AI Training Compute Doubles Every Five Months
New research shows frontier models require significantly more processing power as training durations and chip counts scale.
Updated on Oct. 15, 2026 in Artificial Intelligence
Live Poll
Is the current pace of growth in AI computing power sustainable for the long term?
Frontier AI model training compute is growing at a rate of 5x annually since 2020, with total compute requirements doubling every 5.2 months. Research from Epoch AI, recently visualized by The New York Times, tracks this trajectory alongside a newly released database of global chip infrastructure.
Why it matters
Understanding the rapid escalation in training compute provides a quantitative baseline for the resources required to scale future frontier AI models. This growth trend, if maintained, significantly impacts the industrial demand for high-bandwidth memory and high-performance silicon.
Frontier models currently double their training compute every 5.2 months, driven by hardware quantity increases of 1.69x per year, longer training durations at 1.53x per year, and hardware performance gains of 1.41x per year.
The players
Epoch AI
A research organization focused on forecasting AI development and analyzing the technical requirements of frontier models.
The New York Times
A global media organization that produced a visualization of Epoch AI data regarding AI training compute.
The details
Epoch AI measures these metrics using floating-point operations—standard mathematical calculations used to quantify computer processing power. The researchers attribute total compute growth to the compounding effect of deploying more hardware, extending the duration of model training, and improvements in the underlying semiconductor performance. The Epoch AI chip users explorer catalogs 86 specific sites, which together account for 44% of global AI compute capacity.
Timeline
2010 marks the beginning of the long-term compute growth measurement period.
2020 marks the beginning of the frontier model growth measurement period.
Mid-2026 saw the global AI chip count reach an estimated 20 million units.
September 2026 marked the official launch of the Epoch AI chip users explorer.
October 2, 2026, saw the publication of an assessment report on HBM capacity.
The Tech Race
This data updates the long-term compute scaling laws established by previous Epoch AI research by isolating the specific contributions of hardware performance, training time, and cluster size. It benchmarks current industry trajectory against the 2010-2020 growth period.
The availability of high-bandwidth memory (HBM) is projected to dictate the ceiling for concurrent frontier-model AI agents through 2027. Developers and researchers can anticipate these supply-side capacity constraints to influence the feasibility of large-scale agentic workflows over the next two years.
The takeaway
The global count of AI chips is projected to reach 200 million by the end of 2028, up from 20 million in mid-2026. Readers should monitor the HBM supply chain reports to see if hardware production keeps pace with the demand for concurrent frontier model operations.
Further reading
For more on how computational power influences model capabilities, visit Artificial Intelligence.
Source note: This article includes information reported by Crypto Briefing.
Live Poll
Is the current pace of growth in AI computing power sustainable for the long term?
