Custom AI chips, compared
42 custom chips from 9 owners, 24 of them deployed. Microsoft Maia 200 leads stated FP8 at over 5,000 TFLOPS per chip, Google TPU 8i has the largest fixed HBM entry at 288 GB, and Broadcom is the named partner on 11. Every figure keeps its precision and links its source.
Last verified
| Alibaba T-Head Zhenwu 810E | Alibaba | Training and inference | In-house (T-Head) | — | 961 | 2.771 | — | — | — | — | — | — | — | — | — | Deployed | 2,0262 | 2026-11-11 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Alibaba T-Head Zhenwu M890 | Alibaba | Training and inference | In-house (T-Head) | — | 1441 | — | — | — | — | — | — | — | — | — | 1281 | Announced | 2,0261 | 2026-11-11 |
| Google TPU 8i | Inference | Broadcom | — | 2881 | 8.6011 | — | — | — | — | — | 10,1001 | — | — | 1,1521 | Announced | 2,0261 | 2026-11-11 | |
| Google TPU 8t | Training | MediaTek | — | 2161 | 6.5281 | — | — | — | — | — | 12,6001 | — | — | 9,6001 | Announced | 2,0261 | 2026-11-11 | |
| Meta MTIA 300 | Meta | Training | Broadcom | — | 2161 | 6.11 | — | — | 1,2001 | 6001 | — | — | — | — | 161 | Deployed | 2,0261 | 2026-11-11 |
| Meta MTIA 400 | Meta | Training and inference | Broadcom | — | 2881 | 9.21 | — | — | 6,0001 | 3,0001 | — | — | — | 12,0001 | 721 | Announced | 2,0261 | 2026-11-11 |
| Meta MTIA 450 | Meta | Inference | Broadcom | — | 2881 | 18.41 | — | — | 7,0001 | 3,5001 | — | — | — | 21,0001 | 721 | Announced | 2,0261 | 2026-11-11 |
| Meta MTIA 500 | Meta | Inference | Broadcom | — | — | 27.61 | — | — | 10,0001 | 5,0001 | — | — | — | 30,0001 | 721 | Announced | 2,0261 | 2026-11-11 |
| Microsoft Maia 200 | Microsoft | Inference | — | TSMC 3nm | 2161 | 71 | 5,0002 | — | — | — | — | 10,0002 | — | — | 6,1441 | Deployed | 2,0261 | 2026-11-11 |
| Microsoft Maia 300 | Microsoft | Inference | — | TSMC N2P | — | 181 | — | — | — | — | — | — | — | — | 961 | Announced | 2,0261 | 2026-11-11 |
| OpenAI Jalapeño | OpenAI | Inference | Broadcom | — | — | — | — | — | — | — | — | — | — | — | — | Sampling | 2,0261 | 2026-11-11 |
| AWS Graviton5 | Amazon Web Services | Server CPU | In-house (Annapurna Labs) | TSMC 3 nm | — | — | — | — | — | — | — | — | — | — | — | Deployed | 2,0251 | 2026-11-11 |
| AWS Trainium4 | Amazon Web Services | Training and inference | — | — | — | — | — | — | — | — | — | — | — | — | — | Announced | 2,0251 | 2026-11-11 |
| Baidu Kunlunxin M100 | Baidu | Inference | In-house (Kunlunxin) | — | — | — | — | — | — | — | — | — | — | — | — | Announced | 2,0251 | 2026-11-11 |
| Baidu Kunlunxin M300 | Baidu | Training and inference | In-house (Kunlunxin) | — | — | — | — | — | — | — | — | — | — | — | — | Announced | 2,0251 | 2026-11-11 |
| Google TPU v7 Ironwood | Training and inference | Broadcom | — | 1921 | 7.371 | 4,6141 | — | — | 2,3072 | — | — | — | — | 9,2161 | Deployed | 2,0251 | 2026-11-11 | |
| Huawei Ascend 950DT | Huawei | Training and inference | In-house (HiSilicon) | — | 1441 | 41 | 1,0001 | — | — | — | — | 2,0001 | — | — | 8,1921 | Sampling | 2,0251 | 2026-11-11 |
