AI chips: which startup is ahead?

Which AI chip startups belong in this comparison?

We rank 12 serious AI chip startups, with 10 aimed mainly at data centers and two focused on edge AI.

The data-center group includes Groq, SambaNova, Etched, Tenstorrent, Rebellions, d-Matrix, FuriosaAI, Positron and Fractile, plus Cerebras as a startup-origin company. Cerebras listed on Nasdaq in May 2026, so it is no longer privately held. We keep it in the field because it came out of the same venture-backed race and now gives us the clearest benchmark for what commercial success looks like.

Groq stays in the field, with a caveat. After its December 2025 licensing deal with Nvidia, it now sells inference as a cloud operator and runs Nvidia systems next to its own LPUs. It is still an AI compute startup, but less of a pure chip company than it was.

SiMa.ai and Axelera AI form the edge group. Their chips run inside cameras, robots, drones, vehicles, factories and local servers. A low-power edge accelerator and a giant data-center system solve different problems, so we compare them on the evidence that fits their market rather than pretending that every chip belongs on one identical scale.

We exclude Nvidia, AMD and Intel, the internal chips made by hyperscalers, acquired startups such as Graphcore and Celestial AI, and companies now focused mainly on optical interconnects. Hailo, bought by Microchip in September 2026, and Taalas, which AMD agreed to buy in August 2026, stay in the tables for reference but are no longer ranked. Capital figures below are cumulative and mostly company-stated or reported; private totals are approximate.

Company	Main position	Cumulative capital raised
Cerebras	Wafer-scale training and high-speed inference systems	About $2.8B privately; IPO: $5.55B at pricing, about $6.4B gross with extra shares (company-stated)
Groq	Inference cloud running its own LPUs and Nvidia systems	About $2.75B (estimate), including $350M in August 2026
SambaNova	Enterprise and sovereign-AI inference systems	More than $2.4B
Tenstorrent	AI accelerators, RISC-V processors and licensable IP	About $2.5B, including a $1.5B Series E (announced)
Rebellions	Data-center inference cards, servers and racks	About $850M
Etched	Highly specialized frontier-inference clusters	More than $1.9B (company-stated)
Axelera AI	Edge and enterprise inference processors	More than $450M
d-Matrix	Memory-centric data-center inference accelerators	$450M
SiMa.ai	Physical-AI and embedded inference processors	$355M
Hailo	Low-power edge-AI accelerators	$340M; acquired by Microchip in September 2026
Positron	Memory-rich transformer inference systems	About $1.2B (estimate), including an $875M Series C
FuriosaAI	Energy-efficient data-center inference processors	$246M disclosed; an 800B-won (roughly $550M) Series D is still being raised
Fractile	Memory-heavy chips planned for frontier inference	About $235M
Taalas	Model-specific chips with computation built into silicon	More than $200M; AMD agreed to buy it in August 2026

Is Cerebras still the clear AI chip leader after its IPO?

Yes. Cerebras sits in a different commercial league from every private AI chip startup.

In its second quarter, which ended in June 2026, Cerebras reported $180.1 million in GAAP revenue, 74% more than a year earlier. Its own core revenue measure, which adjusts for items the company treats as non-recurring, reached $209.9 million, up 103%. Cloud and services revenue jumped 281% to $126.0 million. That last figure matters most: Cerebras is building a repeat-use computing service, not just selling boxes.

The order book is even more telling. Cerebras reported $25.4 billion in remaining performance obligations at the end of June. It has a multi-year OpenAI agreement covering 750 megawatts and valued at more than $20 billion, and AWS plans to bring Cerebras-based inference to Amazon Bedrock in the first quarter of 2027. No private rival has disclosed anything close. Renting out computing instead of selling boxes is why we compare which business models are winning in our AI chips market report.

The listing has not been smooth. Shares fell about 20% after the first post-IPO report in June 2026, when Cerebras cut its margin forecast and said it would rent back some systems from a customer while its own data centers come online. In late September 2026 the stock fell below its $185 IPO price as the lockup ended and a report said OpenAI’s next Ultrafast tier would run on Nvidia hardware. The second quarter also showed a GAAP net loss of $450.5 million. Those are real pressures, but they come from a company growing into its backlog, not from a lack of demand.

The market still has one leader. The real argument is who deserves second place, and the answer has moved: SambaNova and Etched now look stronger than Groq as independent chip challengers.

Which AI chip startup has turned its technology into real sales?

Paid demand puts Cerebras first. Among private companies, Etched now has the clearest contracted demand, and Groq the broadest visible usage.

