What are the top AI inference chip startups by revenue today?

The ranking of top startups in the AI inference chips market by revenue

Below is a table ranking all the companies in this market by their current revenue scale. You can find our methodology at the end of this page.

If you want a deeper understanding of the market and its current dynamics, get our report covering the AI inference chips market.

Which AI inference chip startups make the most revenue today?

Cerebras is currently the strongest revenue leader among Western AI inference chip challengers with verified financial results, while Cambricon leads the broader global comparison when publicly traded Chinese AI accelerator companies are included.

Cerebras reported $510 million in 2025 revenue and approximately $681 million in trailing twelve-month revenue through June 2026. Cambricon generated RMB 6.50 billion in 2025, equivalent to roughly $900 million. Ambarella, another established public semiconductor company, reported $390.7 million for its fiscal year ending January 2026.

The private-company picture is less straightforward. Groq, SambaNova, Tenstorrent, Etched and Hailo have all developed specialized AI processors, but their latest recognized annual revenue is either undisclosed or supported mainly by external estimates.

According to Sacra, Groq generated an estimated $90 million in 2024 revenue. That is meaningful commercial activity, although the figure is too old to establish Groq's precise current position against companies reporting more recent financial results.

We also need to distinguish between companies selling chips, those selling access to inference computing through cloud services, and those developing processors for both AI training and inference. Their reported company-wide revenue does not necessarily represent inference-only sales.

Why are AI inference chip startups becoming such a big business now?

AI inference chip startups are attracting more business today because running AI models has become a massive computing expense, particularly as AI assistants and agents handle increasingly complicated tasks.

Training a model requires large amounts of computing power, but inference happens every time someone actually uses it. A coding assistant, for example, may generate hundreds of model requests while completing a single software engineering task.

This creates opportunities for companies that can deliver AI responses faster or at a lower cost than conventional GPU systems.

Recent investments show how seriously companies are taking that opportunity. Groq raised $650 million in June 2026 to expand its inference cloud, while SambaNova announced a $350 million financing round alongside its new SN50 processor in February 2026. Etched also raised $700 million in August 2026 after delivering its first inference rack to Jane Street.

These companies are starting to compete on commercial execution rather than chip demonstrations alone. Customers increasingly care about how many useful tokens an accelerator can generate, how quickly it can respond and how much operating a complete system costs.

Still, the gap between investor expectations and recognized revenue remains enormous. Some companies are valued in the billions before publishing even one full year of meaningful hardware sales.

How much revenue does Cerebras make today?

Cerebras now generates approximately $681 million in trailing twelve-month revenue, making it one of the most commercially established independent AI accelerator challengers.

Its SEC filings show a remarkable expansion. Revenue increased from $24.6 million in 2022 to $78.7 million in 2023, $290.3 million in 2024 and $510 million in 2025. Across those three years, annual revenue multiplied by approximately 20.7.

The company has continued growing since its IPO. Its second-quarter 2026 financial results showed $180.1 million in GAAP revenue, an increase of 74% from $103.3 million a year earlier.

Cloud and other services generated $126 million during the quarter, compared with $33 million in the same quarter of 2025. That represents approximately 282% growth, far faster than Cerebras's overall revenue expansion.

The business has also changed substantially. Hardware represented approximately 70% of total revenue in 2025, but cloud and other services accounted for approximately 70% in the second quarter of 2026.

Cerebras started as a wafer-scale processor developer. Its fastest-growing business now involves selling access to computing capacity rather than simply shipping processors.

Period	Revenue	Growth	Revenue metric
2022	$24.6M	-	Annual revenue
2023	$78.7M	+220%	Annual revenue
2024	$290.3M	+269%	Annual revenue
2025	$510M	+76%	Annual revenue
Q2 2025	$103.3M	-	Quarterly revenue
Q2 2026	$180.1M	+74%	Quarterly revenue
TTM June 2026	~$681M	+112%	Trailing revenue

Is Groq making more money than Cerebras?

Groq may be one of Cerebras's closest competitors in AI inference, but there is currently no verified revenue disclosure establishing that Groq makes more money.

Sacra estimates Groq generated approximately $90 million in 2024 revenue. It also reports that the company guided investors toward more than $500 million in 2025 revenue, after previously projecting more than $2 billion.

