Are custom AI chips like Google's TPUs built to stay useful longer than the typical four-year lifespan of top chips?
Do custom ASICs and new TPU generations extend chip lifespan beyond four years?
This explores whether newer, purpose-built AI chips (like Google's TPUs or other custom ASICs) stay useful for frontier AI work longer than the roughly four-year window seen for leading chips so far.
This explores whether custom-designed AI chips and newer TPU generations can stay in frontier use for more than about four years. The short answer is that this corpus has no direct evidence on ASICs or TPUs specifically. What it does have is the baseline those chips would be measured against, plus some signals about which way the trend is moving.
The baseline comes from Epoch AI's study of 149 frontier models. Leading chips from the V100 era onward stayed in frontier training for a median of 3.9 years, measured from release to their last use on a frontier run How long do leading AI chips stay in frontier training?. The useful surprise is that chip lifespans have already grown: designs from before the V100 lasted only about 2.6 years. In the same period the field settled on fewer chip designs. Post-V100 chips each trained about 9.3 frontier models on average, compared with 2.1 for earlier designs Did AI chip designs consolidate after the V100?. Those two findings likely go together. When labs build very large training clusters around one chip, they have strong reasons to keep using that hardware longer. This suggests chip lifespan depends partly on how a lab invests in hardware, not only on how fast new chips arrive. A custom ASIC deployed at scale inside one company could plausibly benefit from the same effect, though the corpus doesn't test this.
Two other findings pull in opposite directions. The first: high-bandwidth memory (HBM) rose from 52% to 63% of AI chip component spending between early 2024 and late 2025, driven by tight supply and rising prices Is memory becoming the dominant cost in AI chips?. When the costliest part of a new chip is hard to get, replacing hardware gets more expensive, which could stretch out how long existing chips stay in use. The second: the cost of reaching a given capability level drops roughly 5–10x per year, partly because of better hardware Does inference cost remain a persistent economic burden?. Each new chip generation that adds to that drop makes older chips less competitive, which shortens how long they stay at the frontier.
Keep in mind what the four-year figure measures: the last time a chip was used for frontier training. It is not the chip's whole working life. A chip that leaves frontier training may still be useful for other work. The corpus doesn't track that later life, and it has no data on TPUs or other custom chips. For a real answer you'd need lifespan data that breaks out custom chips from Nvidia GPUs, and the collection doesn't have it yet.
Sources 4 notes
Epoch AI's analysis of 149 frontier models shows V100-era and newer chips have a median frontier lifespan of 3.9 years from release to last use, though pre-V100 designs lasted only 2.6 years median.
Epoch AI's analysis of 149 frontier models shows post-V100 chip designs averaged 9.3 trained models each, versus 2.1 for pre-V100 designs, indicating a shift toward large-scale industrial training runs on narrower hardware choices.
Epoch AI's component cost model shows HBM's share of total AI chip spending rose 11 percentage points between Q1 2024 and Q4 2025, driven by price increases and supply tightness. HBM spend nearly tripled in dollar terms while logic held flat and packaging and auxiliary components declined.
Denain argues the inference cost burden is not persistent because the dollar cost to reach a given capability level drops rapidly each year through distillation, algorithmic efficiency gains, and hardware advances. He documents a 3x cost reduction over eight months on a math benchmark and projects 5–10x annual reductions, suggesting today's expensive capabilities become affordable within years.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The rising costs of training frontier AI models
- The Price of Progress: Price Performance and the Future of AI
- The plunging price of thought
- Leading AI chip designs are used for around four years in frontier training
- AI Compute Architecture and Evolution Trends
- Is a compute crunch coming?
- PostTrainBench: Can LLM Agents Automate LLM Post-Training?
- AI chip component costs: memory at 63%