What are the most valuable AI chip startups?

Top startups in the AI chip market ranked by valuation

Here is an updated table that ranks the top startups in the AI chip market based on their latest reported or estimated valuations.

If you need to dig deeper and get more detailed data, please check our report on AI chip funding, valuations and who's ahead.

Which AI chip startups are worth the most right now?

Cambricon currently leads our AI chip valuation ranking at about $105.1 billion, with Cerebras Systems at roughly $50.5 billion and MetaX, Enflame Technology and Moore Threads all above $30 billion.

The top of the market has changed a lot because several companies that recently looked like venture-backed semiconductor startups now trade publicly. Cambricon, MetaX, Enflame, Moore Threads, Iluvatar CoreX and Biren Technology all sit near the top of our ranking on public-market values rather than old private funding rounds.

Etched is the first clearly private specialist near the top. Its latest financing valued it at $21 billion, while SambaNova reached $11 billion. Below that, the market drops into single-digit billions pretty quickly.

China also dominates the top ten. Seven of the ten highest-valued companies in our dataset are Chinese, while Cerebras, Etched and SambaNova represent the United States.

That gives today’s AI chip market an unusual shape: public Chinese semiconductor companies occupy much of the very top, while the fastest-moving American challengers are still largely private.

Company	Current valuation	What it mainly does	Valuation basis
Cambricon	$105.1B	Cloud and edge AI processors	Public market
Cerebras Systems	$50.5B	Wafer-scale AI systems	Public market
MetaX	$34.7B	General-purpose AI GPUs	Public market
Enflame Technology	$32.8B	Cloud AI accelerators	Public market
Moore Threads	$31.1B	Full-function GPUs	Public market
Etched	$19B–$21B	Transformer inference	Private round
Iluvatar CoreX	$14.3B	Data-center GPGPUs	Public market
Biren Technology	$14.2B	Training and inference GPUs	Public market
Kunlunxin	$13B–$16B	Data-center AI accelerators	Listing estimate
SambaNova Systems	$10B–$11B	Dataflow AI systems	Private financing

How many AI chip unicorns are there today?

We count about 35 AI chip companies in the dataset whose valuation reaches at least $1 billion, although several of those figures are estimates rather than valuations disclosed in completed financing rounds.

That is a remarkable number for a business where designing one competitive chip can require hundreds of millions of dollars and several years of work.

The group is also much broader technologically than the word “AI chip” suggests. It includes GPUs such as MetaX and Moore Threads, transformer-specific accelerators such as Etched, wafer-scale systems from Cerebras, RISC-V companies such as Tenstorrent, photonics companies such as Lightmatter, analog compute developers such as d-Matrix and edge specialists such as DEEPX.

The cutoff drops quickly. Around position 20, valuations are already around $2.5 billion to $3.5 billion. Around position 35, Hailo sits close to the $1 billion line.

So the sector has created a large unicorn class, but only a much smaller group has made the jump from “billion-dollar AI chip company” to valuations above $5 billion, $10 billion or $20 billion.

Why are Chinese AI chip companies worth so much right now?

Chinese AI chip companies are getting huge valuations because investors are paying for both rapid AI demand and the scarcity of credible domestic alternatives to foreign processors.

That premium is visible across the ranking. Cambricon is around $105 billion in our snapshot, MetaX roughly $35 billion, Enflame almost $33 billion and Moore Threads about $31 billion. Iluvatar CoreX and Biren are both around $14 billion.

Those values are far above what many of the same companies would have commanded as private startups. Those values sit far above what these companies were worth as private startups; we cover what each one is worth, then and now, in our report on top AI chips companies.

Enflame shows how extreme the public-market repricing can become. Its IPO pricing implied a valuation closer to $9 billion. The shares then surged on their debut, and our later market snapshot puts the company above $30 billion.

MetaX and Moore Threads have experienced a similar enthusiasm around locally designed GPUs. Investors are effectively betting that Chinese cloud companies, AI labs and enterprise customers will spend heavily on domestic hardware as access to the strongest foreign processors remains restricted.

The underlying businesses are growing, but the stock-market values have moved even faster. Chinese AI chip valuations currently contain a very large strategic premium on top of normal revenue expectations.

Is Etched now the most valuable private AI chip startup?

Yes. Etched is currently the highest-valued clearly private AI chip specialist in our dataset, with its latest round valuing the company at $21 billion.