Cerebras generated $126.0 million from cloud and other services and $54.1 million from hardware in its second quarter, on a GAAP basis. Hardware fell year over year while cloud nearly quadrupled. Customers both buy the machines and rent the computing power, and the mix is swinging fast toward rental. New cloud customers named in August 2026 include Cognition and Lovable.

Etched says it holds more than $1 billion in signed contracts with frontier AI companies and cloud providers. In July 2026 it shipped its first rack to Jane Street, which is putting the system into production. That is one disclosed rack for one customer, and Jane Street is also Etched’s lead investor, so the endorsement is not independent. Still, it moves Etched from a reservation story to delivered hardware.

Groq says it serves more than six million developers across 13 data centers. That is far more visible usage than any other private rival. The missing numbers are the same as before: revenue, paid-customer count and gross margin. And Groq now serves part of that demand on Nvidia systems, so the usage no longer proves demand for its own chip.

SambaNova has the strongest named-enterprise evidence through JPMorganChase, which runs on-premises inference on SN40L and SN50 systems, but it has not disclosed the value of that installation. Tenstorrent’s largest deployment is at the Japanese cloud ai&, with more than 120 Galaxy systems reported. FuriosaAI, d-Matrix and Rebellions are shipping, and none has published revenue close to Cerebras.

Company	Best evidence of demand	Main gap in the evidence
Cerebras	$180.1M GAAP quarterly revenue, $25.4B in remaining performance obligations	Heavy exposure to a few large customers
Groq	More than 6M developers across 13 data centers (company-stated)	No disclosed revenue; part of the capacity runs on Nvidia systems
Etched	More than $1B in signed contracts and a first rack delivered to Jane Street	Only one delivery disclosed; revenue recognition has barely begun
SambaNova	JPMorganChase on-premises deployment	Contract size and SN50 volumes are undisclosed
Tenstorrent	More than 120 Galaxy systems at ai& and a 96-system order	Total revenue and repeat orders are unknown
FuriosaAI	Volume hardware and expanding infrastructure access	Customer revenue remains private

Which AI chip startup is growing fastest right now?

At commercial scale, Cerebras is growing fastest. Etched is moving fastest from first silicon to delivered hardware.

Cerebras doubled core revenue year over year in its second quarter and raised its full-year guidance to $880 million to $890 million of core revenue. Management says it plans to more than triple revenue in 2027, which is a plan, not a result. Adding several hundred million dollars in annual sales carries far more weight than doubling a tiny pilot business.

Etched compressed a normal semiconductor journey into a few years. Its first chip came back working from TSMC’s N4P process in early 2026. It then raised $300 million at a $10.3 billion valuation in July and $700 million at $21 billion in August, and shipped its first rack in between. The next test is the second, tenth and hundredth rack.

Groq’s developer count rose from more than two million to more than six million in about a year, according to the company. We still cannot separate paying production use from free or experimental traffic.

SambaNova’s growth is easier to see in products and financing than in sales. It introduced the fifth-generation SN50 in February 2026, raised more than $350 million, and in July completed the first close of a $1 billion Series F at an $11 billion valuation. SN50 shipments were due in the second half of 2026, with SoftBank as first deployment partner. Now it needs to show the revenue curve.

Who has the most mature AI chip product?

Cerebras owns the most mature complete platform. SambaNova has the most mature private system sold for on-premises use.

Cerebras is on its third CS system generation. Customers can install it, rent hosted capacity or reach it through partners. OpenAI uses it for Codex-Spark and, since August 2026, for a preview of an Ultrafast mode of GPT-5.6 Sol, so the product runs demanding live applications rather than trials.

SambaNova’s SN40L systems are commercially available and run inside JPMorganChase, and SN50 is the next generation. Tenstorrent sells cards, workstations and Galaxy clusters. d-Matrix has Corsair in production and in October 2026 announced Squadron compute clusters for data-center scale. Positron says Atlas is shipping and lists its next Asimov chip and Titan system for 2027.

Groq built its LPU into a service across 13 data centers, reachable through an OpenAI-compatible API. That simple route into production explains its user base. But the product people use is now a cloud, not a chip, and Groq has added Nvidia systems to it.

FuriosaAI made the clearest jump among the smaller challengers. It received its first volume batch of RNGD accelerators, launched complete servers and began placing capacity with partners in Korea and Europe. Etched has delivered its first rack, which is a real product but not yet a mature one. Taalas had a working demonstrator and beta API around one small model. Fractile remains pre-product.

Among edge companies, Hailo-10H, SiMa.ai Modalix and Axelera’s Metis platform are commercially available. Axelera launched its Europa architecture in September 2026 with Dell and Supermicro as partners, and SiMa.ai is pushing into drones with partners such as AVerMedia and ARK Electronics.

