Why did the biggest AI labs stop spreading training across many chip types and settle on just a few?

This explores why frontier AI training, the work of building the most capable models, came to rely on a smaller set of chip designs after NVIDIA's V100. The pattern itself is clear. Epoch AI looked at 149 frontier models and found that each chip design from the V100 onward trained an average of 9.3 of them. Earlier designs averaged 2.1 Did AI chip designs consolidate after the V100?. Epoch's own explanation is a shift toward large-scale industrial training runs. Training a frontier model stopped looking like a lab experiment you could run on whatever hardware was nearby. It became more like a factory process built around a few proven designs.

One sign of that change is how long chips stay in use. V100-era and newer chips have a median frontier lifespan of about 3.9 years, from release to the last frontier model trained on them. Earlier designs lasted about 2.6 years How long do leading AI chips stay in frontier training?. Concentration and longevity go together. When a design stays useful for four years, more models get trained on it. When labs commit to fewer designs, each one has more reason to stay in service. A chip generation now works more like long-lived infrastructure than a part you swap out quickly.

Cost helps explain why labs would make that commitment. Epoch's cost breakdown for GPT-3, OPT-175B, GPT-4 and Gemini Ultra puts accelerator chips at tens of millions of dollars per model. R&D staff are the other big cost, at roughly 29 to 49 percent of the total How much do AI staff costs drive frontier model training?. That second number matters more than it first appears. If people are nearly half the bill, the engineering know-how built up around a particular chip (software, cluster setup, debugging habits) is a large investment. Switching designs means paying for much of that again. The corpus doesn't state this link directly, so treat it as a reasonable reading of the cost data rather than an established finding.

The library is thin on the deeper causes. It has nothing on NVIDIA's software ecosystem, supply constraints, or how chip vendors compete, and those are the explanations people usually give. There is also a hint that this concentration may not be permanent. One note finds that a 3B-parameter model, trained with a carefully designed post-training pipeline, matched much larger systems on verifiable math and coding tasks Can small models match frontier reasoning without massive scale?. If more frontier capability comes from training recipes rather than raw compute, the pressure to standardize on the biggest chip deployments could weaken. Don't read too much into it: that result covers only tasks with checkable answers.