's proprietary managed platform, which was never part of the foundation.

When does the deal complete, and can anything still change?

Nscale expects completion in the second half of 2026, subject to closing conditions and regulatory approvals. Until then Anyscale operates under its own brand with its existing customers. Terms you agree during that window carry over to the acquirer, which is why it is the moment to ask for them.

How much did Nscale pay for Anyscale?

Neither company disclosed the value. Bloomberg reported approximately 1.65 billion dollars, citing a person familiar with the deal. For context, Anyscale was last valued at 1.38 billion dollars in its 2022 Series C round.

Vertical integration is not a scandal, and Nscale is doing nothing unusual by buying its way up the stack. The mistake would be to accept a reassurance about Ray as an answer to a question about Anyscale. The framework is governed by a foundation. The platform is governed by whoever owns it, and from the second half of this year that is a company whose margin improves when your job runs on its own machines. Get the placement commitment into the contract while you still have a counterparty that wants your signature.

NscaleAnyscaleRayPyTorch FoundationAI neocloudVendor concentrationChange of control

A Vendor You Never Hired Holds Your Tenant Keys

Inforcer raised 50 million dollars on 30 July to manage Microsoft 365 for managed providers. The firm with the deepest reach into your tenant is one you never signed.

5 min read

Capital

Splitting AI Risk Across Chips And Power Failed

A 20 billion dollar fund was short the chip layer and long the power layer. Both fell together in July. What that means if you buy compute from that cohort.

5 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.

3 min read

← All articles