How fast are frontier AI training costs actually rising?

Epoch AI's cost model finds that the amortized cost to train the most compute-intensive ("frontier") AI models has grown at a rate of 2.4× per year since 2016, with a 90% confidence interval of 2.0× to 2.9×. The most expensive publicly announced training runs to date are OpenAI's GPT-4 at $40M and Google's Gemini Ultra at $30M. The discussion section projects that "if the current trend of 2.4x per year growth continues, then the amortized cost of frontier training runs will exceed one billion dollars by 2027" — a threshold the paper notes Anthropic CEO Dario Amodei has suggested could arrive even sooner, having said frontier developers "are likely to spend close to a billion dollars on a single training run this year, and up to ten billion-dollar training runs in the next two years."

The estimate rests on two complementary approaches that agree closely. A hardware-and-energy approach amortizes AI accelerator chip, server, and networking capital costs — depreciated at r = 0.14 orders of magnitude per year, based on the growth rate of ML GPU price-performance — plus hardware energy consumption. A cloud-price approach instead uses historical cloud rental rates, and "shows a similar growth rate (2.5× per year with a 90% CI of 2.1× to 3.1×)." Both draw on the Notable AI Models database of 796 models, filtered to 41 models that ranked in the top 10 of training compute at release between October 2015 and December 2023, cross-referenced against a hardware-price dataset of 142 entries (52 purchase prices, 90 cloud rental prices) spanning 24 hardware models.

This trend sits alongside other accounts of why AI progress costs what it does. Where Does data really create lasting competitive advantage for startups? describes a cost pressure that erodes the value of what is spent on, this paper's trend describes a cost pressure — compute and hardware — whose price keeps climbing with no offsetting decline in value to the firms buying it. The paper's framing of investment "driving AI progress" also bears on Are AI feedback loops strong enough to sustain recursive self-improvement?: that note treats capability growth as bounded by self-improvement elasticities, while this paper treats it as bounded instead by who can afford the compute. The specific composition behind the 2.4× figure — which spending categories are rising and by how much — is detailed in How much do AI staff costs drive frontier model training?.

The estimate is a model, not a direct measurement: the paper notes its methods "neglect several costs that are potentially significant, including the data center infrastructure apart from the training cluster, and the acquisition of data for model training," and that results carry "uncertainties due to factors such as hardware depreciation rates and pricing dynamics." The 41-model sample covers only models at the computing frontier at release, so the trend "may not generalize to the broader AI landscape, or specific AI domains," and rapid hardware or software efficiency gains could bend the curve in ways "difficult to predict from historical data." If the trend does hold, the paper's own inference is that only "the most well-funded organizations will be able to finance frontier AI models" — a concentration effect it ties to open questions about coordination and oversight rather than treating as settled.