How long do leading AI chips stay in frontier training?

Epoch AI, in a 2025-03-05 data insight by Luke Emberson, Ben Snodin, and David Owen, measures what it calls the "frontier lifespan" of commercial AI chip designs — the time between a chip's public release and the publication date of the last "frontier AI model" (one of the top 10 largest models by training compute at the time) trained on that chip. Restricting the analysis to the 5 chip designs "at least as new as the NVIDIA V100," Epoch AI finds "the median lifespan from release until final use for frontier training is 3.9 years," with lifespans for those chips "rang[ing] from 2.3 to 4.5 years." "Considering all chip designs at once," including the 7 released before the V100, the median falls to 2.7 years, since "older AI chip designs generally saw shorter lifespans": a median of 2.6 years for the pre-V100 group, spanning 0.6 to 3.7 years.

The estimate comes from joining Epoch AI's Notable Models dataset (publication dates, training compute, and chip design per model) with its Machine Learning Hardware dataset (chip release dates), restricted to commercial chip designs and excluding consumer parts "like NVIDIA's GTX/RTX series." This leaves "149 frontier AI models, trained by 14 unique AI chip designs," though the chip design is only identified for "64 of the 88 frontier AI models" in scope. Chip designs used within the prior 7 months are dropped from the lifespan count — Epoch AI's chosen threshold, the 95th percentile gap observed between successive frontier runs on the same chip — because they may still see further use. Epoch AI is explicit that the result is "a weak upper bound," since model publication typically lags training completion "by weeks or months," and that the figure assumes chips like the NVIDIA A100 or Google TPU v4 are now retired from frontier training, an assumption it calls "likely to be correct" but not certain.

The figure sits alongside Did AI chip designs consolidate after the V100?, which traces why the lifespan lengthened after the V100: frontier training concentrated onto fewer chip designs, each reused for more model generations. It also shares a resource lens with Does pretraining loss predict performance after reinforcement learning? — both describe a compute constraint, chip lifespan here and the pretraining-to-RL budget split there, that bounds how much capability gain a given training generation can produce before the next one supersedes it.

The dataset is small — 5 chip designs at or after the V100, 7 before it — so a single long- or short-lived outlier chip can move the median noticeably, and Epoch AI does not explain why consolidation onto fewer, longer-lived designs happened (vendor lock-in, software-stack switching costs, or simply fewer competing architectures reaching frontier scale). The excerpt establishes a historical cadence, not a forecast: it says nothing about whether newer chip classes (custom ASICs, more TPU generations) will extend or shorten the 3.9-year figure going forward. The clearest reading it supports is that a chip bought for frontier-scale training today has, on recent history, roughly a four-year window before it stops appearing in the largest training runs — useful for amortization planning, but Epoch AI's own estimate rather than a fixed hardware constant.