Remove deprecated Fly GPUs docs

[kcmartin]

Summary of changes

GPUs became unavailable after August 1, so this removes the remaining GPU documentation.

Pages removed (the five listed as "Primary GPU Docs" in docs-tracking#241, all of which were left in place carrying deprecation warnings):

• gpus/index.html.md
• gpus/gpu-quickstart.html.markerb
• gpus/getting-started-gpus.html.md
• gpus/python-gpu-example.html.md
• kubernetes/using-gpus.html.markerb

Side nav: removed the Fly GPUs section and the "Use GPUs with FKS" link under Fly Kubernetes.

Stale references cleaned up on pages that remain, either because they described GPUs as an available product or because they carried an "as of 08/01/26" deprecation date that has now passed:

File	Change
about/pricing.html.markerb	Removed the GPU-enabled Machines section and the "Attached GPUs are charged separately" note
about/free-trial.html.md	Removed the GPU machines bullet
about/support.html.md	Removed the GPU Machines supported-product entry
reference/suspend-resume.html.md	Removed the "No GPU configured" requirement
getting-started/troubleshooting.html.md	8GB is now stated flat in both places, without the 50GB GPU exception
deep-dive/index.html.markerb	Cut the AI-on-GPUs framing from the intro
elixir/advanced-guides/clustering-from-home-to-your-app-in-fly.html.md	Removed the AI/ML development section and reworded the summary; the guide still stands on its distributed-app use case
blueprints/shared-nothing.html.markerb	Removed the "runs Whisper with Fly GPUs" aside
getting-started/launch-demo.html.markerb	Removed the "Try out Fly GPUs" link, the only live inbound link that would have 404'd

No redirects, matching how the enveloop and Upstash Kafka removals were handled.

Preview

Check that the Fly GPUs section is gone from the docs side nav, and that Fly Kubernetes no longer lists "Use GPUs with FKS".

Related Fly.io community and GitHub links

Part of the GPU deprecation work tracked in superfly/docs-tracking#241, which deferred these pages when the rest of the audit was completed in #2362.

Notes

volumes/overview.html.markerb keeps its volume placement paragraph, rewritten with a --vm-size=performance-4x example instead of --vm-gpu-kind=l40s. The compute hint is not GPU-specific: flyctl passes every --vm-* flag through as compute_requirements, and the scheduler uses it to pick a host with capacity for the Machine you plan to attach.

Deliberately left alone:

• machines/api/machines-resource.html.markerb and machines/flyctl/fly-machine-run.html.md still document gpu_kind/gpus and --vm-gpu-kind. docs-tracking#241 said to leave the API reference until deprecation, and that gate has now passed, but whether the API and flyctl still accept these is a platform question.
• flyctl/cmd/* will drop the flags on their own when the bot next syncs.
• partials/gpus/_ollama-create.erb, deep-dive/whisper.html.markerb, and python/do-more/add-ollama.html.markerb were deferred to a separate PR in docs-tracking#241. All are published: false or only reachable from a page that is, so the dead /docs/gpus/ link in the partial does not render on the live site.

Separately, volumes/overview.html.markerb does not mention that the API places a five-minute capacity hold on the hinted CPU and memory when a volume is created (rolled out April 2026). That is a real gap, but it is new feature documentation rather than GPU cleanup, so it is not in this PR.