Falling AI costs get pinned on cheaper chips — but how much of that drop is really just competitors racing to undercut each other?

This explores whether a hardware price deflator, meaning an adjustment that strips out falling chip and server prices, can reliably separate technical progress from market effects in AI costs. Straight answer: the collection has no note that builds or tests such a deflator. What it does have is evidence for why that separation is hard, and why a hardware-only adjustment would probably miss most of what's going on.

The most useful starting point is Epoch AI's finding that the price of a given level of AI performance falls fastest right after a capability first reaches state of the art: about 66% per quarter at first, slowing to about 32% per quarter within two years Do AI performance prices fall faster when capabilities debut?. Epoch explains this as a market effect. A new capability starts out with temporary monopoly pricing, and competition wears that premium down over time. That matters here because the steepest part of the decline may be prices coming down from a markup, not costs coming down. A deflator built from hardware prices alone would not see this. The markup sits in what the model provider charges, not in what the GPUs cost.

On the technical side, Denain lists three separate drivers behind the 5–10x yearly drop in cost per capability level: distillation, algorithmic efficiency, and hardware improvements Does inference cost remain a persistent economic burden?. Hardware is only one of the three. So 'technical' is not the same as 'hardware.' A hardware deflator would count software and model-compression gains as if they were something else, possibly as market effects. Even within hardware, the gains depend on the setup. MobileLLM shows that on memory-limited phones, recomputing a block of the model is cheaper than fetching its weights from memory Does recomputing weights cost less than moving them on mobile?. Which resource is the bottleneck changes from one deployment to the next, so a single hardware price index will only roughly track real efficiency.

There is a reason hardware still deserves attention. Epoch's cost model of a one-gigawatt data center finds servers make up about 60% of annualized cost, while energy is only about 7% What actually drives AI data center costs?. Hardware prices clearly move the supply-side cost floor. But that floor is not the price customers pay, and the gap between the two is exactly where market effects show up.

A practical lesson comes from evaluation research. Because inference prices keep falling, benchmarks should report token counts along with dollar costs, so readers can recalculate costs as prices change Should benchmarks report token counts instead of frozen costs?. The same reasoning helps with your question. To separate technology from market, track physical quantities (tokens, compute, memory traffic) apart from prices, rather than relying on one deflator to untangle them afterward. If you want rigorous treatment of deflator methods, such as quality-adjusted price indexes for chips, this collection won't have it. That's a gap, and the economics literature is the better place to look.