Key takeaways
- Top pick: Nvidia, best for: AI labs and enterprises training and running large models.
- Also near the top: AMD, best for: Cloud providers and AI labs that want a second GPU supplier; and Google (TPU), best for: AI labs training and serving large models in the cloud.
- How it was ranked: 10 deep tech companies compared. Each verdict is below.
Best AI Chip Companies: Top 10 Ranked. Nvidia ranks first, with a Lister Score of 97.4.
Nvidia leads globally, with its data centre GPUs and CUDA software behind most AI training.
AMD is the strongest GPU alternative, Google's TPUs are the most proven custom chips in the cloud, and
Broadcom designs custom accelerators for the largest AI labs.
Cerebras is the specialist for fast inference on wafer-scale chips.
Best AI Chip Companies in 2026, compared| RANK | ENTRY | BEST FOR | SCORE | FROM |
|---|
| 1 |  NvidiaNvidia's data centre revenue was $89.0 billion in the quarter ended 26 July 2026, up 117% from a year earlier. | AI labs and enterprises training and running large models | 97.4 | Custom |
|---|
|
| 2 |  AMDAMD's data centre segment, which includes its EPYC CPUs, earned $6.7 billion in the second quarter of 2026, up 107% year on year. | Cloud providers and AI labs that want a second GPU supplier | 95.9 | Custom |
|---|
|
| 3 |  Google (TPU)Anthropic agreed in October 2025 to access up to one million Google TPUs. | AI labs training and serving large models in the cloud | 94.7 | Pay per use on Google Cloud |
|---|
|
| 4 |  BroadcomBroadcom's AI semiconductor revenue was $16.7 billion in the quarter ended 2 August 2026, up 221% year on year. | Hyperscalers and AI labs building their own chips | 93.5 | Custom |
|---|
|
| 5 |  Amazon Web Services (Trainium)AWS activated Project Rainier with about 500,000 Trainium2 chips in October 2025. | Teams training and serving models on AWS at lower cost | 92.2 | Pay per use on AWS |
|---|
|
| 6 |  Huawei (Ascend)Huawei showed its Atlas 950 SuperPoD, built on Ascend chips, at MWC in March 2026. | AI labs and clouds in China that cannot buy top Nvidia chips | 90.8 | Custom |
|---|
|
| 7 |  CerebrasCerebras began trading on the Nasdaq under the ticker CBRS on 14 May 2026. | Teams that need very fast AI inference | 89.5 | Custom |
|---|
|
| 8 |  IntelIntel unveiled the Crescent Island inference GPU, with 160GB of LPDDR5X memory, at the OCP Global Summit in October 2025. | Enterprises seeking lower-cost inference hardware | 88.1 | Custom |
|---|
|
| 9 |  QualcommQualcomm named Humain as the first AI200 and AI250 customer, planning 200 megawatts of deployment from 2026. | Operators running large-scale AI inference | 86.6 | Custom |
|---|
|
| 10 |  TenstorrentTenstorrent raised $1.5 billion at a $6.2 billion valuation on 19 October 2026. | Developers who want open, buyable AI hardware | 85.2 | From $1,399 |
|---|
|
|
How to choose
Start from what you need it for. Where each entry fits best:
- Nvidia (#1): AI labs and enterprises training and running large models
- AMD (#2): Cloud providers and AI labs that want a second GPU supplier
- Google (TPU) (#3): AI labs training and serving large models in the cloud
- Broadcom (#4): Hyperscalers and AI labs building their own chips
- Amazon Web Services (Trainium) (#5): Teams training and serving models on AWS at lower cost
- Huawei (Ascend) (#6): AI labs and clouds in China that cannot buy top Nvidia chips
- Cerebras (#7): Teams that need very fast AI inference
- Intel (#8): Enterprises seeking lower-cost inference hardware
- Qualcomm (#9): Operators running large-scale AI inference
- Tenstorrent (#10): Developers who want open, buyable AI hardware
Each entry's full verdict is below.
The ranking, in detail
01
Data centre GPUs and the CUDA platform
Nvidia, based in Santa Clara, California, designs the Blackwell and Rubin data centre GPUs, along with networking hardware and the CUDA software platform. It reported data centre revenue of $89.0 billion for the quarter ended 26 July 2026.
PROS- CUDA software used across the industry
- Full stack from chips to networking
CONS- Premium pricing
- Subject to US export limits on China sales
- Best for
- AI labs and enterprises training and running large models
- Where it falls short
- Premium pricing and US export limits on China sales may restrict access.
- Pick something else if
- you need lower costs or unrestricted sales in China.
97.4
02
Instinct GPUs and Helios rack systems
AMD, based in Santa Clara, California, designs the Instinct MI series of data centre GPUs and Helios rack-scale systems. In October 2025 it signed a multi-year deal to supply OpenAI with Instinct GPUs, giving OpenAI an option to take up to a 10% stake, and it also has an agreement with Anthropic for up to 2 gigawatts of MI450 GPUs.
PROS- Large memory on Instinct GPUs
- Major supply deals with OpenAI and Anthropic
CONS- Software stack less mature than CUDA
- Much smaller AI share than Nvidia
- Best for
- Cloud providers and AI labs that want a second GPU supplier
- Where it falls short
- Its software stack is less mature than CUDA.
- Pick something else if
- you need a more mature software ecosystem.
95.9
03
Tensor Processing Units on Google Cloud
Google designs Tensor Processing Units, now in their seventh generation, Ironwood, and offers them through Google Cloud. In October 2025 Anthropic agreed to use up to one million TPUs, adding to Google's own use of the chips for Gemini.