| Huawei Ascend 950PR | Huawei | Inference | In-house (HiSilicon) | — | — | — | 1,0001 | — | — | — | — | 2,0001 | — | — | — | Deployed | 2,0251 | 2026-11-11 |
| Huawei Ascend 960 | Huawei | Training and inference | In-house (HiSilicon) | — | — | — | 2,0001 | — | — | — | — | 4,0001 | — | — | 15,4881 | Announced | 2,0251 | 2026-11-11 |
| Huawei Ascend 970 | Huawei | Training and inference | — | — | — | — | — | — | — | — | — | — | — | — | — | Announced | 2,0251 | 2026-11-11 |
| Microsoft Cobalt 200 | Microsoft | Server CPU | — | TSMC N3P | — | — | — | — | — | — | — | — | — | — | — | Preview | 2,0252 | 2026-11-11 |
| Tesla AI6 | Tesla | Training and inference | — | Samsung SF2T (2 nm-class) | — | — | — | — | — | — | — | — | — | — | — | Announced | 2,0251 | 2026-11-11 |
| AWS Trainium3 | Amazon Web Services | Training and inference | Alchip; Marvell | 3 nm | 1443 | 4.93 | — | 2,5171 | — | 6711 | — | — | 2,5171 | — | 1442 | Deployed | 2,0245 | 2026-11-11 |
| Google Axion | Server CPU | — | — | — | — | — | — | — | — | — | — | — | — | — | Deployed | 2,0241 | 2026-11-11 | |
| Google TPU v6e Trillium | Training and inference | Broadcom | — | 321 | 1.6381 | — | — | — | 9181 | — | — | — | — | 2561 | Deployed | 2,0242 | 2026-11-11 | |
| Huawei Ascend 910C | Huawei | Training and inference | In-house (HiSilicon) | SMIC N+2 (7 nm-class) | 1281 | 3.21 | — | — | — | 7521 | — | — | — | — | 3841 | Deployed | 2,0244 | 2026-11-11 |
| Meta MTIA v2 | Meta | Inference | — | TSMC 5nm | 1281 | 0.2051 | — | — | — | 1771 | — | — | — | — | 721 | Deployed | 2,0241 | 2026-11-11 |
| Tesla AI5 | Tesla | Inference | — | Samsung SF2T (2 nm-class) for the first prototypes | — | — | — | — | — | — | — | — | — | — | — | Announced | 2,0244 | 2026-11-11 |
| AWS Graviton4 | Amazon Web Services | Server CPU | In-house (Annapurna Labs) | — | — | — | — | — | — | — | — | — | — | — | — | Deployed | 2,0231 | 2026-11-11 |
| AWS Trainium2 | Amazon Web Services | Training and inference | In-house (Annapurna Labs) | — | 961 | 2.91 | 1,2991 | — | — | 6671 | — | — | — | — | 642 | Deployed | 2,0234 | 2026-11-11 |
| Google TPU v5e | Training and inference | Broadcom | — | 161 | 0.81 | — | — | — | 1971 | — | — | — | — | 2561 | Deployed | 2,0232 | 2026-11-11 | |
| Google TPU v5p | Training and inference | Broadcom | — | 951 | 2.7651 | 4591 | — | — | 4591 | — | — | — | — | 8,9601 | Deployed | 2,0232 | 2026-11-11 | |
| Meta MTIA v1 | Meta | Inference | — | TSMC 7nm | 642 | 0.1762 | — | — | — | — | 51.21 | — | — | — | — | Deployed | 2,0231 | 2026-11-11 |
| Microsoft Cobalt 100 | Microsoft | Server CPU | In-house | — | — | — | — | — | — | — | — | — | — | — | — | Deployed | 2,0231 | 2026-11-11 |
| Microsoft Maia 100 | Microsoft | Training and inference | — | TSMC 5nm | 642 | — | — | — | — | — | — | — | — | — | — | Deployed | 2,0231 | 2026-11-11 |
| AWS Inferentia2 | Amazon Web Services | Inference | In-house (Annapurna Labs) | — | 321 | 0.881 | — | — | — | 1901 | — | — | — | — | — | Deployed | 2,0223 | 2026-11-11 |
