Supply
Estimated AI Compute Shipped by Chip Designer
Quarterly shipments normalized to Nvidia H100-equivalent peak dense 8-bit compute
Quarterly Epoch AI estimates of peak dense 8-bit compute in tracked flagship AI accelerators sold or shipped, normalized to Nvidia H100 equivalents and grouped by designer. This gauges supply entering the hardware pipeline—not unit sales, installed capacity, utilization or company-reported revenue.
Estimated AI Compute Shipped by Chip Designer
- AMD
- Nvidia
As of 2026-06-30, Nvidia represents 3,953,823 H100-equivalent accelerators estimated H100-equivalent accelerators shipped.
Methodology
H100-equivalents compare peak dense 8-bit operations per second. They do not measure realized workload throughput, utilization, installed capacity, chip lifetime, or deployment date.
Keep source records whose start date is on or after 2023-01-01. Group records by the source start date, source end date, and chip designer. Sum chip-level H100-equivalent median estimates within each designer-period. Sum the chip-level 5th-percentile values and 95th-percentile values separately as transparent component-bound sums. Mark a designer-period incomplete when any contributing Epoch record is marked Incomplete. No dates or values were interpolated in this chart transformation. Epoch states that its upstream calendar-quarter records may themselves be interpolated from fiscal-quarter, half-year, or annual inputs by assuming a constant sales rate. The CSV preserves Epoch's published calendar-quarter boundaries without further interpolation.
The lower and upper columns are sums of component-level percentile estimates. They are not independently calculated joint-distribution percentiles and should be presented as indicative uncertainty bounds. Every row is an Epoch historical estimate. No row is labeled observed, and no forecast is included. The incomplete 2026-Q2 rows are retained as a separately flagged status rather than combined with complete periods without disclosure.
Do not interpret the cross-designer total as a consistently covered market series before 2024 because four tracked designers have no published 2023 history.
Frequently asked questions
What does one H100-equivalent mean?
It is the amount of peak dense 8-bit compute Epoch assigns to one Nvidia H100. A quarter with 1 million H100-equivalents represents shipped accelerators whose combined peak dense 8-bit operations per second equal about 1 million H100s.
Are these reported sales figures?
No. Epoch describes all figures as estimates grounded in financial reports, company disclosures, analyst estimates, media reports, price assumptions, and chip specifications. The chart therefore labels every value as a historical estimate.
Does the chart include forecasts?
No. It contains no forecast series. The latest available quarter, 2026-Q2, is an incomplete historical estimate and is explicitly flagged; the latest complete period is 2026-Q1.
Why do some designer series start in 2024?
Epoch's published history begins in 2022 for Nvidia, 2023 for Google, and 2024 for the other tracked designers. This chart starts in 2023 as requested and does not fill unavailable periods with zeros.
Are calendar-quarter values interpolated?
This chart adds no interpolation. Epoch documents that some of its source calendar quarters are upstream interpolations from fiscal-quarter, half-year, or annual information using a constant-rate assumption. Those published quarters are preserved as supplied.
Does shipped compute equal installed compute?
No. Epoch's figures generally reflect delivery or recognized revenue at the chip vendor. Server assembly, delivery to the final owner, installation, and operation can occur later.
Why does estimated shipped compute matter for the AI supply chain?
It puts accelerator generations on a common peak-compute basis, helping track the scale and mix of hardware entering the supply chain. It can indicate vendor concentration or diversification, but it does not show where chips were deployed or how intensively they are used.
Can the designer shares be read as complete market shares?
Not consistently across the full period. Nvidia and Google have 2023 history, while AMD, Amazon, Cambricon and Huawei begin in 2024. Missing earlier records mean the cross-designer total has incomplete coverage before 2024, and Epoch may revise its estimates.