Hardware for local AI inference: real market forming

The AMD Ryzen AI Halo dev kit at $4,000 and the OpenWrt One router at the other end of the price spectrum both point to the same thing: people are buying hardware to run compute locally, for different reasons. The AMD kit thread noted that a Bosgame device at $2,799 does the same thing for less, and that the Halo does not add meaningfully over the existing Strix Halo processor available since spring 2025. The market is fragmenting around price points.

The through-line connecting this to the small models thread and the Ternlight WASM thread is that local inference is not a single product category. It spans in-browser WASM models, on-device consumer hardware, and workstation-class AI dev kits. The use cases are different but the direction is the same: less dependence on cloud endpoints.

The OpenWrt One thread had a useful reality check. Commenters asked who the target audience actually is for a router with only two Ethernet ports and no Wi-Fi 6E. The honest answer is a small but serious group of people who want open, auditable hardware. That is not a mass market, but it is a real one.