SambaNova follows at $11 billion. Fractile is estimated around $5.5 billion to $6.7 billion, Positron is at $5 billion and Lightmatter’s last disclosed valuation was $4.4 billion.

What makes Etched especially unusual is its age. The company was founded in 2022 and chose an extremely narrow bet: hardware designed specifically around transformer inference rather than a broadly programmable GPU.

That approach looks much less eccentric now that transformers remain at the center of large language models and inference spending keeps growing.

Etched also says its first silicon has come back from TSMC, that it is validating rack-scale systems with customers and that it has roughly $1 billion of demand. More recently, Jane Street tested the hardware, received an Etched rack and led the company’s $700 million financing at the $21 billion valuation.

That combination of working silicon, customer testing and real hardware deployment explains why Etched has separated so sharply from most private AI chip companies.

Private company	Latest valuation or range	Funding in our dataset
Etched	$19B–$21B	$1.63B
SambaNova Systems	$10B–$11B	$1.48B
Fractile	$5.5B–$6.7B	$266M
Positron AI	$4.8B–$5.2B	$1.18B
MatX	$3.3B–$6.3B	$605M
Lightmatter	$4B–$5.2B	$850M
Unconventional AI	$4.5B	$475M
Groq	$3.4B–$3.6B	$4.26B
Tenstorrent	$2.5B–$3.5B	$1.02B
Rebellions	$2.3B–$2.7B	$973M

Why did Etched jump from about $5 billion to $21 billion so quickly?

Etched’s valuation rose roughly fourfold because the company moved through several of the hardest stages of building an AI chip in a very short period.

A semiconductor startup can look impressive on paper for years. The real tests come when the chip has to tape out, come back from the fab, run correctly, work inside a complete system and survive customer testing.

Etched moved through several of those steps while raising successive rounds. Its valuation was around $5 billion late last year, then reached $10.3 billion before the latest $700 million round put the company at $21 billion. Etched's valuation rose with each new round, so we look at who has raised the most, and from whom, in our AI chips report.

Jane Street’s involvement is especially useful evidence. The firm tested Etched’s hardware before leading the latest round, and Etched says it has already shipped a rack into Jane Street’s data center.

The company also reports around $1 billion of demand for its systems.

Bookings and demand are obviously different from recognized revenue, so we should not read that figure as $1 billion already earned. But for a four-year-old semiconductor company, customers committing at that scale changes the risk profile substantially.

Etched’s current valuation assumes the next step will work too: producing those systems in large quantities. That remains a much harder test than building the first successful racks.

Is Cerebras still worth more than every private AI chip startup?

Yes. Cerebras is currently worth roughly $50 billion in our dataset, more than twice Etched’s $21 billion private valuation.

Cerebras reached that scale with one of the strangest architectures in modern semiconductors. Instead of cutting a silicon wafer into many normal processors, Cerebras built wafer-scale chips that effectively use almost the whole wafer as one computing engine.

The company spent years proving that this could work commercially. Its eventual public listing then gave investors a daily market price for that bet.

At roughly $50.5 billion, Cerebras is about $29 billion ahead of Etched. It is also worth several times SambaNova, Lightmatter, Positron or Tenstorrent.

Etched is closing ground much faster than Cerebras is adding absolute valuation, but those are two different things. As of now, Cerebras still sits in a separate financial tier among Western AI chip challengers.

Why are AI inference chip startups suddenly worth billions?

AI inference startups are attracting billion-dollar valuations because running AI models is turning into an enormous computing business of its own.

Training gets most of the attention because the clusters are spectacularly expensive. But a model may be trained occasionally and then queried millions or billions of times. Every answer, image, reasoning chain or agent task requires inference.

That is why so many of today’s valuable startups are attacking inference directly.

Etched focuses on transformer inference. Positron builds memory-focused inference hardware. Fractile is developing an SRAM-heavy architecture. Rebellions sells inference accelerators and complete systems. FuriosaAI, d-Matrix, Hailo, DEEPX and SiMa.ai are attacking other parts of the same workload.

The economics become even more important as models spend more compute per answer. Long reasoning chains, AI agents, video generation and multimodal systems can all increase the amount of inference work behind a single user request.

A specialist therefore does not need to replace Nvidia everywhere. Cutting the cost or power use of one huge inference workload can already support a substantial business.

That is why inference now appears again and again among companies valued between $1 billion and $20 billion.

Are photonic AI chip startups actually becoming big companies?

Yes. Photonic AI hardware has moved beyond tiny research startups, with several companies now worth billions of dollars.

