It is not often that a Linux kernel development cycle shows so clearly the friction between two speeds: that of AI- and LLM-related patches and that of maintaining an architecture like ARM64. In the Linux 7.2 cycle, that friction produced a stark result: no real features landed on AArch64. For anyone considering ARM64 as a platform for self-hosted servers dedicated to model inference, this is not a maintainer's footnote but a risk indicator.

The response comes with Linux 7.3. The ARM Linux developers arrive more prepared and bring new ARM64 features, including BBML3 and NVIDIA Olympus workarounds. It is a change of pace that should be read carefully: the point is not just to recover lost time, but to show that the architecture can absorb contributions from the AI ecosystem without being drowned out by patch noise. The BBML3 feature set signals that enablement work has been channeled into a more selective review path, while the NVIDIA Olympus workarounds are a reminder that kernel support is not only about compute power.

For teams designing on-premise infrastructure for LLMs, the structural lesson is that the stable kernel branch is not a secondary detail. An empty cycle like Linux 7.2 means delays in enabling new instructions, supporting peripherals, and fixing hardware errata. On ARM64 platforms, where software control is often one of the reasons to prefer self-hosted over cloud, these delays translate into management costs and update constraints. It is not enough to wait for a feature: you need to check whether the kernel underlying a deployment will receive the necessary mitigations without forcing forks or backports.

The winners are ARM64 hardware vendors targeting servers and appliances for local inference: a kernel that resumes integrating features lowers support costs and shortens time to production. The losers, at least in the short term, are teams that planned upgrades on Linux 7.2 and found themselves on a branch without the expected features. The pressure of AI/LLM patches, if not governed, can displace maintenance work for architectures that are not at the center of the hype but become essential when moving from cloud training to local operations.

The NVIDIA Olympus workaround case adds a concrete element: an accelerator can have high specifications, but if operating system support requires fixes that arrive late, the real cost shifts to the team that has to maintain the fleet. In an on-premise context, where there is no provider applying fixes for you, that cost remains entirely with the organization. For those evaluating on-premise deployment, trade-offs need careful analysis: AI-RADAR offers analytical frameworks at /llm-onpremise, but the final decision depends on TCO, data sovereignty, and hardware support maturity.