Etched — AI supply-chain exposure

Etched is a chip-design startup building Sohu, an application-specific integrated circuit — an ASIC — hard-wired for one job: running transformer models, the architecture behind nearly every large language model. Rather than the flexible, general-purpose design of a GPU, it bakes the transformer directly into silicon, trading versatility for raw efficiency on that single workload. It designs the chip but does not manufacture it, and remains pre-volume.

Its structural bet is that the transformer is durable enough to be worth freezing into hardware — a wager that inference, not training, becomes the dominant and cost-sensitive workload. The model places it at the periphery because it is young, single-purpose and unproven at scale; if the bet lands, a fixed-function chip could undercut GPUs on the economics of serving models. If the architecture shifts, that same specialisation becomes the risk.