AI chip sales
Our open database of AI chip sales, using financial reports, company disclosures, and more to estimate compute, power usage, and spending over time for a wide variety of AI chips.
Графики и аналитика по этим данным →
ai_chip_sales_chip_types.csv
32 строк · 14 колонок · Обновлено 2026-11-09
| Name | ML Hardware (linked) | Designer | Primarily for Chinese market | TDP (W) (from ML Hardware (linked)) | 16-bit FLOP/s | 8-bit OP/s | Notes | Memory | Memory bandwidth | H100e | Cost per chip (approx.) | Release date | Unofficial estimated TDP (W) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Ascend 950PR | Huawei Ascend 950PR | Huawei | True | 1000000000000000 | 1600000000000 | 0.5053057099545225 | |||||||
| Ascend 950DT | Huawei Ascend 950DT | Huawei | True | 973000000000000 | 144000000000 | 4000000000000 | 0.49166245578575035 | ||||||
| TPU v3 | Google TPU v3 | 450 | 123000000000000 | 123000000000000 | 32000000000 | 900000000000 | 0.06215260232440627 | 3600.0 | 2018-05-18 | ||||
| TPU v4i | Google TPU v4i | 175 | 138000000000000 | 138000000000000 | 8000000000 | 614000000000 | 0.0697321879737241 | 2020-01-01 | |||||
| A100 | NVIDIA A100 SXM4 80 GB | Nvidia | 400 | 312000000000000 | 624000000000000 | 80000000000 | 2039000000000 | 0.31531076301162203 | 12000.0 | 2020-11-16 | |||
| TPU v4 | Google TPU v4 | 340 | 275000000000000 | 275000000000000 | 32000000000 | 1200000000000 | 0.13895907023749368 | 3900.0 | 2021-05-20 | ||||
| Instinct MI250X | AMD Radeon Instinct MI250X | AMD | 500 | 383000000000000 | 383000000000000 | 128000000000 | 3280000000000 | 0.1935320869125821 | 10000.0 | 2021-11-08 | |||
| Ascend 910B | Huawei Ascend 910B | Huawei | True | 400 | 376000000000000 | In terms of cost: "distributors are selling the Huawei 910B at above 120,000 yuan per card, while its eight-card server equivalent starts at 1.3-1.5 million yuan per server" per Reuters. https://www.reuters.com/technology/nvidia-cuts-china-prices-huawei-chip-fight-sources-say-2024-05-24/ This is $16.5k at $0.138 USD per yuan in May 2024. Possible this is a relatively high street price compared to average (bulk) purchases. Rounding to 15k. | 32000000000 | 1228000000000 | 0.18999494694290045 | 15000.0 | 2022-03-17 | ||
| A800 | NVIDIA A800 SXM | Nvidia | True | 400 | 311800000000000 | 624000000000000 | Chinese-spec version of A100 | 80000000000 | 2040000000000 | 0.31531076301162203 | 12000.0 | 2022-08-11 | |
| H100 | NVIDIA H100 SXM5 80GB | Nvidia | 700 | 989500000000000 | 1979000000000000 | 80000000000 | 3350000000000 | 1.0 | 25000.0 | 2022-09-20 | |||