| Huawei Ascend 910B | Huawei | Training and inference | In-house (HiSilicon) | SMIC 7 nm (N+2) | 641 | 1.62 | — | — | — | — | 4001 | — | — | — | — | Deployed | 2,0222 | 2026-11-11 |
| Google TPU v4 | Training and inference | Broadcom | — | 321 | 1.21 | — | — | — | 2751 | — | — | — | — | 4,0961 | Deployed | 2,0212 | 2026-11-11 | |
| Tesla Dojo D1 | Tesla | Training | — | TSMC 7 nm | — | — | — | — | — | — | — | — | — | — | — | Discontinued | 2,0211 | 2026-11-11 |
| AWS Trainium | Amazon Web Services | Training | In-house (Annapurna Labs) | — | 321 | 0.821 | — | — | — | 1901 | — | — | — | — | 161 | Deployed | 2,0203 | 2026-11-11 |
| AWS Inferentia | Amazon Web Services | Inference | In-house (Annapurna Labs) | — | 81 | 0.051 | — | — | — | 641 | — | — | — | — | 162 | Deployed | 2,0183 | 2026-11-11 |
| Baidu Kunlunxin P800 | Baidu | Training and inference | In-house (Kunlunxin) | — | 961 | — | — | — | — | — | 3451 | — | — | — | 2563 | Deployed | — | 2026-11-11 |
How this table is built
Full methodologyEvery value comes from the owner’s own announcement, docs or earnings call, a design partner’s statement, a technical paper or named trade press. Each FLOPS column holds one named precision, with sparsity and chip or module scope kept in the precision note. The small number next to a figure opens its source.
One record per chip generation, from Google, AWS, Meta, Microsoft, OpenAI, Tesla, Huawei, Alibaba and Baidu. Each row carries the date we last checked it, and the newest check was 11 Nov 2026. Spot a newer figure? Email hello@customsiliconlist.com with the source.
Free to reuse under CC BY 4.0 with a link to this page: CSV, JSON, llms.txt.
Questions buyers ask
What is a custom AI chip?
An accelerator that a cloud, platform or AI lab specifies for its own workloads instead of buying a merchant GPU. Google TPU, AWS Trainium, Meta MTIA and Microsoft Maia are the best-known; most are built with an ASIC partner such as Broadcom and fabbed at TSMC.
Which companies design custom AI chips with Broadcom?
Meta (MTIA 300 onward) and OpenAI (Jalapeño) name Broadcom themselves. For Google TPUs, Broadcom is reported by Reuters and CNBC and named by Broadcom’s CEO, who also says MediaTek builds TPU 8t. Broadcom handles silicon implementation, packaging and networking; the owner sets the architecture.
Which chip has the largest fixed HBM capacity entry?
Google TPU 8i, at 288 GB per chip. Capacity ranges stay in precision notes. Sort the memory column to compare fixed entries.
Can I rent these chips?
Google TPUs run in Google Cloud, Trainium and Inferentia in AWS EC2, Huawei Ascend in Huawei Cloud and Alibaba Zhenwu in Alibaba Cloud. Meta MTIA, Tesla AI5 and OpenAI Jalapeño stay internal. The "How to get it" field on each chip page names the service.
Why are some FLOPS cells empty?
Each column holds one named precision; the precision note records sparsity and chip or module scope. When an owner publishes only another precision, a sparse figure or a ratio, the cell stays empty and the precision note on the chip page quotes what was published.