PyTorch Foundation's New Member Wave Shows Open Source Has Become GPU Cloud Infrastructure

At PyTorch Conference North America in San Jose this week, the PyTorch Foundation announced nine new members: Nebius, Nscale, and Shopify at the Platinum tier; LightSeek Foundation and Lablup at Silver; and four academic and nonprofit members at the Associate level. Taken alone, a membership roster is a press-release footnote. Taken alongside what each of those three Platinum names actually does, it reads as something more specific — a snapshot of how GPU cloud providers are buying their way into the governance of the open-source stack that runs on top of their chips, rather than just renting the chips underneath it.

The Nscale tell

The most telling name on the list is Nscale, the London-based AI cloud provider that in July signed a definitive agreement to acquire Anyscale, the company founded by the creators of Ray. Ray is the distributed-compute framework used to train and serve large models across thousands of GPUs, and it was donated to the PyTorch Foundation in 2025 alongside PyTorch itself and the inference engine vLLM — meaning Ray's technical direction is no longer Anyscale's to set unilaterally. Nscale and Anyscale's own announcement of that deal was explicit that Nscale would join the Foundation as a Platinum member once the acquisition closed; this week's announcement is that promise coming due. Bloomberg reported the deal at roughly $1.65 billion, citing an anonymous source, against Anyscale's last disclosed private valuation of $1.38 billion from a 2022 round — a modest markup that undersells what Nscale is actually buying, which is less Anyscale's balance sheet than a formal vote in how Ray evolves.

That matters because Ray's governance is genuinely distributed in a way that pure corporate ownership isn't: contributors from Google, NVIDIA, Microsoft, Red Hat, and Alibaba all work on its GPU scheduling and Kubernetes integration, and none of them answer to Nscale. Buying Anyscale doesn't hand Nscale the project; it hands Nscale a seat on the committee that decides the project's direction, at the exact moment Ray is being optimized for exactly the kind of multi-gigawatt GPU buildouts Nscale is selling.

Why a cloud provider wants a vote on a framework

PyTorch Foundation executive director Mark Collier framed the wave of new members in terms of compounding returns: "Open systems learn faster when more people, projects, and organizations are working together, creating compounding gains... Open source ecosystems let breakthroughs travel between organizations instead of staying locked inside one company." That's the sales pitch. The more mechanical explanation is that GPU cloud providers have figured out that owning capacity is no longer the differentiator it was two years ago — the software layer that schedules and serves workloads across that capacity is where the next round of lock-in gets built, and the biggest pieces of that layer are open source and foundation-governed rather than any single vendor's property. Nebius joining at the same tier, with its own pitch about contributing "high-performance infrastructure and applied AI" expertise, suggests this isn't an Nscale-specific maneuver so much as a pattern: neoclouds are converging on PyTorch Foundation membership as the price of influence over frameworks they increasingly depend on and increasingly want to shape toward their own hardware.

The Shopify number worth checking

The announcement's other notable inclusion is a case study from Shopify, which says it uses PyTorch and vLLM to feed production failures from its GraphQL commerce agent, Sidekick, back into model weights on a continual-learning loop the team calls "the flywheel" — and claims the result beats frontier-model quality at a 96% lower serving cost. That figure isn't just a vendor's talking point in a Foundation press release: a separate write-up in ZenML's independently maintained LLMOps case-study database describes the same Sidekick flywheel, the same 96% figure, and the same rubric-based validation process, including a detail the PyTorch post omits — that two top annotators blindly re-score random samples to check inter-annotator agreement before any result is trusted. ZenML's own writeup flags that the 96% figure's evaluation methodology isn't fully disclosed and should be read as directional rather than audited, which is a fair caveat for a number this attractive to repeat uncritically.

Doing that well requires the software and the infrastructure it runs on to be designed together.

That line, from Anyscale (now Nscale) CEO Keerti Melkote in the original July deal announcement, doubles as the thesis connecting all three Platinum additions: the open-source AI stack isn't just a shared set of libraries anymore, it's the contested ground where infrastructure companies are deciding they need a formal vote, not just an API key.