Improvements in training cost efficiency typically lead AI companies to spend more, not less, on training models.

The claim is a general factual statement: when the cost of training a model at a given capability level falls (the cost curve 'shifts'), AI companies typically respond by raising total training spending rather than cutting it. TRUE: evidence up to the statement date shows the pattern holds in general. That means credible data (for example Epoch AI or similar compute and cost trend analyses, company disclosures, or industry reports) showing large, sustained algorithmic or hardware efficiency gains alongside rising frontier training budgets and compute at the leading labs, with little sign that labs typically cut training spend after efficiency gains. MOSTLY TRUE: the pattern holds for frontier labs generally, with minor exceptions or some imprecision in how widely it applies. MIXED: the evidence is split, with some major companies raising spend and a comparable number holding or cutting it after efficiency gains, or the pattern holds only in some periods. MOSTLY FALSE: spending rose, but the evidence shows the rise was mainly driven by things other than efficiency gains, or the pattern holds only for one or two firms and is stated as typical. FALSE: the evidence shows that efficiency gains have typically been followed by flat or lower training spending at AI companies.

Set before the research began.

1. E1 How much does it cost to train frontier AI models? Epoch AI · 3 June 2024

   “Our analysis reveals that the amortized hardware and energy cost for the final training run of frontier models has grown rapidly, at a rate of 2.4x per year since 2016 (95% CI: 2.0x to 3.1x).”

2. E2 Trends in GPU price-performance Epoch AI · 27 June 2022

   “Our higher point estimate for the rate of performance improvements amongst GPUs used for ML research could be explained by relevant labs spending more resources on procuring top GPUs over time.”

3. E3 Trends in GPU price-performance Epoch AI · 27 June 2022

   “we find that the amount of floating-point operations/second per $ (hereafter FLOP/s per $) doubles every ~2.5 years.”

4. E4 Trends in AI Epoch AI · 5 February 2026

   “The consequences are striking. Training costs are climbing by 3.5× annually.”

5. E5 Algorithmic progress in language models Epoch AI · 12 March 2024

   “We find that the level of compute needed to achieve a given level of performance has halved roughly every 8 months, with a 95% confidence interval of 5 to 14 months.”

Before the statement date, Epoch measured hardware efficiency gains, with GPU FLOP/s per dollar doubling about every 2.5 years [E3]. Epoch's later analysis of the period since 2016 finds that the cost of final training runs for frontier models grew about 2.4x per year [E1]. A separate Epoch study finds large algorithmic efficiency gains over the same era [E5]. So falling costs went together with rising frontier training spend, and no evidence shows labs cutting spend. However, the evidence covers frontier labs rather than AI companies in general. It also does not show that the efficiency gains caused the higher spending; E2 offers this only as a possible explanation.