AI on vintage computers is not a thought experiment. At Elyan Labs, transformer language models run today on a Nintendo 64, on PowerPC G4 and G5 Macs, on a 1998 AMD K6-2, and on an IBM POWER8 server — all CPU-only, all on hardware we physically own. This page is a tour of what running AI on retro hardware actually looks like.
AI on vintage computersretro hardware AI sovereign AICPU-only inference PowerPCN64
| Machine | Era | What runs on it |
|---|---|---|
| Nintendo 64 (VR4300, 93.75 MHz) | 1996 | 819K-parameter byte-level transformer, ~60 tokens/sec — see the N64 LLM |
| AMD K6-2 (Cobalt Qube 3) | 1998 | The same engine, ported to C89 and verified logic-exact against the N64 build |
| PowerPC G4 / G5 Macs | 1999–2005 | Quantized model inference using AltiVec — see PowerPC LLM |
| IBM POWER8 S824 | 2014 | Multi-billion-parameter models held entirely in RAM, VSX-accelerated — see POWER8 LLM |
The recipe is the same at every scale. First, size the model to the machine:
a nano-GPT in kilobytes for a console, a small quantized model for a G4, a large one
for a 512 GB server. Second, write (or use) dependency-free portable C — no Python
stack, no framework, nothing that won't compile on an old toolchain. Third, quantize
the weights so they fit in RAM and lean on integer math. Fourth, use whatever the
silicon gives you: the N64's CP0 cycle counter for sampling entropy, AltiVec
vec_perm on PowerPC, VSX on POWER8. The source for the console engine is
open at github.com/Scottcjn/legend-of-elya-n64,
and the practical steps are written up in
How to Run an LLM on Old Hardware.
A 93 MHz console will not replace a frontier model, and we don't claim it will. What small on-device models do well is bounded, local intelligence: game dialogue, classification, control logic, text generation with character. The larger vintage and exotic machines — G5s, POWER8 — run genuinely useful local models for chat and coding assistance. The capability scales with the machine; the principle doesn't change.
The deeper point is sovereignty. Inference that happens on silicon you own cannot be metered, rate-limited, deprecated, or switched off by a vendor. Old hardware proves the point in the starkest way: if language modeling runs on a 1996 console, then "you need the cloud" is a business model, not a law of physics. That argument is made in full in Proof of Physical AI, and the same philosophy drives RustChain, where verified vintage machines earn network rewards through proof of antiquity.
Is this emulation? No. Everything above runs natively on the original silicon. Emulated hardware is explicitly what our attestation work detects and excludes.
Can I try it? Yes — the N64 engine is open source, and the guides linked above walk through PowerPC and generic old-hardware setups.
By Elyan Labs. Every CPU has a voice.