The price of frontier-grade AI collapsed in 2025, faster than most forecasts allowed for. The cause was not any single lab. It was the shape of the contest. Two rivals who are level keep spending to stay level; let one pull clearly ahead and the trailing side stops trying while the leader eases off. Aghion et al., 2005 found this pattern in UK firm data and called it an inverted-U: innovation rises with competition, peaks when firms are neck and neck, then falls away. The US and China are about as evenly matched as the theory's best case, which puts them on the steep part of that curve.
The history in question is one of constraint leading to innovation. Export controls limited Chinese access to advanced chips, and rather than halting progress they redirected it. Bansemer & Miller, 2025 argue that the restrictions pushed DeepSeek toward algorithmic optimisations such as memory management and synthetic data. The final training run for its V3 model cost roughly USD 5.5 million on export-grade hardware, and its R1 reasoning model was released open-weight. The effect on price has been steep. Maslej et al., 2025 report that the inference cost of a system scoring at GPT-3.5 level fell from USD 20.00 per million tokens in November 2022 to USD 0.07 by October 2024, a more than 280-fold reduction, while open-weight models narrowed the gap to closed ones from 8% to 1.7% on some benchmarks in a single year.
How far the underlying gap has however really closed is harder to settle. Qian & Liu, 2025 build a multi-agent economic simulation of the two economies and find a pattern they summarise as the United States leading and China accelerating, driven by agent expansion and rapid capability catch-up. They note that on the MMLU benchmark the distance between the two countries' top models fell from 17.5 percentage points in 2023 to 0.3% by early 2024, though this is a modelled comparison rather than a measure of realised output, and the US retains the lead in depth and ecosystem maturity.
Whether the dividend persists depends on the rivalry staying close. Haag, 2025, in a Federal Reserve cross-country review, finds that the United States still outperforms across most measures of AI capacity while China remains competitive rather than level. The inverted-U is the warning embedded in the same frame that explains the gains: the innovation they rest on is a property of the contest remaining neck-and-neck, not of competition in the abstract. Should one side pull decisively ahead, the curve points downward.