Lightmatter is the clearest example. The company has raised about $850 million and states a $4.4 billion valuation. Its focus has increasingly shifted toward optical interconnects that move data between processors inside huge AI systems. As Lightmatter shifts toward optical interconnects, we watch which technologies are gaining traction in our report on AI chips.

Celestial AI reached an even more concrete outcome. Marvell agreed to acquire the company for roughly $3.25 billion upfront, with additional consideration tied to revenue milestones. Marvell later recorded total purchase consideration of about $3.5 billion when the deal closed.

Lightelligence is around $3.9 billion in our ranking. Q.ANT, Neurophos, Salience Labs, Arago Computing and Lumai sit further down the same broad photonics ecosystem.

The common problem is data movement. AI clusters can contain thousands of expensive processors, and those processors lose useful time when networking and memory cannot feed them fast enough.

Photonics is increasingly valuable because that bottleneck is getting worse as clusters grow. The strongest commercial opportunity today appears to be connecting AI hardware more efficiently rather than replacing every electronic processor with an optical one.

Does raising more money make an AI chip startup worth more?

No clear relationship appears in the dataset between how much an AI chip company has raised and how high investors value it.

Tenstorrent, for example, raised more than $693 million in its Series D at a $2 billion pre-money valuation. Our current estimate puts the company around $2.5 billion to $3.5 billion.

Etched has raised much more and is worth $21 billion, but DEEPX shows the opposite pattern: only about $128 million of funding in our dataset against a valuation above $2 billion.

Groq makes the comparison even messier because its corporate structure and recent financings changed substantially. The company announced $650 million in new growth capital and later another $350 million financing that valued Groq at $3.5 billion.

The useful pattern is what happens between funding rounds. Investors reward companies when fresh capital produces working silicon, customer deployments, credible software, manufacturing capacity or a realistic path to large revenue.

Money buys time in semiconductor development. It does not guarantee a high valuation.

How concentrated are AI chip valuations?

AI chip valuations are extremely concentrated: the market contains almost 100 companies in our dataset, but most of the value sits in a small group near the top.

Cambricon alone is above $100 billion. The fifth company, Moore Threads, is still above $30 billion. By number ten, SambaNova is around $10 billion to $11 billion.

Then the ranking falls quickly.

Tenstorrent, around position 20, is in the $2.5 billion to $3.5 billion range. Hailo, around position 35, sits near $1 billion. Around position 50, companies are generally worth a few hundred million dollars. Near position 80, we are down to tens of millions.

A broad “AI chips are booming” headline hides an enormous gap between the leaders and everybody else.

Investors are currently willing to place very large bets on a small number of architectures while dozens of other teams are still trying to prove that they can turn good semiconductor research into a large company.

Approximate ranking position	Valuation scale
#1	~$105B
#5	~$31B
#10	~$10B–$11B
#20	~$2.5B–$3.5B
#30	~$2B
#40	~$450M–$700M
#50	~$220M–$400M
#60	~$225M
#70	~$90M–$140M
#80	~$40M–$65M
#90	~$5M–$15M

Which AI chip startups have increased their valuations the fastest?

Etched, Positron, SambaNova and Rebellions have produced some of the clearest recent private-market valuation jumps.

Etched moved from roughly $5 billion to $21 billion in less than a year. That is an increase of about 320%.

Positron moved even more sharply from a valuation a little above $1 billion after its Series B to $5 billion after raising $875 million in its Series C.

SambaNova reached an $11 billion post-money valuation in its latest $1 billion financing. Its previous Series D in 2021 had valued the company above $5 billion, so the company has now crossed another major valuation threshold after several years of slower progress.

Rebellions went from $1.4 billion in its Series C to roughly $2.34 billion in its subsequent $400 million pre-IPO round.

We also see dramatic moves in public markets, especially in China, but those price changes are easier to reverse because the stock trades every day.

The private rounds are more useful for understanding how investors are repricing technical progress. Multiple AI hardware companies have recently jumped by billions of dollars between financings. That is aggressive even by AI standards.

Are AI chip startups selling chips or entire AI systems now?

The biggest AI chip startups increasingly sell racks, systems, cloud capacity or networking platforms around their silicon because a good chip alone is difficult for customers to adopt.

Cerebras built its business around complete wafer-scale systems. SambaNova combines its processors with hardware and software infrastructure. Etched now describes its product around frontier inference clusters and has shipped a full rack to Jane Street.

