Ranked by compute delivered per watt — what decides how much AI a megawatt of grid capacity actually buys.
Microsoft Maia 200 leads with 1.4 x 10^13 OP/s per W, ahead of NVIDIA B300 (Blackwell Ultra) (1.3 x 10^13 OP/s per W) and NVIDIA GB300 (Blackwell Ultra) (1.1 x 10^13 OP/s per W).
Efficiency has improved by roughly a third per year across generations. It compounds with the power constraint: at a fixed grid connection, a more efficient chip is directly more compute, which is why efficiency gains show up in data-center capacity rather than in electricity bills.
| # | Name | Manufacturer | Released | Efficiency |
|---|---|---|---|---|
| 01 | Microsoft Maia 200 | Microsoft | Jan 26, 2026 | 1.4 x 10^13 |
| 02 | NVIDIA B300 (Blackwell Ultra) | NVIDIA | Aug 22, 2025 | 1.3 x 10^13 |
| 03 | NVIDIA GB300 (Blackwell Ultra) | NVIDIA | Aug 22, 2025 | 1.1 x 10^13 |
| 04 | AMD Instinct MI350X | AMD | Jun 12, 2025 | 9.2 x 10^12 |
| 05 | NVIDIA B200 | NVIDIA | Nov 15, 2024 | 9 x 10^12 |
| 06 | NVIDIA GB200 | NVIDIA | Feb 15, 2025 | 8.3 x 10^12 |
| 07 | AMD Instinct MI355X | AMD | Jun 12, 2025 | 7.2 x 10^12 |
| 08 | NVIDIA L4 | NVIDIA | Mar 21, 2023 | 6.7 x 10^12 |
| 09 | Google TPU v6e Trillium | May 14, 2024 | 4.8 x 10^12 | |
| 10 | Google TPU v7 Ironwood | Nov 6, 2025 | 4.8 x 10^12 | |
| 11 | Meta MTIA v1 (MTIA 100) | Meta | May 18, 2023 | 4.1 x 10^12 |
| 12 | Meta MTIA v2 (MTIA 200) | Meta | Apr 10, 2024 | 3.9 x 10^12 |
| 13 | Biren BR100 | Biren | Aug 22, 2022 | 3.7 x 10^12 |
| 14 | Cambricon MLU270-S4 | Cambricon | Jun 20, 2019 | 3.7 x 10^12 |
| 15 | Amazon Trainium3 | Amazon AWS | Dec 2, 2025 | 3.6 x 10^12 |
| 16 | GroqChip LPU v1 | — | Oct 15, 2022 | 3.5 x 10^12 |
| 17 | AMD Instinct MI300X | AMD | Dec 6, 2023 | 3.5 x 10^12 |
| 18 | NVIDIA RTX Pro 6000 Blackwell | NVIDIA | Mar 18, 2025 | 3.4 x 10^12 |
| 19 | T-Head Hanguang 800 | Alibaba | Sep 25, 2019 | 3 x 10^12 |
| 20 | NVIDIA H200 SXM | NVIDIA | Nov 18, 2024 | 2.8 x 10^12 |
| 21 | NVIDIA GH100 | NVIDIA | Mar 21, 2023 | 2.8 x 10^12 |
| 22 | NVIDIA H100 SXM5 80GB | NVIDIA | Sep 20, 2022 | 2.8 x 10^12 |
| 23 | NVIDIA H800 SXM5 | NVIDIA | Mar 21, 2023 | 2.8 x 10^12 |
| 24 | NVIDIA GH200 | NVIDIA | Aug 8, 2023 | 2.8 x 10^12 |
| 25 | AMD Instinct MI325X | AMD | Oct 10, 2024 | 2.6 x 10^12 |
| 26 | Amazon Trainium2 | Amazon AWS | Dec 3, 2024 | 2.6 x 10^12 |
| 27 | AMD Instinct MI300A | AMD | Jan 4, 2023 | 2.6 x 10^12 |
| 28 | Cambricon MLU370-S4 | Cambricon | Nov 3, 2021 | 2.6 x 10^12 |
| 29 | Cambricon MLU370-S8 | Cambricon | Nov 3, 2021 | 2.6 x 10^12 |
| 30 | Cambricon MLU370-S4/S8 | Cambricon | Nov 3, 2021 | 2.6 x 10^12 |
| 31 | Cambricon Siyuan 590 | Cambricon | Oct 31, 2024 | 2.6 x 10^12 |
| 32 | NVIDIA A800 PCIe 80 GB | NVIDIA | Nov 3, 2022 | 2.5 x 10^12 |
| 33 | NVIDIA H100 PCIe | NVIDIA | Sep 20, 2022 | 2.2 x 10^12 |
| 34 | Huawei Ascend 910C | Huawei | Oct 15, 2024 | 2.1 x 10^12 |
| 35 | NVIDIA H100 NVL | NVIDIA | Mar 14, 2024 | 2.1 x 10^12 |
| 36 | NVIDIA A100 PCIe | NVIDIA | Jun 22, 2020 | 2.1 x 10^12 |
| 37 | Huawei Ascend 910 (320 TFLOPS) | Huawei | Aug 23, 2019 | 2.1 x 10^12 |
| 38 | Huawei Ascend 310 | Huawei | Oct 10, 2018 | 2 x 10^12 |
| 39 | Baidu Kunlun RG800 | Kunlunxin Baidu | Jan 1, 2023 | 2 x 10^12 |
| 40 | Intel Habana Gaudi3 | Intel | Sep 24, 2024 | 1.9 x 10^12 |
| 41 | Google TPU v5e | Aug 29, 2023 | 1.7 x 10^12 | |
| 42 | Baidu Kunlun II | Kunlunxin Baidu | Aug 18, 2021 | 1.7 x 10^12 |
| 43 | Cambricon MLU370-X4 | Cambricon | Nov 3, 2021 | 1.7 x 10^12 |
| 44 | Baidu Kunlun R100 (75W) | Kunlunxin Baidu | Jan 1, 2022 | 1.7 x 10^12 |
| 45 | Baidu Kunlun I | Kunlunxin Baidu | Jan 1, 2019 | 1.7 x 10^12 |
| 46 | Cambricon MLU270-F4 | Cambricon | Jun 20, 2019 | 1.7 x 10^12 |
| 47 | Baidu Kunlun K200 | Kunlunxin Baidu | Jan 1, 2020 | 1.7 x 10^12 |
| 48 | Baidu Kunlun R200 (150W) | Kunlunxin Baidu | Jan 1, 2022 | 1.7 x 10^12 |
| 49 | Google TPU v5p | Dec 6, 2023 | 1.7 x 10^12 | |
| 50 | Baidu Kunlun R100 (100W) | Kunlunxin Baidu | Jan 1, 2022 | 1.7 x 10^12 |
