AI
AI inference capacity surges 4.6× in a year as Blackwell chips dominate
Analysis of chip shipments shows agentic inference capacity skyrocketing through the Hopper-to-Blackwell transition, though growth rates are expected to decelerate.
Between Q2 2025 and Q2 2026, the total agentic inference capacity available across AI chips sold to date expanded at a dramatic pace—growing 4.6 times over the course of just one year. This acceleration far outpaced gains in raw computing power, which rose only 2.6×, underscoring how architectural improvements and hardware-model co-design are becoming the driver of inference capability.
The Blackwell transition reshapes the market#
The growth surge coincides with the rapid deployment of Blackwell and Blackwell Ultra chips, which now command roughly 73% of total agentic inference capacity despite accounting for a smaller 50% share of peak 8-bit computing power. Chips sold in Q2 2026 alone added 3.3× the inference capacity of those sold in the prior-year quarter—the highest quarterly addition rate seen.
Fitting historical trends across the full span from Q1 2024 through Q2 2026 suggests that inference capacity has been doubling roughly every 4.3 months when accounting for the Blackwell transition window. By contrast, peak compute has doubled every 9.2 months. These numbers reflect what has already happened; they are not forecasts.
Hardware and software working in tandem#
A key finding is that inference capacity per unit of peak compute improved by about 77% over the year—a gain that comes not from faster arithmetic alone but from multiple sources: larger model-serving domains, adoption of lower-precision computation, and architectural shifts that extract more utility from the same hardware. The comparison between successive GPU generations illustrates this trend: modeled inference capacity rises 8× from the H200 to the GB300, but that figure drops to 2.2× when projecting from the GB300 to the next-generation Vera Rubin VR200.
What comes next#
Sustaining the 4.6× annual growth rate would require new chip additions equal to 3.6× the starting capacity each year—a threshold that becomes harder to meet as the installed base grows. Moreover, the Hopper-to-Blackwell shift captured some of the largest efficiency gains available. Epoch's analysis suggests that future generations will yield more modest co-design improvements, though faster shipment growth could partially offset this trend.
The uncertainty analysis, which tests sensitivity to shipment quantities and hardware configurations, narrows the range of outcomes to between 4.2× and 5.1× for inference growth, confirming that the broad direction is robust even if absolute figures shift.