That revision reduced the stated revenue ambition by at least 75%. Sacra attributed the change primarily to data-center capacity constraints that pushed some expected revenue into later periods.

Those were projections. We cannot assume Groq actually recognized $500 million in 2025 simply because management expected to reach that level.

Groq has continued expanding its business. Its August 2026 announcement described 13 operating data centers and more than six million developers, up from five million reported in June.

Its latest financing also provides useful context. Groq raised $350 million in August at a $3.5 billion valuation, following the $650 million raised in June. The two rounds brought $1 billion into the company during a relatively short period.

Groq clearly operates a sizable inference platform. But without a recent revenue disclosure, Cerebras retains the much stronger position in a financial ranking.

Is Cambricon bigger than Cerebras in AI chips?

Cambricon is already bigger than Cerebras based on reported 2025 revenue, although both companies sell AI acceleration technology for more than inference alone.

Cambricon reported RMB 6.497 billion in 2025 revenue, representing approximately $900 million using a representative annual exchange rate. Cerebras generated $510 million during the same year.

That puts Cambricon at roughly 1.8 times Cerebras's revenue.

Cambricon also grew exceptionally quickly. Its 2024 revenue was RMB 1.174 billion, meaning sales increased by 453% in 2025.

The company's commercial expansion has taken place in a Chinese market increasingly focused on domestic AI computing technology. Restrictions on advanced American semiconductor exports have strengthened the incentive for Chinese customers to develop and purchase alternatives.

Cambricon reported approximately RMB 2.06 billion in attributable net income for 2025, compared with a loss the previous year. That's particularly notable in an industry where many AI accelerator challengers remain unprofitable.

However, Cambricon is publicly listed and serves both training and inference workloads. We include it as a major global comparison rather than classifying it as a current private inference startup.

Which AI inference chip companies have disclosed the most revenue?

Cambricon, Cerebras, Ambarella and Moore Threads are among the largest AI accelerator challengers with directly reported annual revenue, while the private specialists generally disclose much less financial information.

The difference becomes striking when we compare the available figures. Cambricon and Cerebras have reached hundreds of millions of dollars in annual sales, whereas companies such as SambaNova, Tenstorrent and Etched have yet to provide equivalent public financial results.

Moore Threads is another important example. Its reported revenue increased from approximately RMB 439 million in 2024 to RMB 1.506 billion in 2025. It then generated RMB 1.736 billion during the first half of 2026 alone, exceeding its entire previous year's sales.

At the smaller end of the market, Blaize reported $38.6 million in 2025 revenue, up from just $1.6 million in 2024. That's an impressive jump, although its product mix and customer concentration deserve closer examination.

We also include Ambarella because inference extends beyond large language models. The company's edge AI processors already serve cameras, vehicles and other physical devices.

Company	Financial evidence	Period	Market focus
Cambricon	~$900M revenue	2025	Training and inference
Cerebras	~$681M revenue	TTM June 2026	AI accelerators and cloud
Ambarella	$390.7M revenue	FY2026	Edge AI inference
Moore Threads	~$209M revenue	2025	GPUs and AI acceleration
Groq	~$90M estimated revenue	2024	Cloud inference
Blaize	$38.6M revenue	2025	Edge and data-center inference
SambaNova	Undisclosed	-	Inference systems and cloud
Tenstorrent	Undisclosed	-	AI processors and licensing
Etched	Undisclosed	-	Transformer inference ASICs
Hailo	Undisclosed	-	Edge AI processors

The figures are deliberately labeled by metric and period. Groq's figure is an external estimate, while the other financial amounts shown here come from company reporting. The table compares commercial scale, not inference-only revenue, and does not place undisclosed businesses below companies with reported sales.

How much revenue does SambaNova generate?

SambaNova has become a serious AI inference hardware competitor, but its latest annual revenue remains undisclosed.

The company reported record bookings and revenue when announcing its SN50 accelerator in February 2026. However, it did not provide the dollar amount needed to compare its financial scale with Cerebras or Groq.

SambaNova has developed several generations of Reconfigurable Dataflow Unit processors, which are designed to handle demanding AI workloads without relying entirely on conventional GPUs.

