Planet Compute: A Verifiable AI Grid from Idle Browsers and Desktops Design, Early Measurements, and a Million-Device Simulation

Planet Compute is a volunteer-computing system that aims to serve AI inference from idle browsers and desktops; we report what is built, measured and only simulated. Built: one job format (WorkPacks) whose integer kernels give bit-identical results as WebAssembly in a browser and as native code in a Tauri 2 desktop program, which also runs whole chat and image prompts in sandboxed llama.cpp and stable-diffusion.cpp engines; commit-then-random-audit leases; and a SherpaHub that cuts matrix multiplies into packs checked by Freivalds’ method, which finished two split jobs of up to 262,144 units with no wrong answer on one operator server, though at only 1.5% of the kernel’s speed, limited by round trips. Measured offline on three NVIDIA GPUs, not yet in the program: diffusion and LLM audits costing 3.3–8.9% of the work. Analysed and simulated: we show that over home links such packs are bounded by upload bandwidth, argue for cutting the work rather than the model and size packs by a square-root law; derive the exact collusion probability of quorum validation, under which a naive “first three matching answers” rule accepts 5.8% wrong packs at 20% colluders (6.1% in simulation); show that spare replicas, network-distinct quorums and retroactive invalidation remove them; and simulate the orchestrator at a million devices, where, with 2% ordinary cheaters, in the final designs no wrong answer reached a consumer uncorrected, and a low-cost design cut the running cost about elevenfold at the price of latency. No volunteer outside the operator’s servers has yet taken part.