AI Chip: what are the top startups now?

Which AI chip startups are leading the market right now?

SambaNova, d-Matrix, Etched, Positron AI, FuriosaAI, Rebellions, Tenstorrent, and MatX are among the strongest private AI chip startups today, although their progress varies enormously.

SambaNova has secured major enterprise customers and reached an $11 billion valuation. d-Matrix has moved its Corsair inference accelerator into full production. Etched, valued at $21 billion, has delivered its first hardware to Jane Street, while FuriosaAI has received an initial production batch of 4,000 accelerators. Billion-dollar valuations before volume sales invite a look at what each company is worth, then and now, in our report on top AI chips companies.

Positron and MatX are attracting large investments in future processors, and Tenstorrent is taking a broader approach that combines AI accelerators with processor licensing. Rebellions is emerging as a serious competitor from South Korea.

Two other names deserve attention: Ayar Labs and Lightmatter. They focus on moving data between AI processors using optical technology, which could become just as valuable as improving the chips themselves.

Cerebras would rank among the strongest companies in this group based on actual revenue. However, its May 2026 IPO means we can no longer classify it as a private startup.

The biggest difference between these companies is how much of their success has already been demonstrated. Some are shipping processors and winning customers; others are worth billions largely because investors expect their next chips to be exceptional.

Why are investors pouring billions into AI chip startups now?

Investors are pouring money into AI chip startups because running AI models has become an enormous business, and specialized processors could capture part of the spending currently flowing to Nvidia.

Nvidia's fiscal 2026 results show the scale of the opportunity. Its data-center revenue reached $193.7 billion, up 68% from the previous year. Total revenue climbed to $215.9 billion.

Much of that spending supports both training and inference. Training produces the model, while inference covers the computing required whenever someone uses it. The growth of AI coding tools, reasoning models, and automated agents means inference can consume enormous computing resources.

Several semiconductor startups have raised unusually large amounts to target that demand.

SambaNova announced a $1 billion Series F in July 2026. Etched raised $700 million in August, and Positron announced financing of up to $875 million in September. MatX raised $500 million in February.

Together, those four financing announcements represent $3.075 billion, although some commitments were structured in stages and should not all be treated as cash already received.

Adding the $400 million financing announced by Rebellions and the $650 million in primary funding reported by Ayar Labs brings the total to approximately $4.125 billion across six companies.

That is a striking concentration of money. The average announced financing amount across those six companies exceeds $680 million.

The investment thesis is straightforward: a company that makes inference materially cheaper could win a meaningful share of a market where customers already spend tens of billions of dollars annually.

The catch is that semiconductor development is expensive, and even a promising architecture can run into serious problems during manufacturing, software integration, or customer deployment.

Is Cerebras still the strongest AI chip challenger?

Cerebras is currently the most commercially proven specialist AI chip challenger in this group, although the company's public listing puts it outside a strict private-startup ranking.

In its second-quarter 2026 financial results, Cerebras reported $180.1 million in GAAP revenue, up 74% year over year. Its core revenue, which excludes certain accounting and pass-through effects, reached $209.9 million.

Its cloud business is especially interesting. Cloud and other services revenue grew 281% to $126 million, accounting for about 70% of total quarterly GAAP revenue.

Cerebras is increasingly making money by providing computing services directly, rather than relying only on selling its unusual wafer-scale processors.

The company has relationships with OpenAI, AWS, AMD, CrowdStrike, and other major technology businesses. It also reported more than 600 megawatts of data-center capacity operating or under contract for delivery by the end of 2027.

Still, the financial results expose a serious weakness. Cerebras reported a GAAP gross margin of just 14% and a GAAP operating margin of negative 265% during the quarter.

Its $25.4 billion in remaining performance obligations is impressive, but this figure represents contracted future obligations, subject to the relevant recognition and delivery conditions. It is not revenue already earned.

Cerebras has established something most private AI chip startups have yet to demonstrate: substantial recognized revenue from customers. Whether it can turn that growth into attractive profits remains a much harder question.

Is SambaNova the best private AI chip startup today?

SambaNova is our strongest overall candidate among private AI chip startups, particularly for enterprise inference, thanks to its customers, integrated systems, and financial backing.

The company announced the first close of a $1 billion Series F in July 2026, giving it an $11 billion post-money valuation. The round followed a $350 million financing in February and a strategic partnership with Intel.

The customer announcement accompanying the latest financing is more useful than the valuation itself. JPMorgan Chase selected SambaNova's SN40 and SN50 systems for on-premises AI inference.

That gives SambaNova a credible customer in a demanding industry where security, reliability, and integration with existing infrastructure are critical.