Its latest product strategy focuses heavily on inference for AI agents. The company claims the SN50 can deliver up to five times the speed of competing chips and a threefold lower total cost of ownership than GPUs under its specified comparisons.

SoftBank was announced as the first customer expected to deploy SN50 chips in next-generation Japanese AI data centers. SambaNova also revealed a planned collaboration with Intel covering cloud-scale inference infrastructure.

These partnerships are commercially significant, but the amounts involved have not been disclosed in a way that establishes annual revenue.

SambaNova could already be generating substantial sales. We can confirm serious customer relationships and product development, but there's not enough financial evidence to give it a reliable revenue rank.

Is Tenstorrent already making serious money from AI chips?

Tenstorrent has attracted substantial investment and major industry attention, but its annual revenue is still not publicly established.

The company is led by semiconductor engineer Jim Keller and develops programmable AI processors built around its own architecture and RISC-V technology.

Unlike companies concentrating entirely on selling cloud inference, Tenstorrent pursues several commercial opportunities. These include processor products, complete computing systems and intellectual-property licensing.

Its December 2024 financing round raised approximately $693 million, demonstrating how much investors were willing to commit to alternative AI computing architectures.

But investment and sales tell us different things about Tenstorrent. Licensing agreements can generate revenue through development fees, milestones and royalties, while hardware revenue generally depends on actual product deliveries.

A large licensing announcement therefore tells us something about future commercial potential without necessarily revealing current recurring revenue.

Tenstorrent remains one of the industry's better-known private AI chip developers, but we would need disclosed sales or licensing revenue before treating it as a confirmed revenue leader.

Could Etched become one of the biggest AI inference chip companies?

Etched has rapidly become one of the most valuable specialized inference-chip startups, but it is still early in proving how much revenue its technology can generate.

In August 2026, Etched announced that it had shipped its first Sohu-powered rack to Jane Street. The same announcement disclosed a $700 million financing round at a $21 billion valuation.

Jane Street confirmed that it had tested the hardware and was operating a rack in its own data center. That gives Etched something more substantial than an engineering demonstration: a real customer deployment.

Etched also takes an unusually focused technical approach. Its Sohu architecture is built specifically for transformer inference instead of attempting to support the broad range of workloads handled by general-purpose GPUs.

That specialization could translate into better performance and lower costs for supported models. The trade-off is a narrower range of workloads and greater exposure to changes in AI model architectures.

The company's valuation is remarkable considering that its first publicly announced rack shipment came only recently. Investors are clearly assigning enormous value to expected future production.

Still, one delivery cannot establish a meaningful annual revenue run rate. Etched's position among the revenue leaders will depend on how quickly production shipments grow from this initial deployment.

Which AI inference chip startups make more than $100 million annually?

The number of AI inference-chip challengers with verified annual revenue above $100 million is smaller than their funding announcements might suggest.

Cerebras, Cambricon, Ambarella and Moore Threads have all reported annual revenue above that threshold, although they are already publicly traded and some sell processors for training as well as inference.

For independent private specialists, the evidence is less complete. Groq's $90 million estimated revenue in 2024 falls below $100 million, while its later projections suggested much higher sales without providing a confirmed actual figure.

SambaNova announced record revenue in 2025 but did not disclose the amount. Tenstorrent and Etched also lack comparable public annual revenue figures.

Look at the valuations. Etched reached a $21 billion valuation shortly after announcing its first rack shipment, while Groq received a $3.5 billion valuation in its August 2026 financing round.

Neither valuation tells us whether the company generated more than $100 million in annual sales.

The confirmed nine-figure group is therefore dominated by businesses with public financial reporting. Several private companies may already qualify, but the published evidence does not establish how many.

Are Chinese AI inference chip startups catching up with American competitors?

Chinese AI accelerator companies are growing extremely fast commercially, with Cambricon and Moore Threads providing some of the clearest evidence.

Cambricon increased its reported annual revenue by 453% in 2025. Moore Threads grew 243% over the same year, then reported a further 147% increase in first-half 2026 revenue compared with the previous year's first half.

Those are unusually high growth rates, even for an expanding semiconductor market.

Other Chinese developers, including Biren, Muxi and Suiyuan, are also pursuing the market for domestic AI computing hardware. The competitive environment has been shaped by restrictions on advanced foreign accelerators and strong demand for locally available alternatives.

