Key takeaways
- Top pick: SambaNova Systems, best for: Enterprises running AI inference on their own premises. SambaNova says JPMorganChase selected it as an inference infrastructure partner and plans to deploy SN40 and SN50 systems.
- Also near the top: Etched, best for: AI companies serving transformer models at scale; and Tenstorrent, best for: Developers who want open, buyable AI hardware.
- How it was ranked: 10 startups compared. Each verdict is below.
Best AI Chip Startups: Top 10 Ranked. SambaNova Systems ranks first, with a Lister Score of 96.8.
SambaNova leads for enterprise inference, with JPMorganChase deploying its systems, and
Etched carries the highest valuation on this list for its transformer-only Sohu chip.
Tenstorrent is the choice for teams that want to buy accelerator cards today, while Positron and
MatX are the best-funded newcomers for inference and large-model training.
Best AI Chip Startups in 2026, compared| RANK | ENTRY | BEST FOR | SCORE | FROM |
|---|
| 1 |  SambaNova SystemsSambaNova says JPMorganChase selected it as an inference infrastructure partner and plans to deploy SN40 and SN50 systems. | Enterprises running AI inference on their own premises | 96.8 | Custom |
|---|
|
| 2 |  EtchedJane Street has tested Sohu and runs its own Etched rack in its data centre, according to Yahoo Finance. | AI companies serving transformer models at scale | 95.6 | Not public |
|---|
|
| 3 |  TenstorrentTenstorrent says it licenses its TT-Ascalon RISC-V CPU and Tensix AI cores to chip designers including Samsung and LG. | Developers who want open, buyable AI hardware | 94.4 | From $1,399 |
|---|
|
| 4 |  Positron AIReuters reported Positron's $875 million Series C at a $5 billion valuation in September 2026. | Data centres cutting the power cost of AI inference | 93.1 | Not public |
|---|
|
| 5 |  MatXMatX plans to finalise its chip design in 2026 and ship in 2027, manufacturing with TSMC, according to DCD. | AI labs training large language models | 91.9 | Not public |
|---|
|
| 6 |  RebellionsRebellions' March 2026 round valued the company at $2.3 billion, according to TechCrunch. | Data centres seeking an alternative inference accelerator | 90.6 | Custom |
|---|
|
| 7 |  LightmatterLightmatter's Series D brought its total funding to $850 million, according to DCD. | Chipmakers and hyperscalers scaling large AI clusters | 89.3 | Custom |
|---|
|
| 8 |  d-Matrixd-Matrix's Series C was co-led by Bullhound Capital, Triatomic Capital and Temasek, with Microsoft's M12 fund taking part. | Data centres running low-latency generative AI inference | 88.0 | Custom |
|---|
|
| 9 |  FuriosaAILG AI Research has adopted RNGD to run its EXAONE foundation models. | Enterprises running large language models efficiently | 86.7 | Custom |
|---|
|
| 10 |  Axelera AIAxelera says it has attracted more than $450 million in equity, grants and venture debt since it was founded. | Edge and industrial AI, such as computer vision | 85.4 | From €264.95 |
|---|
|
|
How to choose
Start from what you need it for. Where each entry fits best:
- SambaNova Systems (#1): Enterprises running AI inference on their own premises
- Etched (#2): AI companies serving transformer models at scale
- Tenstorrent (#3): Developers who want open, buyable AI hardware
- Positron AI (#4): Data centres cutting the power cost of AI inference
- MatX (#5): AI labs training large language models
- Rebellions (#6): Data centres seeking an alternative inference accelerator
- Lightmatter (#7): Chipmakers and hyperscalers scaling large AI clusters
- d-Matrix (#8): Data centres running low-latency generative AI inference
- FuriosaAI (#9): Enterprises running large language models efficiently
- Axelera AI (#10): Edge and industrial AI, such as computer vision
Each entry's full verdict is below.
The ranking, in detail
01
Dataflow chips and systems for AI inference
SambaNova, based in Palo Alto, California, designs Reconfigurable Dataflow Unit chips and sells SN40 and SN50 systems for running large AI models. In July 2026 it announced a $1 billion first close of its Series F at an $11 billion valuation, led by General Atlantic.
PROS- Named enterprise customer in banking
- Large recent funding round
CONS- Price not published
- Competes directly with Nvidia systems
- Best for
- Enterprises running AI inference on their own premises
- Where it falls short
- It does not publish pricing, making cost estimation difficult for buyers.
- Pick something else if
- you need transparent upfront pricing.
96.8
02
Sohu chip built for transformer models
Etched, founded in 2022 and based in San Jose, California, designs Sohu, an inference chip built only for transformer models. In August 2026 it raised $700 million at a $21 billion valuation, led by Jane Street, weeks after a $300 million Series C at $10.3 billion.
PROS- Highest valuation on this list
- Early customer running its own rack
CONS- Runs transformer models only
- Early commercial stage
- Best for
- AI companies serving transformer models at scale
- Where it falls short
- It runs transformer models only, limiting use for other AI architectures.
- Pick something else if
- you need support for non-transformer workloads.
