Which AI chip startups are generating the most revenue today?

Why is it so hard to tell which AI chip startup makes the most money?

AI chip startup revenue is unusually hard to rank today because the companies with the cleanest financial data are quickly becoming public, while many private startups disclose bookings, contracts or forecasts instead of actual sales.

Cerebras is the clearest example. It generated $510 million of revenue in 2025 and another $373.5 million in the first half of 2026, but it has now completed its IPO. Several Chinese challengers followed the same path. Moore Threads, MetaX, Biren Technology and Iluvatar CoreX all came out of the startup world, yet their recent listings now give us much better financial disclosure than we get from most private US competitors.

The private companies are harder to compare. Groq has talked about very large Saudi infrastructure commitments, while people familiar with its finances later told Forbes that annual revenue around its Nvidia licensing deal was closer to $100 million. SambaNova says it finished 2025 with record revenue and bookings but does not publish a full income statement. Tenstorrent previously projected roughly $200 million of 2025 revenue, but later financial disclosure has not confirmed that it reached the target.

Etched shows the problem even more clearly. The company says it has more than $1 billion in signed customer contracts, yet its first rack-scale systems are still being validated with customers. That tells us Etched has serious demand, but not how much revenue has already been recognized.

So there are really two rankings here: which startup-origin AI chip company has built the biggest business, and which company that is still private is making the most money today.

What should we actually count as an AI chip startup?

For this ranking, an AI chip startup is an independent company whose core business depends on proprietary AI processors, accelerator systems or compute built around its own silicon.

That includes companies such as Cerebras, Groq, SambaNova, Tenstorrent, d-Matrix, Etched, Rebellions, FuriosaAI and Axelera AI. Their main technical differentiation comes from their own accelerator architecture.

We also look at Moore Threads, MetaX, Biren Technology and Iluvatar CoreX because they show how far AI-chip challengers can actually scale commercially, even though all four have now crossed into public markets.

We exclude Nvidia, AMD, Broadcom and Marvell because they are established semiconductor companies with much broader businesses. Google TPU, Amazon Trainium, Microsoft Maia and Meta's MTIA are also outside the ranking because they are internal hyperscaler chips rather than products from independent startups.

Hailo would once have fit the definition, but Microchip agreed to acquire the Israeli edge-AI company after Hailo ran into financing pressure. Blaize is another borderline case because it came from the startup ecosystem but is already Nasdaq-listed.

The distinction changes the ranking a lot. Startup-origin companies have already reached several hundred million dollars of annual revenue. Among companies that are still private today, even crossing $100 million remains relatively rare.

Is Cerebras actually the biggest AI chip startup by revenue?

Cerebras is the clearest revenue leader from the modern AI-chip startup wave, even though it is now public rather than technically a startup.

Cerebras reported $290.3 million of revenue in 2024 and $510 million in 2025, which means sales grew 76% in one year. The growth continued in 2026: revenue reached $193.4 million in the first quarter and $180.1 million in the second, taking first-half GAAP revenue to $373.5 million.

More interestingly, Cerebras is becoming less dependent on hardware sales. In the second quarter, hardware generated $54.1 million while cloud and other services generated $126 million. A year earlier, cloud and services had contributed only $33 million in the same quarter.

Cloud and services revenue rose 281% year over year in the second quarter while hardware revenue declined. Cerebras can increasingly make money every time customers consume AI compute instead of waiting for another large system sale.

Cerebras also reported $25.4 billion of remaining performance obligations after signing major capacity agreements, including a multiyear OpenAI deal worth more than $20 billion. We are not counting those obligations as current revenue, but they make the existing sales base look much more repeatable than it did a year ago.

Among recent startup-origin AI chip companies, nobody else has the same clean combination of hundreds of millions in recognized revenue, continued growth and contracted future demand.

If you want more recent data on this point, please see our latest AI chip market report.

Are Chinese AI chip companies already making more revenue than the famous US startups?

Yes. Several Chinese AI-chip challengers are now generating more recognized revenue than most of the private American names that dominate the conversation.

