# Astral Shuts Down pyx, Open-Sources the Part That Mattered


Astral is winding down [pyx](https://astral.sh/blog/introducing-pyx), its hosted commercial registry, and open-sourcing the GPU index and prebuilt-wheel infrastructure underneath it, with the artifacts free. Charlie Marsh confirmed the plan on [Talk Python #552](https://talkpython.fm/episodes/show/552/astral-joins-openai), recorded about a month after the [OpenAI acquisition closed](https://pydevtools.com/blog/openai-acquires-astral.md).

The commercial layer was the part Astral could afford to lose. The open tools and the GPU-packaging engine are what survive. For anyone who feared the OpenAI deal would lock the good parts behind a paywall, that ordering is the reassuring signal.

## What pyx sold

pyx launched in beta in August 2025 to fix the problem [uv](https://pydevtools.com/handbook/reference/uv.md) users keep hitting on GPU machines: installing PyTorch, CUDA, and libraries like FlashAttention that build against both. It served prebuilt wheels with hardware-aware selection, so the registry figured out which build matched the machine. It was the revenue answer for a venture-funded company giving away uv and [Ruff](https://pydevtools.com/handbook/reference/ruff.md), and it went unmentioned in the [March acquisition announcement](https://siliconangle.com/2026/03/19/openai-acquires-open-source-python-tooling-startup-astral/). Now its future is settled, with the service gone and the engineering public.

## Why open-sourcing it beats a paywall

The usual fear when a big company buys critical open-source tooling is enclosure: the free thing gets starved while value moves behind a paywall. Shutting down pyx runs the other way. Marsh's account is that venture funding forced the commercialization, and OpenAI's backing removes the pressure.

> I genuinely think it's possible that we end up writing more open source here than we did at Astral.

Open GPU packaging is the piece worth the most. Installing CUDA-linked wheels correctly is one of the sharpest edges in Python, which is why the handbook keeps separate guides for [PyTorch](https://pydevtools.com/handbook/how-to/how-to-install-pytorch-with-uv.md), [vLLM](https://pydevtools.com/handbook/how-to/how-to-use-vllm-with-uv.md), and [RAPIDS](https://pydevtools.com/handbook/how-to/how-to-install-rapids-with-uv.md), and why the [wheel variants](https://pydevtools.com/handbook/explanation/what-are-wheel-variants.md) proposal exists. A free, open index helps every uv user on a GPU, not just the enterprises pyx would have billed.

## The tools kept shipping

Marsh said uv's release cadence held through the acquisition, [ty](https://pydevtools.com/handbook/reference/ty.md) is on track for a stable release later in 2026, and the team is clearing its feature backlog. Whether the trajectory holds is the open question, and one quarter under a new owner is thin evidence; the handbook's [long-term analysis of uv](https://pydevtools.com/handbook/explanation/can-you-trust-uv-long-term.md) lays out the licensing and fork options that hold whatever OpenAI decides. Marsh's answer is that trust gets settled by actions, not announcements, and open-sourcing pyx's engine is one of them.

## Learn More

- [Talk Python #552: Astral joins OpenAI](https://talkpython.fm/episodes/show/552/astral-joins-openai)
- [OpenAI to Acquire Astral](https://pydevtools.com/blog/openai-acquires-astral.md)
- [Can you trust uv long-term?](https://pydevtools.com/handbook/explanation/can-you-trust-uv-long-term.md)
- [What are wheel variants?](https://pydevtools.com/handbook/explanation/what-are-wheel-variants.md)
- [Introducing pyx (Astral blog, August 2025)](https://astral.sh/blog/introducing-pyx)
- [Simon Willison on pyx](https://simonwillison.net/2025/Aug/13/pyx/)