95.6
03
RISC-V AI processors you can buy
Tenstorrent, founded in 2016 in Toronto and led by chip architect Jim Keller, sells Wormhole and Blackhole AI accelerator cards and licenses its RISC-V CPU and Tensix AI cores to other chip designers. It raised a $1.5 billion Series E at a $6.2 billion valuation on 19 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.
94.4
04
Power-efficient AI inference hardware
Positron, founded in 2023 and based in Reno, Nevada, builds Atlas inference systems and Asimov, its AI inference chip platform. It raised an $875 million Series C at a $5 billion valuation in September 2026, seven months after a $230 million Series B.
PROS- Rapid growth in funding and valuation
- Built around performance per watt
CONS- Young company
- Price not published
- Best for
- Data centres cutting the power cost of AI inference
- Where it falls short
- Buyers cannot see the price because it is not published.
- Pick something else if
- you need to know costs upfront.
93.1
05
Processors for large language models
MatX, founded in 2022 by former Google TPU engineers, is designing MatX One, a processor for training and running large language models. It raised a $500 million Series B on 24 February 2026, co-led by Jane Street and Situational Awareness, with Marvell among the backers.
PROS- Founding team built Google's TPUs
- Backed by Jane Street and Marvell
CONS- Chips not shipping yet
- Unproven at volume
- Best for
- AI labs training large language models
- Where it falls short
- MatX chips are not yet shipping and remain unproven at volume.
- Pick something else if
- you need immediate hardware availability.
91.9
06
AI inference accelerators from Korea
Rebellions, founded in 2020 in Seoul, makes AI inference accelerators including ATOM-Max and Rebel100, sold in servers and racks. It raised $400 million in a pre-IPO round in March 2026 at a $2.3 billion valuation.
PROS- Products already sold in servers and racks
- Well funded for an Asian chip startup
CONS- Smaller software ecosystem than Nvidia
- Price not published
- Best for
- Data centres seeking an alternative inference accelerator
- Where it falls short
- Buyers cannot see the price because it is not published.
- Pick something else if
- you need to know costs upfront.
90.6
07
Photonic interconnect for AI chips
Lightmatter, an MIT spinout based in Mountain View, California, builds Passage, a photonic engine that moves data between AI chips using light. It raised a $400 million Series D in October 2024 at a $4.4 billion valuation.
PROS- Deep funding and high valuation
- Targets the data-movement limit in AI clusters
CONS- Builds interconnect rather than compute chips
- Price not published
- Best for
- Chipmakers and hyperscalers scaling large AI clusters
- Where it falls short
- It builds interconnects rather than compute chips.
- Pick something else if
- you need compute chips.
89.3
08
In-memory compute for AI inference
d-Matrix, founded in 2019 in Santa Clara, California, makes Corsair inference accelerator cards that compute inside memory, alongside JetStream network cards and Aviator software. It closed a $275 million Series C in November 2025 at a $2 billion valuation.
PROS- Backed by Temasek and Microsoft's venture fund
- Cards, networking and software from one vendor
CONS- No named customers published
- Price not published
- Best for
- Data centres running low-latency generative AI inference
- Where it falls short
- Buyers cannot evaluate costs because the company does not publish its pricing.
- Pick something else if
- you need to know exact costs before committing.
88.0
09
RNGD inference chip for data centres
FuriosaAI, founded in 2017 in Seoul, makes RNGD, an inference accelerator for large language models. It reportedly turned down an $800 million acquisition offer from Meta in March 2025 and raised a $125 million Series C bridge round that July, valuing it at about $735 million.
PROS- Named customer in LG AI Research
- Stayed independent after a reported Meta offer
CONS- Smaller funding than US rivals
- Price not published
- Best for
- Enterprises running large language models efficiently
- Where it falls short
- It does not publish its pricing.
- Pick something else if
- you need to know costs upfront.
86.7
10
AI inference chips for edge and server
Axelera AI, founded in 2021 in Eindhoven, Netherlands, makes Metis AI inference accelerators sold as embedded modules, edge cards and server cards. It announced more than $250 million in new funding in February 2026.
PROS- Cards on open sale with published prices
- Covers edge devices as well as servers
CONS- Smaller than US data centre rivals
- Newer to server-class products
- Best for
- Edge and industrial AI, such as computer vision
- Where it falls short
- It is smaller than US data center rivals.
- Pick something else if
- you need a larger US-based provider.
85.4
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 Startups?
SambaNova Systems ranks first on Lister. SambaNova, based in Palo Alto, California, designs Reconfigurable Dataflow Unit chips and sells SN40 and SN50 systems for running large AI models. In July 2026 it announced a $1 billion first close of its Series F at an $11 billion valuation, led by General Atlantic.
What are the top 10 AI Chip Startups?
1. SambaNova Systems; 2. Etched; 3. Tenstorrent; 4. Positron AI; 5. MatX; 6. Rebellions; 7. Lightmatter; 8. d-Matrix; 9. FuriosaAI; 10. Axelera AI.
Some links on Lister are affiliate links. Rankings are not affected.
Related rankings
