Training costs for top AI models are doubling every five months — what could actually bend that curve?

This explores which efficiency gains could slow the rise in frontier training costs, and how far each one gets you. Start with how fast costs are rising. Epoch AI's analysis of 41 frontier models found that amortized training costs grew about 2.4× a year since 2016, which means they double roughly every five months How fast are frontier AI training costs actually rising?. To slow that curve, an efficiency gain has to be large, and it has to target the part of the bill that actually dominates.

That second point changes which gains matter. Energy might look like the obvious target, but Epoch's cost model for a one-gigawatt data center puts servers at about 60% of annualized cost, around $5B a year. Energy comes to only about $0.6B What actually drives AI data center costs?. Cheaper power or better cooling helps only a little. What matters is getting more useful work out of each expensive chip before it has to be written off. The same picture shows up at the company level: compute makes up 54–62% of operating expenses at Anthropic, Minimax and Z.ai, while staff costs stay under 25% Do compute costs really dominate AI company budgets?. Whatever cuts compute per unit of capability cuts the bill.

The strongest efficiency number in the collection comes from inference (running a trained model), not training. After removing the effect of falling hardware prices and price competition, researchers estimate that better algorithms alone make inference about 3× more efficient each year, and independent energy measurements back that figure up How fast is AI inference actually becoming more efficient?. On paper, 3× a year is faster than 2.4× cost growth. But training costs keep rising anyway, which suggests labs spend their efficiency gains on bigger, more capable runs rather than on cheaper ones. A related lever is smaller models trained on better data. Occamy-1.0, a 35B-parameter model trained on execution-grounded data and long multi-step tasks, competes with much larger models at a fraction of the cost Does model efficiency matter more than peak capability for real work?. The lesson is that targeted training data can stand in for raw scale on some kinds of work.

The most speculative lever is AI automating AI research itself. Claims that automated research could squeeze four or five years of progress into one rest on unproven assumptions: that research results can be checked at scale, and that skill on small tasks carries over to research that matters Could automated AI research compress years of progress into months?. Even the benchmarks used to measure research agents raise questions. They score agents on doing better within a fixed evaluation budget, but that doesn't show whether real research costs per discovery go down Do fixed-budget efficiency gains translate to real research progress?.

One caveat. This collection is strong on cost accounting and inference efficiency but thin on training-specific techniques like data efficiency, sparse architectures or better optimizers, so it can't put a number on how much those slow training costs. The less obvious takeaway is that efficiency and cost aren't really fighting each other. Algorithms are improving faster than costs are rising, and costs are rising anyway.