2025–20267 chipsGlobal median ~$39K (2) · China median ~$18.6K (2)
| Region | Headline spec | Trend | Details | |||
|---|---|---|---|---|---|---|
| Nvidia B300 / GB300 (Blackwell Ultra)Blackwell Ultra · launched 2025-03 · Shipping | Nvidia | Global | FP4 Tensor Core (rack): 1,440 PFLOPS with sparsity / 1,080 PFLOPS dense (15 PFLOPS dense per GPU) | ~$53K2025-10-25 · analyst estimate · B300 card | ||
| Ascend 910CAscend · launched 2025-05 · Shipping | Huawei | China | ~800 TFLOPS (estimated) | ~$27.6Kfrom ¥200,000 at 2026 avg 7.25 CNY/USD2026 · street price · card | ||
| AMD Instinct MI355XCDNA 4 · launched 2025-06 · Shipping | AMD | Global | MXFP4 compute: 10.1 PFLOPS (dense) | ~$25K2026-08 · analyst estimate · card | ||
| Ascend 950PRAscend · launched 2026-03 · Shipping | Huawei | China | FP4 compute: 1.56 PFLOPS (Atlas 350 card level), ~2.8x NVIDIA H20 inference | ~$9.7Kfrom ¥70,000 at 2026 avg 7.25 CNY/USD2026-04 · street price · card | ||
| T-Head Zhenwu M890Zhenwu · launched 2026-05 · Limited availability | Alibaba T-Head (Pingtouge) | China | Memory: 144 GB | Not disclosed | ||
| T-Head PPU AI inference acceleratorZhenwu · launched 2025-09 · Shipping | Alibaba T-Head (Pingtouge) | China | Memory: 96 GB HBM2e | Not disclosed | ||
| Ascend 950Ascend · launched 2026-01 · Shipping | Huawei | China | FP8 compute: 1 PFLOPS | Not disclosed |
20246 chipsChina median ~$9K (3) · Global median ~$17.5K (2)
| Region | Headline spec | Trend | Details | |||
|---|---|---|---|---|---|---|
| Nvidia B200Blackwell · launched 2024-03 · Shipping | Nvidia | Global | FP4 Tensor Core: 18 PFLOPS with sparsity / 9 PFLOPS dense per GPU | ~$25K2026-08 · analyst estimate · GPU | ||
| Kunlunxin P800Kunlun · launched 2024 · Shipping | Kunlunxin (Baidu) | China | 345 TFLOPS | ~$13.2Kfrom ¥96,000 at 2026 avg 7.25 CNY/USD2026-08-27 · reseller listing · card | ||
| AMD Instinct MI325XCDNA 3 · launched 2024-10 · Shipping | AMD | Global | 81.7 TFLOPS vector / 163.4 TFLOPS matrix | ~$10K2026-08 · analyst estimate · card | ||
| MLU590 (思元590)MLU · launched 2024 · Shipping | Cambricon | China | ~256 TFLOPS (unofficial) | ~$9K2026-08 · analyst estimate · chip | ||
| Nvidia H20Hopper · launched 2024-03 · Discontinued | Nvidia | China | ~1 TFLOPS | ~$7.2Kfrom ¥52,000 at 2026 avg 7.25 CNY/USD2026-08-27 · street price · complete 8-card machine-month (H20-141G, Jinan carrier equipment room, 1-month deposit + 3 months prepaid) | ||
| Nvidia GB200 NVL72Blackwell · launched 2024-03 · Shipping | Nvidia | Global | FP4 Tensor Core (rack): 1,440 PFLOPS with sparsity / 720 PFLOPS dense (20 / 10 PFLOPS per GPU) | Not disclosed |
2022–202316 chipsChina median ~$7.2K (10) · Global median ~$17.5K (4)
| Region | Headline spec | Trend | Details | |||
|---|---|---|---|---|---|---|
| Nvidia H800Hopper · launched 2023-03 · Discontinued | Nvidia | China | ~1,979 TFLOPS with sparsity (roughly H100 SXM parity) | ~$35K2023-12 · street price · card | ||
| Nvidia H200 SXMHopper · launched 2023-11 · Shipping | Nvidia | Global | 34 TFLOPS | ~$31K2026-08 · street price · card | ||
| Nvidia H100 SXMHopper · launched 2022-03 · Shipping | Nvidia | Global | 34 TFLOPS | ~$25K2026-08 · street price · card | ||
| Ascend 910BAscend · launched 2023 · Shipping | Huawei | China | ~320 TFLOPS (dense); China Telecom Cloud lists 376 TFLOPS for 910B instances — figures vary by variant B1/B2/B3 | ~$13.9Kfrom ¥100,000 at 2024 avg 7.20 CNY/USD2024-05-24 · reported contract · card | ||
| MTT S4000MTT · launched 2023-12 · Shipping | Moore Threads | China | 25 TFLOPS | ~$10.9Kfrom ¥78,851 at 2026 avg 7.25 CNY/USD2026-08-25 · reseller listing · card | ||
| Nvidia A800 80GBAmpere · launched 2022-09 · Discontinued | Nvidia | China | 9.7 TFLOPS (same as A100) | ~$10K2023 · street price · card | ||