Moore Threads generated roughly RMB1.51 billion in 2025 after growing more than 240%. It then reported RMB1.736 billion in the first half of 2026 alone, so six months of revenue had already exceeded its entire previous year.

MetaX reached around RMB1.64 billion in 2025, or roughly $230 million. More than RMB1.6 billion came from GPU products and accessories, which means the number mostly reflects the core AI-chip business rather than unrelated services.

Biren Technology generated RMB1.035 billion in 2025, up from RMB336.8 million in 2024. Almost the full amount came from intelligent-computing products and solutions. Biren then said it expected RMB1.15 billion to RMB1.30 billion of revenue in the first half of 2026.

Iluvatar CoreX followed the same broad path. Revenue climbed from RMB539.5 million in 2024 to RMB1.034 billion in 2025, then reached RMB945.7 million in the first half of 2026. GPGPU products generated RMB916.1 million of that latest six-month total.

The pattern is hard to ignore: four former startups have crossed roughly RMB1 billion of annual or half-year revenue within a relatively short period. There is no equally deep cluster of still-private US AI-chip companies above $100 million in confirmed annual sales.

Domestic substitution is helping. US restrictions on advanced Nvidia products have created customers that actively need Chinese alternatives, while state-backed and private data-center spending gives those companies a large home market. That does not prove the chips would win the same business in a completely open global market, but it does prove they are selling at real scale today.

Company	Latest meaningful revenue evidence	What it tells us
Cerebras	$373.5M in H1 2026; $510M in 2025	Clear startup-origin revenue leader
Moore Threads	RMB1.736B in H1 2026	Six-month revenue already above full-year 2025
MetaX	~RMB1.64B in 2025	GPU sales already in the hundreds of millions of dollars
Biren Technology	RMB1.15B-1.30B expected in H1 2026	Revenue still accelerating after crossing RMB1B
Iluvatar CoreX	RMB945.7M in H1 2026	Nearly all latest revenue came from GPGPU products

So which private AI chip startup is making the most money today?

Among AI chip companies that are still private, Groq, SambaNova and Tenstorrent form the most credible top group today, but there is no undisputed revenue leader with the same level of evidence we have for Cerebras.

Groq has the strongest externally reported current number. Forbes spoke with people familiar with the company who put annual revenue around the time of its Nvidia transaction at roughly $100 million.

SambaNova appears to be in a similar range. The company said it closed 2025 with record bookings and revenue, while outside private-company databases put sales around $100 million. We treat that figure as an estimate because SambaNova does not publish audited annual results.

Tenstorrent could be larger. A company presentation projected about $200 million of 2025 revenue after more than $40 million expected in 2024. The same material referred to roughly $150 million of IP bookings over eight months. The problem is simple: we have the forecast, but not later disclosure proving the company actually recognized $200 million.

d-Matrix looks like the next tier. Several outside estimates put annual revenue around $65 million to $70 million. That would make d-Matrix commercially meaningful, but those estimates are much less reliable than public-company filings.

Rebellions is smaller in absolute terms but easier to verify. Korean financial reporting put 2025 sales around KRW35 billion, up sharply from roughly KRW10.3 billion in 2024.

The private leaders appear to be around $100 million to the low hundreds of millions in annual revenue, not billions.

Private company	Best current revenue evidence	Confidence
Tenstorrent	~$200M 2025 company forecast	Medium-low
Groq	Around $100M annual revenue reported by Forbes sources	Medium-high
SambaNova	Around $100M estimated 2025 revenue; company confirms record year	Medium
d-Matrix	Roughly $65M-$70M external estimate	Low-medium
Rebellions	~KRW35B 2025 sales	High
FuriosaAI	~KRW5.7B 2025 sales	High

If you want more recent data on this point, please see our latest AI chip market report.

Is Groq really a $500 million revenue company?

Groq does not look like a $500 million revenue company today; the best current reporting puts annual revenue much closer to $100 million.

