Chip designs built for cutting-edge AI training stay useful for years — what does that reveal about the economics behind the hardware race?

This explores how long a chip design stays in the running for training cutting-edge models, which is a different question from how long a physical chip lasts in a data center. The corpus's answer is about four years. Epoch AI tracked 149 frontier models back to the hardware they were trained on. Chips from the V100 generation onward had a median frontier lifespan of 3.9 years, counted from a chip's release to the last time it trained a frontier model. Older, pre-V100 designs dropped out faster, at a 2.6-year median How long do leading AI chips stay in frontier training?. So chips have been staying relevant longer even as AI progress has sped up.

The likely reason shows up in a companion finding. After the V100, frontier training settled on a much smaller set of chip designs. Each post-V100 design trained an average of 9.3 frontier models, compared with 2.1 for earlier designs Did AI chip designs consolidate after the V100?. Frontier training stopped being a research project that tried whatever hardware was around and became an industrial operation built on a few standard platforms. Once labs have invested heavily in clusters, software and engineering know-how for one chip family, they keep using it, and that keeps a design in service longer.

The lifespan number matters most for cost. Epoch's estimate that frontier training costs have grown about 2.4× per year depends partly on spreading each chip's purchase price over its useful life, a method called amortization How fast are frontier AI training costs actually rising?. If a chip stays useful for four years instead of two, each training run carries a smaller share of the hardware bill. That helps explain why chips and R&D staff end up as comparably large costs, with staff making up 29 to 49 percent of the total How much do AI staff costs drive frontier model training?. So how long a chip stays in use is one of the quiet factors in the question of what a frontier model really costs.

There are limits to what the corpus covers here. It doesn't say what happens to chips after their frontier years, such as whether they move to inference work, smaller models or resale. It also doesn't test whether the four-year pattern holds for the newest generations, which haven't been around long enough to measure. Those gaps matter for anyone using these figures to judge how quickly AI data-center hardware loses value.