Epoch: Inference Capacity Grew 4.6x vs 2.6x for Compute

The estimated agentic inference capacity of AI chips sold to date grew 4.6 times between Q2 2025 and Q2 2026, while their peak 8-bit computing power grew 2.6 times, Epoch AI said on 18 November, using its new Inference Performance Index (IPI).

Our number, from Epoch’s own table: Google TPUs went from 5.5% to 15.9% of that agentic inference capacity over the year, while Nvidia’s share fell from 84.2% to 76.8% even as Nvidia supplied most of the growth (our arithmetic).

Business Pill · GPUs Explained

An evergreen Business Engineer explainer on why AI runs on GPUs. It teaches the general idea only and does not discuss Epoch’s index.

The key insight: By Epoch AI’s new Inference Performance Index, the capacity of AI chips sold to date to run agentic workloads grew 4.6 times in a year against 2.6 times for peak compute; Blackwell supplied 74% of the gain, and Google TPUs rose from 5.5% to 15.9% of capacity by our arithmetic from Epoch’s table.

What Epoch Measured

Epoch says the IPI “compares hardware on a fixed set of workloads with long, cached contexts”. It applies version 1 IPI scores, in GB300 equivalents, to chip-sales estimates from its AI Chip Sales dataset, and the index “accounts for memory, inter-chip communication, and lower-precision computation, including both input processing and output generation”.

Capacity is the cumulative number of chips sold in each category multiplied by its per-chip IPI score. By that measure, capacity rose from about 1.45 million GB300 equivalents in Q2 2025 to about 6.67 million in Q2 2026. Epoch gives 90% uncertainty intervals of 4.2 to 5.1 times for inference growth and 2.5 to 2.7 times for peak compute.

Epoch AI estimates the agentic inference capacity of AI chips sold to date rose 4.58 times from Q2 2025 to Q2
Epoch AI estimates the agentic inference capacity of AI chips sold to date rose 4.58 times from Q2 2025 to Q2 2026, against 2.59 times for peak 8-bit compute. Data: Epoch AI.

What Was Posted on X

Epoch AI posted on 18 November 2026 that agentic inference capacity grew 4.6 times from Q2 2025 to Q2 2026, against 2.6 times for compute, by its Inference Performance Index, with Nvidia’s Blackwell generations accounting for 74% of the added capacity.

Blackwell Supplied Most of the Gain

Epoch says Nvidia’s Blackwell and Blackwell Ultra families “account for 74% of the added capacity”, with GB300 systems delivering “roughly 8× the inference capacity of H200s per GPU”. By Q2 2026 they were about 73% of total agentic inference capacity, against 50% of total peak 8-bit computing power.

That ties to Nvidia’s own numbers this week: its next platform, Rubin, made nearly $20 billion in its first quarter, per its call. Epoch’s table runs to Q2 2026 and does not yet include Rubin.

Where the Share Shifted

In Epoch’s data, Google’s TPU v7 went from zero to about 905,000 GB300 equivalents by Q2 2026, about 13.6% of the total on its own. Amazon’s Trainium fell from about 7.1% to 4.4% of capacity, and AMD GPUs held near 2.1% (our arithmetic from Epoch’s table).

Epoch says inference capacity per unit of peak compute rose by about 77% over the year, which it attributes broadly to “hardware–model co-design”. Our division of the two growth multiples, 4.58 over 2.59, gives the same 1.77 times.

Why Growth May Slow

Chips sold in Q2 2026 added 3.3 times the inference capacity of those sold in Q2 2025, less than the 4.6 times growth in the total, so Epoch writes “we expect total growth to slow”. It notes that sustaining 4.6 times a year would require annual additions equal to 3.6 times the capacity at the start of each year.

As we read it alongside Epoch’s separate report this week, in which no model met its staff standard on real Epoch work, the rise in capacity measures what can be served, not what the models can do.

The Structural Read

As we read Epoch’s table, raw compute understates how fast serving capacity is growing, which matters for anyone pricing inference by FLOP/s.

As we read the share shift, Nvidia still supplied most of the new capacity, but Google’s TPU v7, from zero to 13.6% of capacity, made the second-largest share gain of any chip family after Blackwell Ultra (our arithmetic).

Epoch AI, 18 Nov 2026

“Continued hardware–model co-design should nevertheless keep inference capacity growing faster than peak computing power.”

Three Implications

FOR CHIP BUYERS Epoch says GB300 systems deliver roughly 8x the inference capacity of H200s per GPU on its workloads.

FOR CAPACITY PLANNING Epoch expects total growth to slow, since Q2 2026 additions were 3.3x those of Q2 2025.

FOR VENDORS By our arithmetic from Epoch’s table, Google TPUs reached 15.9% of capacity while Nvidia fell to 76.8%.

The Business Engineer Lens

This story maps onto the Business Engineer framework The AI Capex Map & The State of AI Hyperscalers.

The analysis puts it this way: “This is not a continuation of the cloud capex cycle.”

As we read Epoch’s index, capacity per unit of compute rising 77% in a year is one way the spending converts into served AI. The analysis predates Epoch’s data insight and does not discuss it.

What Is Not Established

These are Epoch’s estimates from chip-sales data and a fixed workload basket; Epoch says the fitted rates “describe historical growth, not forecasts”. The shipment data combine B200 with GB200 and B300 with GB300. Vendor shares above are shares of estimated capacity, not of revenue.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The AI Capex Map & The State of AI Hyperscalers

A Business Engineer map of who is spending on AI compute, how much, and who carries the demand.

Read the Map of AI →

The Bottom Line

By Epoch’s index, the capacity to serve agentic AI grew 4.6 times in a year against 2.6 times for raw compute; Blackwell supplied 74% of the gain, and Google TPUs nearly tripled their share to 15.9% (our arithmetic).

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. All capacity and compute figures, method notes and quotes are from Epoch AI’s 18 November 2026 data insight by Jason Li and its data table, read in full, and Epoch’s X post. Vendor shares of capacity (Google TPUs 5.5% to 15.9%, Nvidia 84.2% to 76.8%, Trainium 7.1% to 4.4%, AMD about 2.1%) and the 1.77 times ratio are our arithmetic from Epoch’s table. The Rubin figure is from our earlier report on Nvidia’s call. We did not contact Epoch.

Sources: Epoch AI, ‘Agentic inference capacity grew 4.6x in a year, versus 2.6x for compute capacity’, 18 Nov 2026 · Epoch AI on X, 18 Nov 2026

Scroll to Top
FourWeekMBA