| AMD Instinct MI300XCDNA 3 · launched 2023-12 · Shipping | AMD | Global | 81.7 TFLOPS vector / 163.4 TFLOPS matrix | ~$10K2026-08 · analyst estimate · card | ||
| Hygon DCU DeepComputing II (深算二号)DeepComputing · launched 2023 · Shipping | Hygon | China | 192 TFLOPS | ~$7.8Kfrom ¥56,000 at 2025 avg 7.19 CNY/USD2025 · street price · card | ||
| MetaX C500 (曦云C500)XiYun (曦云) · launched 2023 · Shipping | MetaX (沐曦) | China | Compute positioning: FP32 to INT8 full-precision support; single-card compute targeting NVIDIA A100, exceeding H20 | ~$6.6Kfrom ¥47,500 at 2025 avg 7.19 CNY/USD2025-06-30 · reported contract · card (ASP) | ||
| Nvidia GeForce RTX 4090Ada Lovelace · launched 2022-10 · Discontinued · Gaming GPU | Nvidia | Global | 82.6 TFLOPS | ~$3.7Kfrom ¥26,000 at 2023 avg 7.08 CNY/USD2023-11 · street price · card (China) | ||
| Nvidia L20Ada Lovelace · launched 2023-11 · Shipping | Nvidia | China | 239 TFLOPS/TOPS (dense) | ~$3.4K2026-08 · reseller listing · card | ||
| MLU370-X8MLU · launched 2023 · Shipping | Cambricon | China | 256 TOPS | ~$2.1Kfrom ¥15,000 at 2026 avg 7.25 CNY/USD2026-08 · reseller listing · card | ||
| Atlas 300I DuoAtlas · launched 2022 · Shipping | Huawei | China | 280 TOPS | ~$1.4K2025-10-20 · analyst estimate · card | ||
| MTT S80MTT · launched 2022-11-11 · Shipping · Gaming GPU | Moore Threads | China | 14.7 TFLOPS (4096 MUSA cores @ 1.8 GHz) | ~$166from ¥1,195 at 2024 avg 7.20 CNY/USD2024-10 · street price · card | ||
| Biren BR104BR · launched 2022-08 · Export restricted | Biren Technology | China | ~512 TFLOPS (half of BR100) | Not disclosed | ||
| Biren BR100BR · launched 2022-08 · Export restricted | Biren Technology | China | 1024 TFLOPS dense / 2048 TFLOPS sparse | Not disclosed |
2021 and earlier7 chipsChina median ~$1.9K (3) · Global median ~$13K (2)
| Region | Headline spec | Trend | Details | |||
|---|---|---|---|---|---|---|
| Nvidia A100 80GB SXMAmpere · launched 2020-11 · Shipping | Nvidia | Global | 9.7 TFLOPS | ~$18K2026-08 · street price · card (80GB SXM) | ||
| Iluvatar CoreX TianGai 100 / BI seriesBI (天垓) · launched 2021-03 · Shipping | Iluvatar CoreX (天数智芯) | China | 147 TFLOPS | ~$11.3Kfrom ¥79,999 at 2023 avg 7.08 CNY/USD2023 · reseller listing · card | ||
| AMD Instinct MI250XCDNA 2 · launched 2021-11 · Shipping | AMD | Global | 47.9 TFLOPS (vector) / 95.7 TFLOPS (matrix) | ~$8K2026-08 · street price · card | ||
| Enflame Suiyuan/CloudBlazer DTU (current: S60, L600)CloudBlazer · launched 2021-07 · Shipping | Enflame Technology | China | 160 TFLOPS | ~$1.9Kfrom ¥14,000 at 2024 avg 7.20 CNY/USD2024-12-31 · reported contract · accelerator card/module (ASP) | ||
| Jingjia Micro JM9 seriesJM · launched 2021 · Shipping · Graphics GPU | Jingjia Micro (景嘉微) | China | ~2 TFLOPS-class, 1.8 GHz (positioned near GTX 1080; FP32 slightly below) | ~$61from ¥439 at 2026 avg 7.25 CNY/USD2026-07 · reseller listing · card | ||
| T-Head Hanguang 800 (含光800)Hanguang · launched 2019-09 · Limited availability | Alibaba T-Head (平头哥) | China | 820 TOPS peak | Not disclosed | ||
| Kunlunxin R480 / R300 (Kunlun Gen-2)Kunlun · launched 2021-08 · Shipping | Kunlunxin (Baidu) | China | 256 TOPS x8 modules | Not disclosed |
One table per generation within each category — similar chips rank against their contemporaries, and categories with different pricing units (cards, CPUs, memory stacks, market indices) never share a sort. All prices are shown in USD, converted from the source currency at that observation year's average exchange rate (the underlying figure is printed under every converted price). Click any row for full specs and the dated price history; every observation links to its source.
Every observation is dated and source-linked. Also see the CPU tables, memory, and edge AI pages, or the value map and price-gap overview.