The financial growth is real, but we should separate it from claims about technical superiority. Revenue can expand quickly when customers have fewer alternative suppliers, even if performance, software support and manufacturing capabilities remain behind global leaders.

For Western inference specialists, the implication is significant. Chinese competitors are already reporting substantial sales, not simply preparing to challenge American suppliers in some distant future.

Their growth also shows why a ranking limited to American startups would miss a large part of the emerging AI accelerator market.

Is Ambarella one of the largest AI inference chip businesses?

Ambarella is already a substantial edge AI inference-chip business, with $390.7 million in fiscal 2026 revenue and approximately $1 billion in cumulative edge AI revenue.

The company reported 37.2% annual revenue growth, up from $284.9 million the previous year. Its management also stated that edge AI systems-on-chip accounted for approximately 80% of fiscal 2026 revenue.

That implies roughly $313 million of annual sales associated with its edge AI processor portfolio.

Ambarella has reached a scale that many better-funded AI chip startups have yet to demonstrate. It reported shipping approximately 45 million edge AI systems-on-chip cumulatively and supporting more than 200 AI model architectures.

Its customers operate in markets such as automotive cameras, security systems, industrial devices and robotics. These applications require AI processing directly inside physical products rather than at remote cloud data centers.

The company's June 2026 long-term agreement with Hanwha Vision also highlights how edge inference chips can become part of multi-generation product designs.

Ambarella is a mature public semiconductor business, so it belongs in the broader industry comparison rather than a strict startup list. Its revenue nevertheless makes it an important benchmark for the commercial maturity of edge inference.

How much money is Blaize making from AI inference chips?

Blaize generated $38.6 million in 2025 revenue, a major increase from $1.6 million in 2024, but its business is still at an early stage.

Its full-year results showed revenue multiplying by nearly 25 times. Fourth-quarter revenue reached approximately $23.8 million, more than double the previous quarter.

That gives us a better indication of how rapidly commercial deliveries accelerated during the year.

However, Blaize's regulatory filings contain two important details. Some 2025 revenue came from third-party hardware sold alongside its own semiconductor products, and approximately $35.2 million of annual revenue came from China.

China therefore accounted for roughly 91% of total reported sales.

That's a lot of geographic concentration, making the company's financial performance sensitive to a relatively narrow commercial market.

Blaize has moved beyond negligible sales, but the durability of its revenue will depend on repeat orders, broader customer adoption and a larger contribution from its own technology.

Are edge AI inference chips a bigger business than cloud inference?

Edge AI inference already supports hundreds of millions of dollars in annual semiconductor sales, although available company disclosures do not establish that the total edge market is larger than cloud inference.

Ambarella provides a clear example. Its edge AI products are used in cameras, robots and other devices where processing needs to happen locally.

Cloud-focused companies such as Groq and Cerebras operate differently. Their customers can purchase inference capacity without installing proprietary hardware, making revenue dependent on ongoing computing consumption.

The two approaches also serve different economic needs.

A security camera may need to recognize objects continuously while operating within strict power limits. A cloud coding assistant may instead need to generate thousands of tokens rapidly while handling many simultaneous requests.

Hailo and SiMa.ai are developing specialized processors for local AI workloads, while SambaNova and Groq target larger computing systems and cloud deployments.

These companies are often grouped together because they all accelerate inference, but their customer bases, sales cycles and revenue models differ substantially.

Ambarella's established sales show that edge inference is already commercially mature in several device categories. Cloud inference, meanwhile, is attracting some of the most aggressive new infrastructure investments.

Did Nvidia's deal with Groq change the AI inference chip market?

Nvidia's licensing agreement with Groq was a major strategic development because it gave Nvidia access to specialized inference technology while allowing Groq's cloud business to continue independently.

The agreement was announced in December 2025. Groq founder Jonathan Ross and several senior employees joined Nvidia as part of the arrangement.

Nvidia's subsequent annual filing described approximately $17 billion of consideration, including $13 billion paid at closing and $4 billion payable within one year.

Importantly, Nvidia stated that it did not acquire Groq's existing products, customer contracts or equity interests.