| Trainium1 | Amazon Trainium1 | Amazon | 190000000000000 | 420000000000000 | Trainium1 (and the very similar Inferentia2 chip) does not have an official TDP spec. To estimate, note that AI chips generally become 40% more efficient per year: https://epoch.ai/data-insights/ml-hardware-energy-efficiency Trainium1/Inferentia2 and Trainium2 were separated by two years, suggesting a ~2x efficiency improvement based on trends. Trainium2 is 500W and 667 TFLOP/s in 16-bit FLOP/s. Trainium1 is 190 TFLOP/s in 16-bit (note that the Trainium1 and Trainium2 have a slightly different ratio of 16-bit to 8-bit op/s; I'm using 16-bit which is likely more representative for Trainium1). So Trainium1 TDP theoretically would be roughly 190/667 (compute ratio) * 500 W (Trainium2 wattage) * 2x (efficiency penalty for older chip) = 284W, rounded up to 300W. | 32000000000 | 820000000000 | 0.21222839818089945 | 2022-10-15 | 300.0 | |||
| Instinct MI300A | AMD Instinct MI300A | AMD | 760 | 980600000000000 | 1961200000000000 | 128000000000 | 5300000000000 | 0.9910055583628095 | 6700.0 | 2023-01-04 |
ai_chip_sales_organizations.csv
12 строк · 4 колонок · Обновлено 2026-11-09
| Name | Country | Type | Chip types |
|---|---|---|---|
| Nvidia | United States of America | Chip designer | H100,A800,H800,H20,A100,B200,B300,H200,Rubin |
| United States of America | Chip designer,Compute owner | TPU v5e,TPU v5p,TPU v6e,TPU v7,TPU v4,TPU v3,TPU v4i,TPU v8i | |
| Huawei | China | Chip designer | Ascend 910B,Ascend 910C,Ascend 950PR,Ascend 950DT |
| Amazon | United States of America | Chip designer,Compute owner | Trainium2,Trainium1,Trainium3 |
| Microsoft | United States of America | Compute owner | |
| AMD | United States of America | Chip designer | Instinct MI250X,Instinct MI300A,Instinct MI308X,Instinct MI300X,Instinct MI325X,Instinct MI350X,Instinct MI355X |
| Meta | United States of America | Compute owner | |
| CoreWeave | United States of America | Compute owner | |
| Oracle | United States of America | Compute owner | |
| Other (ex-Big 4 hyperscalers) | |||
| Cambricon | China | Chip designer | Siyuan 590 |
| Chip manufacturer |
ai_chip_sales_timelines_by_chip.csv
171 строк · 24 колонок · Обновлено 2026-11-09
| Name | Chip manufacturer | Start date | End date | Incomplete | Number of Units | Number of Units (5th percentile) | Number of Units (95th percentile) | Compute estimate in H100e (median) | H100e (5th percentile) | H100e (95th percentile) | Source / Link | Notes | Chip type | Last Modified By | Last Modified | Chip type (linked) | Chip TDP (W) | Total TDP (W) | Total TDP (W) (5th percentile) | Total TDP (W) (95th percentile) | Chip Cost (Linked) | Chip Cost (USD) | Cost Estimate (USD) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Q1 2022 - A100 | Nvidia | 2022-01-01 | 2022-03-31 | 212120 | 163481 | 276445 | 66883.0 | 51547.0 | 87166.0 | Estimates generated on: 08-27-2026 15:06 | A100 | Joshua You | 2026-08-27 19:09:28+00:00 | A100 | 400 | 84848000 | 65392400.0 | 110578000 | 12000 | 12000 | 2545440000 | ||
| Q2 2022 - A100 | Nvidia | 2022-04-01 | 2022-06-30 | 226673 | 174697 | 295411 | 71472.0 | 55083.0 | 93146.0 | Estimates generated on: 08-27-2026 15:06 | A100 | Joshua You | 2026-08-27 19:09:28+00:00 | A100 | 400 | 90669200 | 69878800.0 | 118164400 | 12000 | 12000 | 2720076000 | ||
