The story is not the support for Voodoo 3, 4 and 5 cards or vintage Atari computers in itself, but the context in which it arrives. Linux kernel 7.3, expected in 2026, will include improvements for these platforms while elsewhere the project continues to drop older hardware drivers, overwhelmed by LLM-generated bug reports and patches for components almost impossible to use with modern versions of the mainline kernel.

It is a dynamic worth looking at closely: noise produced by automation is not neutral. When a maintainer receives dozens or hundreds of plausible but useless reports, the review cost rises and the ability to focus on real cases drops. The fact that Voodoo and Atari still receive improvements is not a contradiction, but an exception: someone has kept working on them with real understanding, probably because the code is isolated enough and tests can be run on physical hardware or reliable emulation.

For anyone evaluating self-hosted or on-premise deployment, the lesson is not about nostalgia. The mainline kernel is the foundation for distributions also used in industrial and lab environments. If driver tree maintenance becomes more selective, the incentive grows to stay on LTS branches or maintain vendor kernels with specific patches: a choice that reduces dependence on external contributions but shifts update and security costs onto internal teams. Software sovereignty is not only about licenses, but about managing the life cycle of components the mainline may decide to abandon as the signal-to-noise ratio worsens. For those weighing on-premise deployment choices, AI-RADAR offers on /llm-onpremise analytical frameworks to compare trade-offs between staying on mainline and maintaining a long-term stable kernel.

The link to LLMs is direct: producing a well-formatted patch has become trivial, but understanding whether that patch solves a real problem on physical hardware is another matter. The kernel is a project built on trust between maintainers and contributors; the arrival of synthetic contributions erodes that trust incrementally. No spectacular incident is needed: a steady stream of almost-correct changes is enough to raise cognitive load and push maintainers to close entire subsystems.

In this sense, Voodoo 3/4/5 and vintage Atari are the symptom of an ecosystem reacting. It is not necessarily good for everyone: every dropped driver makes obsolete machines that could still have a role in isolated networks, control systems or simple data acquisition workstations. But it is a reminder that free software is not an infinite warehouse: it has maintenance costs, and whoever generates automated noise is accelerating its downsizing.