Which AI chip startup actually performs best?

On production evidence, Cerebras performs best. Taalas still holds the most extreme single-model speed claim.

Cerebras says OpenAI’s GPT-5.6 Sol Ultrafast preview runs at up to 750 output tokens per second on its hardware, with the same weights and precision as the standard endpoint. On a sample of GDPval tasks, Cerebras says the work finished 5.6 times faster. Codex-Spark has been reported above 1,000 output tokens per second. These are company figures, but they come from a live product used by a frontier lab.

Taalas says its HC1 demonstrator can run Llama 3.1 8B at about 17,000 tokens per second for one user. The chip has the model built into silicon and uses aggressive low-bit quantization, and Taalas acknowledges some quality loss. It shows that radical specialization can unlock another order of magnitude, not that Taalas makes the best general AI chip. Silicon built around one model is an extreme bet, so we follow which technologies are gaining traction in our report covering AI chips.

SambaNova is the strongest speed challenger across a broader range of accessible models. Artificial Analysis measured its MiniMax M2.7 service at about 401 output tokens per second, while smaller models have exceeded 700. Its design pairs other hardware for prefill with SambaNova RDUs for decoding. d-Matrix is now taking a similar split approach, pairing Corsair with Nvidia H100 and adopting Nvidia’s NVLink Fusion for its Raptor chip.

A Harvard-led academic comparison of Cerebras, SambaNova, Groq, Intel Gaudi, Google TPU, Nvidia and AMD found different winners as batch size, sequence length and model size changed, and measured idle power on Cerebras, SambaNova and Gaudi at 10% to 60% above Nvidia and AMD systems. A fast specialist chip needs steady use before its peak efficiency turns into lower electricity bills.

Etched has not published independent performance comparisons for its delivered rack. Tenstorrent has credible results in narrower workloads, including a production text-to-speech study that found roughly four times lower accelerator cost than an Nvidia L40S at equal throughput. Customers should benchmark their own workload rather than trust a generic leaderboard.

Which AI chip startup gives customers the best value?

For large cloud workloads, Cerebras still has the best proven economics. Tenstorrent offers the cheapest serious entry into alternative hardware.

Cerebras reported a core gross margin of 40.6% in its second quarter and guides 41% to 43% for the full year. Its GAAP gross margin was much lower, at 14.2%, and the company had already cut its margin forecast in June while it rents capacity it has not yet built. Buyers are paying enough for the speed to support the product, but the margin is thinner than it looked earlier in the year.

Groq can be attractive because customers use an API and avoid buying racks or maintaining a new stack. We cannot tell whether the pricing works for Groq because it does not publish revenue or margins, and its costs now include Nvidia equipment.

Tenstorrent lets a developer buy a Blackhole card for $999 and move into larger systems later. That is a far easier start than a Cerebras or SambaNova installation. Its text-to-speech results and its October 2026 partnership with Smallest.ai for on-premises voice AI suggest it can beat mainstream GPUs on carefully tuned workloads.

FuriosaAI, d-Matrix, Positron and the edge specialists sell mainly on electricity, memory and infrastructure savings. Furiosa’s air-cooled RNGD fits ordinary enterprise racks. SiMa.ai and AVerMedia showed a 50 TOPS drone module below 10 watts in August 2026. Hailo and Axelera can replace larger GPUs inside cameras and industrial equipment. Most published comparisons still come from the vendors, so real deployments carry more weight than theoretical tokens per dollar.

Etched and Taalas promise large efficiency gains from specialization. Customers gain speed and power savings but lose much of the flexibility that makes GPUs convenient. The value could be excellent for a stable, high-volume model and poor for a model that changes every few months.

Which AI chip startup can actually deliver at scale?

No rival can support deployments as large as Cerebras, and its manufacturing expansion is widening the gap.

Cerebras says it will grow manufacturing capacity more than tenfold in 2026 through lines at Flex, Sanmina and Rocket EMS, with TSMC wafer supply secured. It reports more than 600 megawatts of data-center capacity live or under contract for delivery by the end of 2027. About $8.6 billion in cash and investments at the end of June gives it the means to attempt that buildout.

The OpenAI commitment remains a huge operational test. Supplying 750 megawatts requires sites, electricity, cooling, networking, wafers, packaging and trained teams. Cerebras is renting back systems from a customer while its own capacity comes online, a sign that the buildout is tight.