Its processors use a reconfigurable dataflow architecture designed to run AI workloads without following the same execution model as Nvidia GPUs. SambaNova sells complete systems and inference services, allowing customers to adopt the technology without building their own accelerator infrastructure.

The company's trajectory also deserves attention. SambaNova raised $676 million in 2021 at a valuation above $5 billion, before the generative AI boom. Its latest valuation is roughly twice that level, after several years spent developing hardware and software.

For comparison, newer companies such as Etched and Positron have seen much faster valuation increases, but have less publicly documented commercial history.

SambaNova's main weakness is financial transparency. The company does not publish audited revenue figures comparable to those available from Cerebras.

We can confidently place SambaNova among the leaders in private enterprise inference hardware. Calling it the largest private AI chip business by sales would go beyond the available evidence.

Is d-Matrix becoming a real Nvidia competitor?

d-Matrix is now one of the most credible production-stage AI inference competitors, and its recent developments make the company more interesting than its funding total suggests.

Its Corsair accelerator entered full production in June 2026. The company said volume deliveries were beginning for priority customers, including hyperscalers, neocloud providers, and frontier AI laboratories.

Corsair uses an SRAM-based in-memory computing architecture. Instead of repeatedly moving large quantities of information between separate processing and memory components, the design performs certain operations closer to where data is stored. Computing closer to where data sits is one bet on how chips get better, and we look at which technologies are gaining traction in our report on AI chips.

Its manufacturing choices are also deliberate. Corsair uses TSMC's established 6-nanometer manufacturing process, LPDDR5 memory, and packaging that avoids some of the complexities associated with advanced HBM-based systems.

That could make manufacturing and supply more predictable.

More recently, d-Matrix has expanded beyond a standalone accelerator. It acquired Wallaroo.ai's technology to improve inference deployment, adopted Nvidia's NVLink Fusion infrastructure, and introduced Squadron computing clusters designed for data-center-scale inference.

d-Matrix wants its processors to fit into customers' existing AI infrastructure, and much of its recent work has been directed toward making that possible.

The company has cited inference speed improvements exceeding tenfold in particular configurations that combine specialized accelerators with GPUs. Those results are workload-specific and should not be interpreted as Corsair being ten times faster than every Nvidia product.

The combination of hardware production and practical integration is compelling. d-Matrix is tackling the messy part of introducing a new accelerator into an established computing environment, something that can matter more to customers than another impressive benchmark.

It still needs to demonstrate major recurring sales. But today, d-Matrix has stronger evidence of product readiness than many startups carrying larger valuations.

Is Etched really worth $21 billion already?

Etched has become one of the most valuable private AI chip startups, but its $21 billion valuation is still an extremely aggressive bet on future commercial growth.

The company raised $700 million in August 2026, less than a month after a financing round that valued it at $10.3 billion.

Its valuation more than doubled in a matter of weeks.

There is genuine progress behind the increase. Etched delivered its first rack to Jane Street in July, and the trading firm began deploying the technology in its workloads. Jane Street subsequently led the August financing.

Etched also reported more than $1 billion in customer contracts and cumulative fundraising of approximately $1.9 billion.

The company has functioning hardware, an identifiable initial customer, and commitments from other buyers. That's considerably more than an architectural concept and a nice presentation.

The Sohu chip's central idea is particularly bold. It specializes in transformer inference, allowing Etched to dedicate hardware resources to operations used by transformer models rather than supporting every kind of neural network.

That specialization could deliver impressive performance and cost advantages on compatible workloads.

However, there is a risk that future AI architectures evolve in ways that reduce the value of such narrowly designed hardware.

Etched's reported customer contracts also need careful interpretation. Signed commitments can support future revenue, but the company must manufacture, deliver, and operate the systems successfully before the commercial opportunity is fully realized.

Etched has earned a place in the first tier of AI chip startups. Its extraordinary valuation, though, is far ahead of the publicly disclosed revenue evidence.

Is Positron AI growing too fast for its own good?

Positron AI's valuation has increased exceptionally quickly, and investors are now betting heavily on a processor generation that has yet to reach mass production.

According to Reuters and PitchBook, Positron raised $230 million in February 2026 at a valuation of $1.06 billion.

In September, the company announced an $875 million financing package at a $5 billion valuation.

That is a roughly 4.7-fold increase in seven months, equivalent to valuation growth of approximately 372%.

The financing was structured in two tranches, including a $375 million Series C and a Series C-1 of up to $500 million. That distinction matters when comparing announced commitments with capital actually available to spend.

Positron's approach emphasizes memory capacity and bandwidth, which can determine how efficiently a processor serves large language models.