Groq's growth has still been huge. Forbes previously reported that the company made only about $3 million of revenue in 2023 while losing roughly $88 million. By 2024 and 2025, Groq had become one of the most visible alternatives to Nvidia for high-speed LLM inference.

The confusion came from much larger commercial ambitions. Groq's Saudi partnership became associated with approximately $1.5 billion of activity, and the company reportedly revised an earlier 2025 revenue ambition of around $2 billion down toward roughly $500 million.

Those numbers were never the same thing as recognized annual sales. People familiar with Groq's finances later told Forbes that annual revenue around the Nvidia transaction was closer to $100 million. Former employees also described the Saudi headline figure as including infrastructure, chips and future compute economics.

That still implies extraordinary growth from 2023. Going from $3 million toward roughly $100 million in around two years would put Groq among the fastest-growing private AI-chip companies we can identify.

Groq also says its cloud now serves more than six million developers and runs across 13 data centers. The company has clearly built a real inference business. Old $500 million targets should not be treated as though Groq had already booked that amount as revenue.

Is SambaNova already making more than $100 million a year?

SambaNova looks like a roughly $100 million-scale AI chip business today, although the public evidence is not strong enough to claim a much higher number.

The company said it ended 2025 with record bookings and record revenue. Outside private-market databases generally place annual sales around $100 million, which fits the scale of SambaNova's growing mix of inference cloud, enterprise systems and strategic deployments.

There are also signs that the business improved materially after a difficult period. SambaNova spent part of 2025 exploring strategic options and was reportedly discussing a possible Intel acquisition at a valuation well below its old $5 billion peak.

Instead of being sold, SambaNova changed direction. It announced $350 million of strategic financing in early 2026 and later raised another $1 billion at an $11 billion valuation. SoftBank also committed to use SambaNova's SN50 technology in next-generation AI data centers in Japan.

Funding and valuation do not tell us revenue directly, so we should not use them to manufacture a larger sales figure. But together with SambaNova's own statement about record revenue, they make a collapse in commercial activity look unlikely.

For now, around $100 million is the sensible anchor. Anything materially above that needs better disclosure.

Did Tenstorrent actually reach $200 million of revenue?

We cannot confirm that Tenstorrent reached $200 million of revenue, because the number we can trace is a company forecast rather than a reported annual result.

Tenstorrent does have substantial commercial activity. One company presentation described more than $40 million of expected 2024 revenue, approximately $150 million of IP bookings over eight months and roughly $200 million of expected 2025 revenue.

Its business model also makes Tenstorrent different from most AI-chip startups. The company licenses both AI and RISC-V processor IP to customers designing their own chips, while also selling Blackhole accelerator cards, workstations and larger systems. Executives told EE Times that IP deals historically accounted for most bookings, with customers including LG and Hyundai.

That gives Tenstorrent three ways to make money: intellectual-property licenses, accelerator hardware and complete systems. A large architecture deal can therefore generate meaningful revenue before Tenstorrent ships huge volumes of its own chips.

The hardware side still looks earlier. Tenstorrent cut part of its enterprise sales organization and shifted more attention toward developers, while CEO Jim Keller said the company still had work to do before optimizing entirely for large enterprise customers.

We would still place Tenstorrent in the top private tier. We just would not put a confirmed $200 million number beside its name until the company publishes evidence that the forecast was actually achieved.

Are Korea's AI chip startups actually making real revenue?

Rebellions is now generating meaningful AI-chip revenue in Korea, while FuriosaAI remains much smaller financially despite attracting plenty of strategic attention.

Rebellions generated about KRW35 billion of revenue in 2025, according to Korean reporting based on its accounts and year-end financial information. That was roughly 3.4 times the KRW10.35 billion reported for 2024 and more than ten times its roughly KRW2.73 billion of 2023 revenue.

The growth came from actual deployments. Rebellions sold thousands of ATOM and ATOM-Max accelerators, with products used by companies including SK Telecom, KT Cloud and LG Electronics.

The company still missed its own much more aggressive target. Rebellions had previously aimed for more than KRW100 billion of 2025 revenue. Reaching roughly KRW35 billion therefore counts as strong growth, but not the commercial breakout management originally expected.

