H100 equivalent explained: how AI compute is counted, and why the yardstick is rougher than it looks

When someone says a company has a million H100s, they usually do not mean a million H100s. They mean the equivalent. H100-equivalent, shortened to H100e, is the unit Epoch AI uses to compare chips from different makers, and it appears in most serious reporting on the AI compute race. It is useful. It is also rougher than the tidy number suggests.

What H100e means

Epoch's AI Chip Sales documentation defines H100e capacity as the total computing power measured in terms of the equivalent number of Nvidia H100 GPUs. The metric is peak dense 8-bit operations, in FP8 or INT8 format, chosen because 8-bit formats are now widely used for inference and for training at major AI labs.

The calculation is simple. Divide a chip's peak dense 8-bit operations by the H100's specification, then multiply by the number of chips. Epoch's own example: because a TPUv7 can perform about 2.3 times as many operations as an H100, one million TPUv7s equal about 2.3 million H100e. The H100 became the reference because it was the most widely used AI chip in 2023 and 2024.

Why a common unit exists

Without it, a Nvidia Blackwell, a Google TPU, an AMD Instinct and a Huawei Ascend cannot be added up. Epoch's page for Huawei's Horinger data center shows the conversion at work. An Ascend 910C has a theoretical peak of about 1.504 × 10¹⁵ operations per second at INT8, roughly 0.76 of an H100. An Nvidia B200 or B300 delivers about 5 × 10¹⁵, roughly 2.5 H100e.

With that conversion, Epoch can say Horinger holds about 81,000 H100e, with a 90% interval of 36,000 to 141,000, against roughly 1.1 million for Colossus 2. We covered the finding in our report on Huawei's Horinger data center.

What the latest totals say

Epoch's October 29, 2026 data insight, by Isabel Juniewicz and Josh You, estimates that new global AI computing capacity is doubling every 9 months. Quarterly additions grew from 827,914 H100e in Q1 2024 to 6,327,416 in Q2 2026. The fitted trend is 2.5 times per year, with a 90% interval of 2.3 to 2.7, and a doubling time of 9.1 months. The analysis uses median estimates as of October 19, 2026 and covers six designers: Nvidia, Google, AMD, Amazon, Huawei and Cambricon.

Growth is slowing from the 2023 peak. Quarterly additions grew about 6 times in 2023 and about 2.5 times per year since 2024. Year-over-year growth in quarterly additions fell from 2.8 times in Q2 2025 to 2.2 times in Q2 2026, and cumulative capacity grew 2.6 times from Q2 2025 to Q2 2026 against 3.5 times a year earlier. The authors write that "Some evidence suggests further deceleration in growth over this period, but it is not statistically significant." Their p-value on the quarter-over-quarter slowdown is 0.26.

What the number leaves out

Epoch is candid about the limits. Its documentation calls H100e an imperfect proxy based on paper specifications. Memory and bandwidth, utilization and software maturity, and how chips are networked into clusters can all change real-world performance. A cluster of 100,000 chips that cannot be linked tightly will not behave like one training machine, and Epoch makes that point about Horinger, where compute may be split across clusters.

There are other gaps. The totals track chips when they are sold, not when they are operational, and chips can take several more months to come online. Omitted makers include Intel, Meta, Microsoft, Cerebras and Groq, and non-data-center GPUs; Epoch writes that "We believe these volumes are minor and would not meaningfully change this trend." The confidence intervals reflect only the trend fit, not uncertainty in the underlying sales estimates, which come from company disclosures, analyst estimates and media reports.

Efficiency is another blind spot. Two sites with the same H100e can draw very different power, which is why Epoch pairs the metric with megawatts. The Ascend 910C draws about 1.56 kilowatts of IT power per chip, while a B200 is about 2.5 times more energy-efficient.

How to read a claim

When a headline cites H100e, ask four questions. Which chips went into the sum, and how were they converted? Is the capacity in service, or merely shipped? Can it work as one cluster? And who produced the estimate, and what is its range? Our explainer on AI compute costs and capex covers how those chips translate into training budgets.

An H100e figure with an interval, a date and a named estimator is a good number. A bare "equivalent to a million H100s" in a press release is a marketing unit.

A worked example

Take Epoch's Horinger figures. It estimates about 105,000 Ascend 910C chips at roughly 0.76 H100e each. Simple multiplication gives about 80,000 H100e from the 910C alone. Epoch's published total of about 81,000 reflects its own chip counts, the 910B and other adjustments, so do not expect the arithmetic to reproduce it exactly. The point is the shape of the calculation: chip count, times a per-chip ratio from paper specifications, with a stated range. If a source gives you a total and none of those three inputs, you cannot check it.

Our take

H100e is the best common yardstick we have, and Epoch deserves credit for documenting its method and its limits. Use it to compare orders of magnitude, such as 81,000 against 1.1 million. Do not use it to rank two sites within 20% of each other, and do not treat a single date as the whole story, since the chips sold in one quarter are the compute online in the next.