That structure explains why Groq was able to continue operating and raise additional capital afterward. Nvidia obtained technology and expertise, while the remaining company concentrated on expanding GroqCloud.

The relationship became more concrete when Nvidia introduced its next-generation LPX platform incorporating Groq inference technology.

A multibillion-dollar licensing commitment from the world's dominant GPU supplier says quite a lot about the strategic value of specialized inference architectures.

But that licensing consideration should never be confused with Groq's annual operating revenue.

Are AI inference chip startups making more money from chips or cloud services?

Among the leading Western AI inference specialists, cloud services are becoming an increasingly important revenue source, but the industry still contains several very different business models.

Cerebras illustrates the change most clearly. Its hardware generated $358.4 million in 2025, compared with $151.6 million from cloud and other services.

By the second quarter of 2026, the proportions had effectively reversed.

Groq has also concentrated on inference as a managed cloud service. Customers purchase computing capacity through GroqCloud rather than necessarily owning the underlying processors.

SambaNova takes a mixed approach, combining accelerators, complete systems and cloud inference. Tenstorrent adds intellectual-property licensing to the mix.

The economics are different. Selling hardware can generate large amounts of revenue when systems are delivered, while operating an inference cloud can create repeated customer spending over time.

Cloud businesses also carry substantial infrastructure costs, and their financial results depend heavily on how efficiently installed capacity is used.

The shift toward cloud services could eventually give inference specialists more recurring revenue. Cerebras already provides clear reported evidence of this transition, while other private competitors still need to disclose comparable figures.

Can AI inference chip startups really beat Nvidia on price?

Specialized AI inference processors can outperform Nvidia GPUs on selected workloads, but no company has demonstrated a universal cost advantage across every major AI application.

Groq's architecture focuses on fast, predictable token generation. Cerebras uses wafer-scale processors to deliver high-speed inference, while SambaNova's SN50 was introduced with claims of lower total operating costs for agentic AI workloads.

These are meaningful technical approaches, but performance depends on the actual model and deployment configuration.

A chip that produces tokens extremely quickly for one transformer model may be less attractive for another workload requiring different memory capacity, precision or software support.

The full cost of inference also includes electricity, networking, server utilization, engineering and the ability to keep expensive hardware busy.

Nvidia benefits from CUDA, a broad software ecosystem and support across a wide range of models and applications. That flexibility can compensate for performance differences on individual benchmarks.

For customers, the useful comparison is the total cost of running a specific production workload at the latency and reliability they need.

Specialized hardware will likely win some of those comparisons. The evidence available today doesn't justify claiming that inference startups have broadly displaced Nvidia on cost.

Are AI inference chip startups profitable yet?

Profitability remains a major dividing line in the AI inference-chip market, even among companies that already generate substantial revenue.

Cambricon reported approximately RMB 2.06 billion in attributable net income for 2025, showing that a domestic AI accelerator manufacturer can reach profitability at scale.

Cerebras reported $237.8 million in GAAP net income for 2025, but the result included significant accounting effects. Excluding specified items, the company reported a non-GAAP net loss of $75.7 million.

Its more recent operating results show how expensive expansion can become. In the second quarter of 2026, Cerebras reported a GAAP operating margin of negative 265%, compared with a core non-GAAP operating margin of negative 16%.

Ambarella reported a fiscal 2026 GAAP net loss of approximately $75.9 million despite generating nearly $391 million in revenue.

Moore Threads also remained loss-making in 2025, although its first-half 2026 results showed a much smaller attributable net loss of approximately RMB 11.6 million.

Strong demand can support rapid semiconductor revenue growth, but profitability depends on research spending, manufacturing costs, gross margins and the amount of infrastructure needed to serve customers.

For cloud inference providers, keeping expensive processors highly utilized is especially important. Revenue growth alone cannot tell us whether a business is making attractive returns.

Which AI inference chip startups have the biggest customer contracts?

Cerebras currently has some of the strongest publicly disclosed commercial commitments among alternative AI accelerator companies, with $25.4 billion in remaining performance obligations reported in its second-quarter 2026 results.

That figure is extraordinary compared with the company's current revenue, but it represents future contractual obligations rather than sales already earned.

A major part of the balance relates to Cerebras's agreement with OpenAI. Its quarterly filing described a multiyear services arrangement and additional capacity options that could significantly expand the relationship.