FuriosaAI is much earlier. Korean financial information shows revenue of about KRW5.7 billion in 2025, up from approximately KRW3.0 billion the previous year. Its operating loss remained around KRW62.5 billion.

So the Korean revenue race is easy to call right now. Rebellions is well ahead of FuriosaAI, even though both companies are still small beside the leading Chinese challengers or Cerebras.

Company	2024 revenue	2025 revenue	What changed
Rebellions	~KRW10.3B	~KRW35B	Revenue more than tripled as accelerator shipments increased
FuriosaAI	~KRW3.0B	~KRW5.7B	Sales grew, but remain small relative to spending

If you want more recent data on this point, please see our latest AI chip market report.

Does Etched's $1 billion in AI chip contracts make it a revenue leader?

Etched is one of the most important future AI-chip contenders, but its $1 billion-plus contract book does not make it one of today's biggest revenue generators.

The company says its first A0 silicon, manufactured on TSMC's N4P process, has returned successfully and that it is now validating full frontier-inference systems with customers. Etched also disclosed more than $1 billion in signed customer contracts.

Those two facts need to be read together. Customers are apparently willing to commit very large sums before broad production, which is unusual and commercially significant for a semiconductor startup.

But the same company update makes clear that systems are still going through customer validation as Etched prepares to fulfil those contracts. Revenue normally arrives as products or services are delivered under the contract terms.

Etched was only founded in 2022, has already built a team of more than 400 people and has raised enough money to reach a valuation above $10 billion. If it converts most of the existing contract book into shipments, annual revenue could move into the hundreds of millions very quickly.

For now, Etched belongs in the "watch closely" category rather than at the top of a current revenue ranking.

How are AI chip startups actually making money today?

The strongest AI chip companies now make money from much more than chip sales: cloud inference, complete systems and processor IP are becoming just as important.

Cerebras shows the shift most clearly. In the second quarter of 2026, hardware accounted for $54.1 million of revenue while cloud and other services produced $126 million. Around 70% of quarterly revenue therefore came from outside straightforward hardware sales.

Groq has taken a similar route through GroqCloud. Customers can consume inference through an API instead of buying LPUs and building their own infrastructure. The company's 13-data-center footprint makes Groq look increasingly like a specialized AI cloud built around its own processor.

SambaNova combines enterprise systems with hosted inference. Tenstorrent leans much more heavily on processor IP and architecture licenses while also selling Blackhole accelerator products.

The economics are very different. A hardware system can create a large chunk of revenue immediately but requires manufacturing and customer capital expenditure. IP deals can carry attractive margins but depend on a smaller number of large design wins. Cloud inference can create recurring revenue, although the provider has to finance or secure expensive data-center capacity.

Which startup sells the most AI chips and which AI chip startup makes the most revenue are no longer the same question. The companies growing fastest are monetizing the whole compute stack around their architecture.

Is AI inference where AI chip startups are finally making real money?

Yes. AI inference has become the clearest commercial opening for independent chip startups, and most of today's private revenue leaders are leaning heavily into it.

Groq built its LPU around fast, predictable inference. SambaNova's latest SN50 positioning focuses on agentic inference. d-Matrix designed its architecture around inference using digital in-memory compute. Etched is building frontier-inference clusters. Rebellions primarily targets inference deployments, while Axelera AI focuses on efficient inference at the edge.

Cerebras gives us the strongest financial proof. Its cloud and services revenue jumped from $33 million to $126 million year over year in the second quarter. That growth came from customers consuming compute on Cerebras infrastructure rather than buying another machine outright.

Inference gives startups a narrower fight against Nvidia. A customer can continue training models on Nvidia GPUs while moving selected inference workloads to Groq, Cerebras, SambaNova or another accelerator if the price, latency or throughput is better.

The workload also repeats constantly. Training a frontier model happens occasionally and is concentrated among a small number of companies. Inference happens every time a user chats with a model, an agent executes a workflow, software generates code or a multimodal system handles another request.