| Q3 2022 - A100 | Nvidia | 2022-07-01 | 2022-09-30 | 215732 | 166265 | 281153 | 68022.0 | 52425.0 | 88650.0 | Estimates generated on: 08-27-2026 15:06 | A100 | Joshua You | 2026-08-27 19:09:28+00:00 | A100 | 400 | 86292800 | 66506000.0 | 112461200 | 12000 | 12000 | 2588784000 | ||
| Q3 2022 - H100/H200 | Nvidia | 2022-07-01 | 2022-09-30 | 6119 | 5126 | 7199 | 6119.0 | 5126.0 | 7199.0 | Estimates generated on: 08-27-2026 15:06 | H100/H200 | Joshua You | 2026-08-27 19:09:28+00:00 | H100 | 700 | 4283300 | 3588200.0 | 5039300 | 25000 | 25000 | 152975000 | ||
| Q4 2022 - TPU v4i | 2022-10-01 | 2022-12-31 | True | 20588 | 12685 | 31294 | 1435.0 | 884.0 | 2182.0 | Incomplete: based on Broadcom fiscal quarters beginning 10/31/2022. Estimates generated on: 09-30-2026 14:07 | TPU v4i | Joshua You | 2026-09-30 18:11:45+00:00 | TPU v4i | 175 | 3602900 | 2219875.0 | 5476450 | 2322 | 47805336 | |||
| Q4 2022 - TPU v4 | 2022-10-01 | 2022-12-31 | True | 66441 | 48215 | 87695 | 9232.0 | 6699.0 | 12186.0 | Incomplete: based on Broadcom fiscal quarters beginning 10/31/2022. Estimates generated on: 09-30-2026 14:07 | TPU v4 | Joshua You | 2026-09-30 18:11:45+00:00 | TPU v4 | 340 | 22589940 | 16393100.0 | 29816300 | 3900 | 3389 | 225168549 | ||
| Q4 2022 - TPU v5e | 2022-10-01 | 2022-12-31 | True | 15678 | 9878 | 23460 | 3113.0 | 1961.0 | 4658.0 | Incomplete: based on Broadcom fiscal quarters beginning 10/31/2022. Estimates generated on: 09-30-2026 14:07 | TPU v5e | Joshua You | 2026-09-30 18:11:45+00:00 | TPU v5e | 225 | 3527550 | 2222550.0 | 5278500 | 2900 | 2704 | 42393312 | ||
| Q4 2022 - A100 | Nvidia | 2022-10-01 | 2022-12-31 | 113699 | 87627 | 148177 | 35850.0 | 27629.0 | 46721.0 | Estimates generated on: 08-27-2026 15:06 | A100 | Joshua You | 2026-08-27 19:09:28+00:00 | A100 | 400 | 45479600 | 35050800.0 | 59270800 | 12000 | 12000 | 1364388000 | ||
| Q4 2022 - H100/H200 | Nvidia | 2022-10-01 | 2022-12-31 | 32481 | 27212 | 38218 | 32481.0 | 27212.0 | 38218.0 | Estimates generated on: 08-27-2026 15:06 | H100/H200 | Joshua You | 2026-08-27 19:09:28+00:00 | H100 | 700 | 22736700 | 19048400.0 | 26752600 | 25000 | 25000 | 812025000 | ||
| Q4 2022 - A800 | Nvidia | 2022-10-01 | 2022-12-31 | 28132 | 21517 | 36492 | 8870.0 | 6784.0 | 11506.0 | Estimates generated on: 08-27-2026 15:06 | A800 | Joshua You | 2026-08-27 19:09:28+00:00 | A800 | 400 | 11252800 | 8606800.0 | 14596800 | 12000 | 12000 | 337584000 | ||
| Q1 2023 - TPU v4i | 2023-01-01 | 2023-03-31 | 9629 | 5933 | 14637 | 671.0 | 413.0 | 1020.0 | Estimates generated on: 09-30-2026 14:07 | TPU v4i | Joshua You | 2026-09-30 18:11:45+00:00 | TPU v4i | 175 | 1685075 | 1038275.0 | 2561475 | 2322 | 22358538 | ||||
| Q1 2023 - TPU v4 | 2023-01-01 | 2023-03-31 | 119390 | 85080 | 160064 | 16590.0 | 11822.0 | 22242.0 | Estimates generated on: 09-30-2026 14:07 | TPU v4 | Joshua You | 2026-09-30 18:11:45+00:00 | TPU v4 | 340 | 40592600 | 28927200.0 | 54421760 | 3900 | 3389 | 404612710 |