The company also reported more than 600 megawatts of data-center capacity either operating or under contract for delivery by the end of 2027.

Groq provides a different kind of evidence. As seen earlier, its cloud platform already serves millions of developers across a global network of data centers, although it does not publish an equivalent contracted-revenue figure.

SambaNova has announced SoftBank as a customer for its new SN50 systems, while Etched's Jane Street deployment represents a first step into production use.

Cerebras has substantial financial obligations attached to future services, Groq has large operating usage, and Etched has demonstrated an initial commercial shipment.

These companies are at different stages of turning demand into recognized revenue. Cerebras currently provides the clearest financial visibility.

Could cheaper AI inference hurt chip startups?

Cheaper inference could put serious pressure on AI chip startups, especially companies selling cloud computing capacity, unless customers increase their usage fast enough.

A cloud provider charging customers per token faces a simple challenge. If the average price of processing tokens falls by 50%, it must double token volume just to maintain the same revenue, assuming its pricing mix remains unchanged.

That creates very different outcomes depending on the company's ability to attract more customers and process workloads efficiently.

For Groq and Cerebras, cheaper inference can make AI applications economically viable at much larger volumes. But it can also reduce revenue per unit of computing capacity.

Hardware suppliers face a related challenge. They need to make their processors faster and cheaper while competing against new generations of GPUs and alternative accelerators.

The biggest commercial risk is sustained price pressure combined with expensive infrastructure expansion.

If inference demand grows faster than prices fall, the market can still support much larger businesses. But the companies with the lowest costs and highest utilization will have a considerable advantage over those competing mainly on advertised processing speed.

Which AI inference chip startups could reach $1 billion in revenue next?

Cerebras is currently the clearest Western candidate to reach $1 billion in annual revenue, while Groq and SambaNova remain potential contenders whose latest financial scale is much less transparent.

Cerebras's approximately $681 million trailing revenue means it needs roughly 47% additional growth to reach $1 billion.

That's a demanding target, but it is below the company's recent historical growth rates. Its large customer commitments and planned capacity expansion also provide evidence of potential demand beyond the next few quarters.

Groq has increased its cloud footprint and developer adoption, but its latest recognized annual revenue remains undisclosed. We therefore cannot calculate a reliable distance to $1 billion.

SambaNova's SN50 rollout could increase commercial sales if announced customer deployments translate into substantial deliveries. Tenstorrent's licensing model offers another potential route to revenue, although the company has not published enough financial information to forecast a billion-dollar milestone.

Etched faces a different challenge. Its first rack delivery establishes that the hardware is reaching customers, but moving from an initial shipment to a billion-dollar business requires sustained large-scale production.

The next major revenue winner will likely be the company that can turn growing demand into repeatable sales while expanding manufacturing and computing capacity without losing control of costs.

So, what are the top AI inference chip startups by revenue today?

Cerebras is the most convincingly established revenue leader among Western AI inference-chip challengers, while Cambricon leads the broader comparison of reported 2025 revenue when Chinese public AI accelerator companies are included.

The financial evidence is strongest for companies already reporting audited or regularly filed results. Cambricon generated approximately $900 million in 2025 revenue, Cerebras reached around $681 million on a trailing basis, and Ambarella reported nearly $391 million for its latest completed fiscal year.

Moore Threads has also grown quickly, with first-half 2026 revenue already exceeding its full-year 2025 figure.

Among private companies, Groq stands out through its cloud infrastructure, customer usage and historical revenue estimates. SambaNova, Tenstorrent and Etched remain important competitors, but public financial disclosures are still too limited to rank them confidently by current annual sales.

Financial success and strategic value can also diverge dramatically. Nvidia's Groq technology agreement and Etched's multibillion-dollar valuation show how highly investors and technology companies value specialized inference architectures.

Those valuations may eventually be supported by large recurring revenue streams, but that has not yet been established for every company.

Cerebras offers the strongest combination of independently reported revenue growth, expanding cloud activity and large customer commitments among Western AI accelerator challengers. Groq is one of its most important competitors, while Chinese companies and established edge AI suppliers make the global market considerably larger and more competitive than a private US startup ranking would suggest.