Rebellions has introduced RebelRack and RebelPOD systems. Positron combines its accelerators with complete inference systems. Groq operates an inference cloud across multiple data centers. Selling full systems or a cloud is a different business from selling chips alone, so we compare which business models are winning in our report covering AI chips.

This reflects Nvidia’s real competitive advantage. CUDA gets most of the attention, but Nvidia also supplies networking, systems, libraries, server designs and a giant developer ecosystem.

A startup that hands customers a standalone chip leaves them with a lot of integration work.

These days, the stronger challengers are removing more of that burden themselves. That pushes them closer to infrastructure companies than traditional semiconductor vendors and also gives them more ways to earn revenue from each customer.

Can any AI chip startup actually challenge Nvidia?

Several AI chip startups can take valuable workloads from Nvidia, but none is currently close to Nvidia’s overall scale, software reach or market position.

The valuation gap alone is enormous. Cambricon, the biggest company in our ranking, is worth around $105 billion. Etched, the leading private specialist, is at $21 billion. Nvidia operates at a completely different scale.

The more interesting competition happens workload by workload.

Etched is trying to make transformer inference dramatically more efficient. Cerebras attacks large AI workloads with wafer-scale systems. Lightmatter wants to fix the optical interconnect bottleneck. Rebellions and FuriosaAI focus on inference. Chinese GPU companies are building domestic alternatives for customers that cannot rely indefinitely on Nvidia hardware.

A company can therefore become worth $5 billion, $20 billion or potentially much more without taking Nvidia’s overall crown.

This is also why older AI chip acquisitions suddenly look cheap. Intel paid about $2 billion for Habana Labs and hundreds of millions for Nervana. Lenovo bought Kinara for $307 million. Companies with comparable ambitions can now reach multibillion-dollar private valuations before an acquisition even enters the discussion.

The likely competitive picture is several specialists taking pieces of a huge market rather than one startup replacing Nvidia outright.

Why are AI chip startups worth billions before they make much revenue?

AI chip startups can reach huge valuations before large-scale revenue because investors have to place their bets before the commercial ramp is fully visible.

Semiconductor development works differently from ordinary software. A team may spend years designing an architecture, building compilers, taping out silicon and arranging manufacturing before customers can buy meaningful volumes.

By the time a chip company has mature revenue, a lot of the technical risk has already disappeared.

Investors therefore look for earlier proof: working silicon, benchmark results, customer testing, purchase commitments, production agreements and successful deployments.

The problem is that a company can clear several of those milestones and still fail commercially. Manufacturing may prove too expensive. Software can be painful to use. Nvidia may cut prices or improve its own hardware. A new model architecture can weaken the startup’s technical advantage.

That is why a $5 billion valuation for a young AI chip company should be read as a large bet on future scale rather than evidence that a $5 billion operating business already exists.

Are edge AI chip startups worth as much as data-center AI companies?

No. Edge AI chip companies generally receive much lower valuations than companies selling compute into large AI data centers.

DEEPX is one of the strongest edge examples in our ranking at roughly $2.1 billion to $2.3 billion. Hailo is around $1 billion to $1.4 billion, SiMa.ai roughly $850 million to $1.1 billion and Kneron around $700 million to $1 billion.

Further down, Mobilint is worth roughly $330 million to $630 million, MemryX around $130 million to $180 million and Blaize about $83 million.

Compare those numbers with Etched at $21 billion, SambaNova at $11 billion or Cerebras around $50 billion.

The reason is mainly economic. A handful of hyperscale data-center customers can spend enormous amounts on accelerators, memory, networking and power. Edge chips may ship in much higher unit volumes, but each processor usually sells for far less.

For now, investors clearly value exposure to giant AI infrastructure budgets more highly than exposure to cameras, industrial machines, vehicles and other edge devices.

Edge AI company	Approx. valuation	Main focus
DEEPX	$2.1B–$2.3B	Low-power edge NPUs
Hailo	$1.0B–$1.4B	Edge inference processors
SiMa.ai	$850M–$1.1B	Physical-AI edge inference
Kneron	$700M–$1.0B	Neural processors
Mobilint	$330M–$630M	Low-power edge NPUs
Kinara	$284M acquisition value in dataset	Programmable edge NPUs
MemryX	$130M–$180M	Edge accelerators
Blaize	~$83M	Programmable edge processors
Perceive	$80M acquisition value	Ultra-low-power inference

Are there too many AI chip startups?

Yes, there are almost certainly too many AI chip companies for all of them to become large independent businesses.