That is why the revenue leaders increasingly look less like traditional semiconductor vendors and more like AI-compute providers with proprietary chips underneath.

If you want more recent data on this point, please see our latest AI chip market report.

Does having hundreds of AI chip customers mean a startup has high revenue?

No. Customer count tells us whether an AI chip startup is getting real adoption, but it is a poor way to estimate how much money the company actually makes.

Axelera AI says it has shipped products to more than 500 customers worldwide, up sharply from around 100 a year earlier. Those deployments span robotics, industrial manufacturing, retail, public safety, agriculture and other edge-AI applications.

That is strong commercial evidence because many chip startups spend years showing benchmark charts without getting meaningful hardware into customer hands.

Still, 500 edge customers can generate less revenue than one large sovereign AI project. An accelerator inside a camera, robot or industrial computer may sell for a fraction of the value of a data-center rack.

Groq and Cerebras face the opposite dynamic. A relatively small number of large infrastructure customers can move annual revenue by tens or hundreds of millions of dollars.

So Axelera's customer count tells us the company has crossed from prototype to actual market adoption. It does not give us enough information to put Axelera near the top of a revenue ranking today.

How far behind Nvidia are all these AI chip startups combined?

AI chip startups are still tiny compared with Nvidia financially, even after several challengers reached hundreds of millions of dollars in annual sales.

Cerebras generated $510 million in 2025. MetaX produced roughly $230 million. Moore Threads was around $210 million. Biren and Iluvatar each reached roughly the $150 million range. Groq and SambaNova appear to be around $100 million, while Tenstorrent may be higher if it achieved its earlier forecast.

Nvidia operates on a completely different scale. Its fiscal 2026 revenue was $215.9 billion, with Data Center revenue alone reaching $193.7 billion.

Even if we combine Cerebras, Moore Threads, MetaX, Biren, Iluvatar, Groq, SambaNova, Tenstorrent, d-Matrix and Rebellions using the best annual numbers available, the total still lands in only the low billions. Some of those numbers are estimates or forecasts rather than audited sales.

That gap tells us what this market really looks like today. The startup race is about proving that a new architecture can build a durable business beside Nvidia, not about taking Nvidia's revenue crown anytime soon.

Crossing $100 million shows that customers will deploy a new chip architecture. Crossing $1 billion would show that a challenger has built something much harder to dismiss. Very few companies from this generation have reached that second level yet.

Which AI chip startups are generating the most revenue today?

Cerebras is the clear revenue leader from the modern AI-chip startup cohort, while Groq, SambaNova and possibly Tenstorrent sit at the top among companies that are still private.

Cerebras generated $510 million in 2025 and $373.5 million in the first half of 2026. Its increasingly large cloud-inference business also makes the company less dependent on occasional hardware deals.

The next surprise is China. Moore Threads, MetaX, Biren and Iluvatar CoreX have all reached roughly RMB1 billion or more of annual or half-year revenue, which puts several of them ahead of the best-known private US startups on recognized sales.

Among companies that remain private, the numbers are less clean. Groq appears to be around $100 million rather than the old $500 million target. SambaNova looks to be in roughly the same range. Tenstorrent had forecast approximately $200 million, but we still cannot verify that it actually recognized that amount. d-Matrix probably sits in the high tens of millions, while Rebellions reached around KRW35 billion.

Etched could change the ranking quickly. More than $1 billion in signed contracts gives it one of the largest forward order books in the group, but the company is only beginning to validate its first full systems with customers.

The clearest takeaway today is that billion-dollar valuations are far more common than billion-dollar revenues in AI chips. Only a small number of independent architectures have crossed even the $100 million level with evidence strong enough to trust.

Cerebras has already moved well beyond that threshold. China's former startups now form the deepest cluster of commercial challengers. The private US race remains open, and the next real test is simple: which company can turn today's bookings, design wins and infrastructure deals into several hundred million dollars of repeatable recognized revenue.

If you want more recent data on this point, please see our latest AI chip market report.

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.