Epoch AI: AI performance costs drop 13x yearly while infrastructure spending approaches $1T
A new Epoch AI report says the cost of achieving a fixed level of AI performance has fallen about 13-fold per year since 2023. The decline is roughly 47% per quarter, with costs dropping fastest right after a capability becomes state-of-the-art. Despite cheaper AI, U.S. and Chinese firms are projected to invest nearly $1 trillion in AI infrastructure this year.
Epoch AI's September 2026 report found that reaching a fixed AI performance level became about 13 times cheaper annually since 2023, with quarterly costs falling roughly 47%. On GPQA Diamond, OpenAI's o3 scored 75% in January 2025 for about $0.30 per question; GPT-5.6 Luna later matched that score for $0.0004, a 725-fold reduction.
Cost declines are steepest just after a capability reaches the frontier, averaging about 66% per quarter, then easing to roughly 32% after two years. AI's cost curve outpaces DNA sequencing, general computing, and lithium-ion batteries. Pure algorithmic gains alone improve efficiency about threefold yearly. US and Chinese hyperscalers and neoclouds may spend nearly $1 trillion on AI infrastructure in 2026, almost double the prior year.
Cheaper AI performance could widen access for smaller developers, researchers, and businesses that previously could not afford advanced models. It may also accelerate automation and data-intensive services, changing how some jobs are organized. At the same time, nearly $1 trillion in infrastructure spending could concentrate control and returns among large US and Chinese firms, while raising questions about energy use, capacity utilization, and whether demand grows quickly enough to justify the buildout. These shifts may affect workers, consumers, investors, and communities hosting data centers.