Our dataset contains 98 names. Even a very large accelerator market is unlikely to support dozens of independent companies each spending heavily on advanced silicon, software and manufacturing. With 98 names and room for far fewer, we weigh the likely future winners in our AI chips market report.

The crowding is especially obvious in inference. Etched, Positron, Fractile, Rebellions, FuriosaAI, d-Matrix, NeuReality, VSORA and several others are all trying to reduce the cost or improve the performance of running AI models.

Still, the 98 companies are not building the same thing. We found GPUs, RISC-V processors, analog compute, photonics, FPGA architectures, in-memory computing, networking chips, edge NPUs, model-specific ASICs and even superconducting designs.

The long tail also shows what usually happens in semiconductors. Many technically serious companies remain below $500 million in valuation, while others eventually get acquired.

We therefore see the current field as dozens of competing architecture bets. A much smaller number should eventually become durable independent platforms.

Does a high AI chip valuation mean the technology is already winning?

No. A high AI chip valuation mainly tells us that investors think the technology has a credible chance of becoming important.

Etched is already worth $21 billion, but its large-scale manufacturing ramp is still ahead. Positron is worth $5 billion while its next-generation silicon is still moving toward broader production. Fractile commands a multibillion-dollar estimated valuation despite being much earlier commercially than Nvidia or AMD.

Public markets can get ahead of the operating numbers too.

Enflame’s market value expanded dramatically after listing, even though the company remained at an early stage compared with established global semiconductor suppliers. Chinese AI chip stocks are being valued partly on how much strategic importance investors assign to domestic compute capacity.

We therefore keep technical progress, commercial adoption and valuation separate.

A working chip lowers technical risk. Customer deployment lowers commercial risk. Large recurring revenue would remove another layer. A valuation can move far ahead of all three when investors believe the eventual prize is big enough.

What could make AI chip valuations fall hard?

AI chip valuations could fall quickly if specialized accelerators fail to save customers enough money or power to justify leaving Nvidia’s ecosystem.

That is the biggest common risk across many companies in this ranking.

Nvidia does not need to make every alternative accelerator useless. It only needs to keep improving price-performance fast enough that switching becomes unattractive for many customers.

Model architecture is another risk. Etched has made an unusually concentrated bet on transformers. That concentration is powerful while transformers dominate frontier inference, but a major architectural shift could make some of today’s optimization choices less valuable.

Manufacturing creates another layer of risk. Startups need advanced wafers, packaging, memory and networking components, and they need those pieces in enough volume to serve large customers. Successful prototypes do not guarantee a smooth production ramp.

Public Chinese companies have an additional valuation problem because several trade at levels that already assume a lot of future growth. Revenue can keep rising while the share price falls if investors simply decide to pay a lower multiple.

So the dangerous scenario does not require AI spending to collapse. Slower growth, cheaper incumbent hardware or a few disappointing deployments could be enough to reset some of the most aggressive valuations.

Are AI chip startup valuations actually justified right now?

AI chip valuations are genuinely experiencing a major repricing, but the strongest evidence supports only a small group of today’s huge values with mature commercial scale.

We see more than a few isolated funding rounds. Cambricon is above $100 billion in our snapshot. Cerebras is around $50 billion. MetaX, Enflame and Moore Threads are above $30 billion. Etched has reached $21 billion privately. SambaNova is at $11 billion. Roughly three dozen companies reach or approach unicorn territory.

There is also a coherent reason for the excitement. AI spending keeps creating new bottlenecks in inference, memory, networking, power and data movement. Different startups are attacking different pieces of that stack. Each new bottleneck in memory, networking and power opens a gap, and we map where the biggest opportunities are in our AI chips report.

Yet the ranking becomes much less spectacular below the leaders. By the 20th company we are around $3 billion, by roughly the 35th we are near $1 billion, and much of the 98-company field is worth only a few hundred million dollars or less.

The sharpest current opportunity appears in inference. Etched, Positron, Fractile, Rebellions, FuriosaAI and d-Matrix all benefit from the view that serving models will become an enormous market of its own.

China adds another dynamic. Domestic GPU and accelerator companies are getting valuations that reflect commercial growth plus the strategic value of having homegrown AI compute.

Our final judgment is that the AI chip valuation boom is real, broad enough to matter and much larger than a few headline deals. The market has already decided that alternatives to Nvidia could become extremely valuable.

What remains unresolved is which companies can turn those expectations into tens of billions of dollars of durable business. Today’s valuations are placing that bet well before the answer is known.