NVIDIA's RTX Spark architecture — the first laptop with over 100 GiB of GPU-accessible unified memory — shipped with Windows-only support. An Arch Linux package maintainer decided that was a suggestion, not a constraint. The result: a fully bootable Arch Linux ARM setup on the ASUS ProArt P16 with 128 GB RAM, 6,144 CUDA cores, and a 2 TB SSD, running local AI inference at parity with Windows configurations. The technical lift was non-trivial but not heroic. NVIDIA had already done kernel work for Ubuntu, providing a starting point. The developer incorporated that, fixed ACPI override allocation and framebuffer mapping for console output, patched hid-asus for keyboard support, wrote a machine driver for audio, and extended Arch's installer to handle the hardware packages. Build inputs are hash-pinned and signed with an official Arch Linux Package Maintainer key. This is reproducible work, not a weekend hack. The CPU benchmarks tell a clean story: Linux matches or slightly beats Windows. Single-thread 7-Zip compression sits within 1.2% of a comparable Windows HP OmniBook Ultra 16 with the same 20-core N1X chip. At 20 threads, Linux pulls ahead by 5% on compression and 2.7% on decompression. A separate ComputerBase Windows test on the same ProArt P16 model landed inside the Linux run's three-run range. The ARM silicon doesn't care about the OS. The memory story is where things get interesting. Linux reports 116.0 GiB as both system RAM and CUDA's total pool — the same physical memory accessible to both CPU and GPU workloads without rebooting. Windows splits it into dedicated GPU, shared, and CPU-only regions. The best Windows configuration reported by Frandroid reached 109.7 GiB for CUDA (including shared memory) with a 91 GiB dedicated reservation. That's 6.3 GiB less than Linux, and Windows requires a reboot to change the split between workloads. Local inference performance validates the setup. Running NVIDIA's Qwen3.8-Flash-Next-FP4 checkpoint through vLLM with B12X CUDA kernels, the laptop allocated nearly 90 GiB to model and KV cache. Code review tasks generated at 30.9 tokens/s median — matching Frandroid's Windows llama.cpp result of 29.4-31.4 tokens/s. A 10-minute sustained run completed 150 requests with zero failures, and median response times grew less than 1.6% from first ten to last ten repetitions. The broader pattern here is familiar: hardware vendors ship with Windows-only support, and the Linux community reverse-engineers compatibility within weeks. What's different is the stakes. Unified memory architectures that let local AI models access 100+ GiB without discrete GPU constraints represent a genuine capability shift for developers. Locking that behind a single OS — one that fragments the memory pool and requires reboots to reallocate — is a choice that serves Microsoft's platform control, not the hardware's potential. The developer's patches and build scripts are public. The work is signed, hash-pinned, and documented well enough for others to reproduce. This isn't a proof of concept; it's a bootable installer. The question now is whether NVIDIA and ASUS will upstream proper Linux support or leave the community to maintain parallel infrastructure indefinitely.
