Training stays on GPUs; cloud giants move their own inference to in-house chips — resolves 2027-01-31

The debate: whether NVIDIA's GPUs will be replaced by the custom chips cloud giants design for themselves (ASICs: chips built for one job, cheaper and more power-efficient but not general-purpose).

Why it matters: the answer decides where hundreds of billions of dollars of chip purchases go over the next few years, and how much pricing power NVIDIA keeps.

Who is on each side: one camp says custom chips will take the whole market and the other says GPUs cannot be replaced for training; our call is that the two split by job and coexist.

GPUs and ASICs diverge rather than displace each other: training stays on GPUs, while hyperscalers move their own inference to ASICs.

Jouppi et al. — Google — arXiv (ISCA 2017)
“This paper evaluates a custom ASIC---called a…”

Reviewed in H1 2027: if GPUs still hold more than 80% of frontier training runs and the ASIC share of hyperscaler inference keeps rising, the call stands.

What would overturn it

Any ASIC winning a major third-party frontier training contract, or the ASIC share of inference flat for two consecutive years.

Correction to the settlement rule

Added on 2026-09-13 to the conditions that would overturn this call: besides the original two, if a third party with no Google partnership measures ASICs ahead on performance per dollar after TorchTPU's public release, while adoption by buyers other than Google itself still does not move, the cause of the coexistence-and-divergence pattern is recorded as supply terms rather than price-performance. The settlement condition was not changed.