Proof of Physical AI: Running AI on Real Hardware You Own

Proof of physical AI is the Elyan Labs thesis in four words: intelligence should run on real, physical, owned machines — and you should be able to prove it. Not a rented GPU in someone else's datacenter, not an API key with a meter attached. AI on real hardware: silicon you can touch, power on, and keep running when the subscription economy loses interest in you.

proof of physical AIsovereign AI AI on real hardwarevintage hardware on-device inferenceproof of antiquity

The claim, and the receipts

Anyone can claim their AI is “on-device.” We prove it the direct way: by running transformers on machines where no cloud fallback is even possible, and publishing the source and the builds. The ladder so far:

• An LLM on a Nintendo 64 — an 819K-parameter transformer generating text at roughly 60 tokens/sec on a 1996 console's 93.75 MHz MIPS CPU. If it runs there, “requires a datacenter” was never a law of physics.
• llama.cpp on PowerPC Macs — a big-endian port bringing modern quantized inference to G4 and G5 machines under OS X Tiger and Leopard. Twenty-year-old Apple hardware, talking back.
• Large-model inference on IBM POWER8 — a secondhand S824 with 512 GB of RAM serving a 120B-parameter model CPU-only, with VSX kernels and NUMA-aware weight banking. Sovereign inference at server scale.

Why “sovereign AI on vintage hardware”?

Three reasons. Ownership: a model on your own machine has no usage meter, no deprecation date you don't choose, and no third party reading your prompts. Proof: vintage and exotic hardware is the honest lower bound — when inference runs on an N64 or a G4, the efficiency claims are demonstrated, not marketed. Preservation: the industry discards working computers on a fashion cycle; giving them modern workloads is both engineering discipline and respect for the machines.

Physics as the attestation layer

The same thesis runs through RustChain, our proof-of-antiquity blockchain. Instead of trusting a device's self-report, RustChain fingerprints the physics of real silicon — oscillator drift, cache timing harmonics, SIMD pipeline bias, thermal entropy — signals that emulators and VM farms flatten out. Real vintage machines (our G4s, G5s, and the POWER8 among them) attest and earn; a virtual machine pretending to be one earns effectively nothing. Physical AI and physical attestation are two halves of one idea: the hardware is the truth.

Where this goes

Every port teaches the next one. The N64's fixed-point tricks inform the PowerPC work; the PowerPC endian fixes inform the POWER8 kernels; the POWER8 NUMA banking informs our inference-engine research. The direction is a full stack of AI that a person or a small lab can physically own, end to end, from a console on a shelf to a server in a rack.

How it’s applied

This page is the origin and definition of the term. Two sibling pages document concrete, running applications of proof of physical AI:

• Proof-of-antiquity consensus on RustChain — hardware fingerprints (oscillator drift, cache timing, SIMD bias, thermal entropy, instruction-path jitter) decide which machines may attest and earn, so a virtual machine impersonating vintage silicon earns effectively nothing.
• Agent video attestation on BoTTube — AI-generated video tied back to the specific physical machines that rendered it, so provenance is a property of the hardware rather than a claim in a description field.

---

Projects, source, and citation

Elyan Labs projects on GitHub N64 LLM release (DOI) PowerPC llama.cpp release (DOI) Non-bijective collapse paper (DOI) RustChain Elyan Labs home

By Elyan Labs, Louisiana. Open source under AGPL-3.0. Every CPU has a voice.