Its first-generation Atlas systems have progressed beyond development. Reuters reported that the company was deploying more than 50 racks at Oracle Cloud Infrastructure.

The larger technological bet is Asimov, its next-generation processor, with production targeted for the second half of 2027.

Positron has also described Titan systems combining several Asimov chips to serve extremely large models and context windows.

The company has achieved early deployment with Atlas, while much of its new valuation depends on Asimov's future performance and adoption.

At $5 billion, Positron is one of the most interesting private AI chip companies. We would rank it below the strongest production-stage competitors on proven commercial execution, but its memory-focused architecture gives it a real chance to move up.

Which AI chip startups have actually shipped products?

The AI chip startups with the strongest physical deployment evidence currently include FuriosaAI, d-Matrix, Etched, SambaNova, and several established accelerator suppliers.

That is a more useful distinction than simply ranking companies by how much funding they have raised. A processor can look excellent in simulation and still face manufacturing problems, unexpected power requirements, software limitations, or production delays.

The available evidence also differs in quality. Receiving manufactured chips is not the same as delivering systems to paying customers, and neither proves strong recurring sales.

Company	Latest meaningful milestone	What the evidence establishes
FuriosaAI	Initial batch of 4,000 RNGD accelerators received	Volume manufacturing achieved
d-Matrix	Corsair entered full production	Product ready for volume deliveries
Etched	First rack delivered to Jane Street	Initial customer hardware deployment
SambaNova	JPMorgan selected SN40 and SN50 systems	Major enterprise customer adoption
Positron AI	More than 50 Atlas racks being deployed at Oracle	Early data-center deployment
Rebellions	Rebel100 and integrated inference infrastructure	Commercial hardware platform available
Tenstorrent	Accelerator products and development platforms	Working hardware and broader IP offering
MatX	MatX One development and financing	Major product milestones still ahead
Ayar Labs	Optical technology moving toward high-volume manufacturing	Advanced component commercialization
Lightmatter	Optical switch and accelerator-system validation hardware	Advanced system demonstrations

Can Tenstorrent succeed without beating Nvidia?

Tenstorrent can become a major AI semiconductor company without matching Nvidia's GPU sales, because its business includes processor technology and intellectual-property licensing.

The company is led by Jim Keller, whose previous engineering roles include work at AMD, Apple, and Tesla.

Tenstorrent develops AI accelerators alongside RISC-V processor technology, giving it several potential customers: companies buying complete computing systems, semiconductor businesses licensing processor designs, and developers building specialized hardware. Selling systems, licensing designs and equipping developers are three revenue paths; we compare which business models are winning in our report covering AI chips.

Its financial backing is substantial. In December 2024, Tenstorrent raised $693 million in a Series D round backed by investors including Samsung Securities, AFW Partners, and Bezos Expeditions.

The financing brought its cumulative reported capital raised to approximately $1 billion.

Tenstorrent's strategy is interesting because Nvidia's strength comes partly from controlling so much of its hardware and software ecosystem. Customers interested in more open designs may find RISC-V and alternative accelerator architectures attractive.

But developing AI processors, licensing intellectual property, and selling complete systems are different businesses. Tenstorrent must prove that it can turn its broad technology portfolio into repeatable commercial demand.

Compared with Etched or MatX, Tenstorrent is less narrowly dependent on one future accelerator becoming a huge success.

That diversification gives Tenstorrent more ways to win, although the lack of detailed public revenue reporting prevents us from ranking it among the commercial leaders.

Can MatX build a better AI chip than Nvidia?

MatX has a credible team and an ambitious processor design, but we still do not have enough production evidence to say that its chips can outperform Nvidia at scale.

The company was founded by former Google TPU engineers Reiner Pope and Mike Gunter. Their experience matters because Google's TPUs are among the few alternatives to Nvidia accelerators already used extensively for advanced AI workloads.

In February 2026, MatX announced a $500 million Series B, bringing its cumulative reported funding to approximately $605 million.

Funding databases have estimated its valuation at around $4.65 billion.

Its MatX One architecture targets high-performance computing for large language models, including both training and inference.

MatX argues that specializing its hardware around the operations used by modern AI models can deliver major improvements over existing processors.

However, the company is still earlier in its commercialization process than d-Matrix, FuriosaAI, or Etched. Its next important milestones concern chip manufacturing, system validation, and expected deliveries in 2027.

Its founders know how to build advanced AI processors, and investors have provided enough capital to pursue an ambitious design. That's a strong starting position, but customers still need to see what MatX One can actually deliver.

We consider MatX one of the most promising next-generation chip startups. Placing it ahead of companies already shipping substantial quantities of hardware would be premature.