Why did nearly every top AI model end up trained on just a handful of chip designs from one or two companies?

This explores why frontier AI training has narrowed onto a few chip designs instead of spreading across many. The clearest evidence in the collection describes the pattern more than its causes. Epoch AI's count of 149 frontier models shows a sharp change around NVIDIA's V100: each chip design released after it was used to train an average of 9.3 frontier models, against 2.1 for earlier designs Did AI chip designs consolidate after the V100?. Before the V100, frontier work was spread across many different pieces of hardware. After it, labs converged on a handful of designs and used each one heavily. Epoch reads this as the shift from research-scale experiments to industrial-scale training runs.

The second clue is how long those chips stay in use. Leading designs from the V100 era onward stay in frontier training for a median of 3.9 years, compared with 2.6 years for older designs How long do leading AI chips stay in frontier training?. So chips are not only shared by more models, they also last longer. Once a lab has built its clusters, software and engineering habits around one chip, that chip keeps earning its place for years. Fewer designs used for longer is what consolidation looks like once switching hardware gets expensive.

The cost numbers help explain why switching gets expensive. Amortized frontier training costs have grown about 2.4× per year since 2016, roughly doubling every five months How fast are frontier AI training costs actually rising?. The cost isn't only hardware, either. R&D staff make up 29–49% of total development cost for models like GPT-4 and Gemini Ultra How much do AI staff costs drive frontier model training?. Together these suggest a lot of what labs pay for is people's time and expertise tuned to particular hardware. When a single run costs tens of millions and your engineers already know one platform deeply, betting on an unfamiliar chip is a large risk. This is an inference from the cost data, not something the collection measures directly.

There is a possible counterweight. Branch-Train-MiX trains separate expert models in parallel with no synchronization between them, then merges them into a single mixture-of-experts model. It beats tightly synchronized training on the tradeoff between accuracy and efficiency Can asynchronous expert training beat synchronized distributed LLM training?. Today's frontier runs need huge, tightly connected clusters of identical chips. Methods that loosen that requirement could make it easier to train on mixed or less concentrated hardware. Nobody has yet shown this weakening vendor lock-in at frontier scale.

The collection has gaps here. It documents that consolidation happened and gives the cost pressures behind it. It doesn't directly cover the factors people usually name, such as software ecosystems like CUDA, supply constraints or vendor pricing power. Treat the explanation above as what the evidence supports, not a complete account.