PROS- Mature design over many generations
- Large external customers
CONS- Mainly available through Google Cloud
- Less flexible than general-purpose GPUs
- Best for
- AI labs training and serving large models in the cloud
- Where it falls short
- It is less flexible than general-purpose GPUs.
- Pick something else if
- you need general-purpose GPU flexibility.
94.7
04
Custom AI accelerators for hyperscalers
Broadcom, based in Palo Alto, California, designs custom AI accelerators, which it calls XPUs, for large customers including Google, Anthropic and OpenAI, along with networking chips for AI clusters. It reported AI semiconductor revenue of $16.7 billion for the quarter ended 2 August 2026.
PROS- Designs chips for several of the largest AI labs
- Strong AI networking line
CONS- Not sold off the shelf
- Revenue concentrated in a few customers
- Best for
- Hyperscalers and AI labs building their own chips
- Where it falls short
- Not sold off the shelf.
- Pick something else if
- you need standard products available immediately.
93.5
05
Trainium chips for training and inference on AWS
Amazon Web Services designs its own AI chips through Annapurna Labs, including Trainium for training and inference. Project Rainier, an AWS cluster of about 500,000 Trainium2 chips used by Anthropic, came online in October 2025, and AWS launched Trainium3 in December 2025.
PROS- Lower cost than GPU instances, by AWS's figures
- Anthropic as anchor customer
CONS- Only available inside AWS
- Smaller software ecosystem than CUDA
- Best for
- Teams training and serving models on AWS at lower cost
- Where it falls short
- It is only available inside AWS.
- Pick something else if
- you need a multi-cloud strategy.
92.2
06
Ascend AI chips and Atlas SuperPoD systems
Huawei, based in Shenzhen, designs the Ascend AI chips used in its Atlas SuperPoD clusters. Its published roadmap puts the Ascend 950 series in 2026, the 960 in 2027 and the 970 in 2028.
PROS- Very large clustered systems
- In-house memory and interconnect
CONS- Limited availability outside China
- Constrained by US export controls
- Best for
- AI labs and clouds in China that cannot buy top Nvidia chips
- Where it falls short
- US export controls limit availability outside China.
- Pick something else if
- you need global availability.
90.8
07
Wafer-scale AI chips and inference cloud
Cerebras, based in Sunnyvale, California, makes the Wafer-Scale Engine, a single chip the size of a silicon wafer, sold in CS-3 systems and through its own inference cloud. It listed on Nasdaq in May 2026 after pricing its IPO at $185 a share.
PROS- Built for fast inference
- Sells systems and a cloud service
CONS- Relies on a few large customers
- Unusual hardware needs specialist support
- Best for
- Teams that need very fast AI inference
- Where it falls short
- Unusual hardware needs specialist support.
- Pick something else if
- you lack specialist support resources.
89.5
08
Gaudi accelerators and inference GPUs
Intel, based in Santa Clara, California, designs the Gaudi 3 AI accelerator and is developing Crescent Island, a data centre GPU for AI inference with customer sampling planned for the second half of 2026.
PROS- Large manufacturing and sales reach
- Cost-focused inference parts
CONS- Small share of AI accelerator sales
- AI roadmap has been reset before
- Best for
- Enterprises seeking lower-cost inference hardware
- Where it falls short
- It holds a small share of AI accelerator sales.
- Pick something else if
- you need a market leader in AI accelerators.
88.1
09
AI200 and AI250 inference racks
Qualcomm, based in San Diego, makes the Cloud AI 100 inference accelerator and announced the AI200 and AI250 rack-scale inference systems in October 2025, due for commercial availability in 2026 and 2027. Saudi AI company Humain is the first customer.
PROS- High memory per card for inference
- Power-efficiency experience from mobile chips
CONS- New to data centre AI at scale
- Few named customers so far
- Best for
- Operators running large-scale AI inference
- Where it falls short
- It has few named customers so far.
- Pick something else if
- you need a proven track record with many clients.
86.6
10
RISC-V AI processors you can buy
Tenstorrent, based in Toronto and led by chip architect Jim Keller, designs RISC-V based AI processors sold as Blackhole cards, QuietBox workstations and Galaxy servers. It raised $1.5 billion at a $6.2 billion valuation in October 2026.
PROS- Open-source software stack
- Cards sold directly at public prices
CONS- Software less mature than CUDA
- Much smaller scale than the chip giants
- Best for
- Developers who want open, buyable AI hardware
- Where it falls short
- Its software is less mature than CUDA.
- Pick something else if
- you need CUDA compatibility.
85.2
Put your company on this ranking
Ask us for a quote for one, two or three years. Nothing recurring.
Positions #4 to #10 are available on request. Contact us.
Not on this ranking?Tell us about your company and we will reply by email.
Frequently asked questions
What are the best AI Chip Companies?
Nvidia ranks first on Lister. Nvidia, based in Santa Clara, California, designs the Blackwell and Rubin data centre GPUs, along with networking hardware and the CUDA software platform. It reported data centre revenue of $89.0 billion for the quarter ended 26 July 2026.
What are the top 10 AI Chip Companies?
1. Nvidia; 2. AMD; 3. Google (TPU); 4. Broadcom; 5. Amazon Web Services (Trainium); 6. Huawei (Ascend); 7. Cerebras; 8. Intel; 9. Qualcomm; 10. Tenstorrent.
Some links on Lister are affiliate links. Rankings are not affected.
Related rankings