Groq operates 13 data centers across North America, Europe, the Middle East and Asia-Pacific, with 54 megawatts today and a company target above 200 megawatts in 2027. Its geographic reach is still the best among private rivals, but some of that capacity will run Nvidia hardware rather than Groq chips.

Etched now has the money to build. Its August 2026 round funds factories, supply chains and fleet software, and the company says it is building three hardware generations in parallel. One delivered rack does not prove volume production. FuriosaAI received 4,000 RNGD accelerators in its first volume batch, Tenstorrent says its Series E will fund volume manufacturing, Rebellions sells RebelRack, and d-Matrix is moving to rack-scale Squadron clusters.

Taalas showed that model-specific chips face a different scaling problem: each major model change may require fresh silicon. Fractile has no shipping product to scale.

Which AI chip startup has the best customers?

OpenAI and AWS give Cerebras the strongest customer position, while SambaNova has the cleanest blue-chip enterprise proof among private hardware companies.

OpenAI runs live products on Cerebras and has committed to a much larger capacity build. AWS plans disaggregated inference on Bedrock that pairs other hardware for prefill with Cerebras for decoding. Cerebras also named Block, Figma, GSK and CrowdStrike among its users in August 2026. These relationships involve product integration and long-term planning, which gives them far more weight than a pilot.

Cerebras does have a concentration problem. Its filings name OpenAI, G42, Mohamed bin Zayed University of Artificial Intelligence and AWS as significant customers, and G42 and MBZUAI produced 86% of its 2025 revenue. A delay from one of them would hurt. Relying on a couple of customers for most revenue is a weak spot; we dig into the biggest unsolved problems in our AI chips report.

SambaNova’s JPMorganChase deployment is the strongest private-enterprise reference. A global bank running inference on its own premises sets demanding requirements on security and control. SoftBank, Saudi Aramco and several Japanese groups offer other routes into large installations, although contract values remain private.

Etched’s only named customer is Jane Street, which is also its lead investor; its other contracts are unnamed. Groq has broad developer reach. Tenstorrent and Rebellions both signed with the Japanese cloud ai& in September 2026, and Rebellions ran SK Telecom’s A.X K1 model on its servers in July, which shows its Korean telecom ties are turning into real workloads.

What can the leading AI chip startups do that rivals cannot easily copy?

Cerebras is the hardest architecture to copy. Tenstorrent offers the clearest escape from replacing one closed ecosystem with another.

Cerebras builds one processor from almost an entire silicon wafer. The hard part includes the chip, on-wafer memory, cooling, power delivery, compilers and years of manufacturing knowledge. Copying one circuit would barely begin to reproduce the system.

Tenstorrent spreads its bets across accelerators, RISC-V processors, licensable chip designs and open-source software. TT-Forge supports PyTorch, JAX and ONNX. Customers can buy the hardware, license the IP or work with the software, which gives Tenstorrent more routes into the market than a single closed accelerator.

Groq’s moat has narrowed the most. Nvidia took a non-exclusive license to its LPU technology in December 2025 and hired founder Jonathan Ross, president Sunny Madra and other engineers. Groq says chief executive Adam Winter now leads it, and Nvidia joined its August 2026 round. Nvidia now sells the licensed technology as Groq 3 LPX. Groq’s edge is now its cloud footprint and developer base more than its silicon.

SambaNova’s moat sits in its full dataflow system and enterprise deployment stack. Etched gains speed by specializing around transformer-style inference. Taalas goes further by turning one model into a chip. Both approaches can be extremely efficient, but their durability depends on how fast model designs change.

Across the field, CUDA remains the practical barrier, and more startups are choosing to work with Nvidia rather than against it: d-Matrix adopted NVLink Fusion in September 2026, and Groq now runs Nvidia systems. Cerebras hides its hardware behind cloud interfaces, Tenstorrent opens more of the stack, and SambaNova controls the full deployment. The winners make it easy to move a working model. Nobody buys a chip just to admire the silicon.

Is the best-funded AI chip startup using its money efficiently?

On the numbers we can verify, Cerebras turns capital into revenue far better than the other heavily funded challengers.

Cerebras raised about $2.8 billion privately before its IPO and produced roughly $510 million of revenue in 2025. That is a revenue-to-private-capital ratio near 18%. It is not a return calculation, but it says more than comparing valuations. The IPO priced at $185 a share and raised $5.55 billion, Reuters reported in May 2026. Cerebras reports about $6.4 billion in gross proceeds, a figure that appears to include the extra shares sold to underwriters.

Groq is the hardest case. It raised about $2.4 billion through June 2026 and another $350 million in August. The Nvidia licensing deal, reported at around $20 billion, was paid out to investors, and the remaining company was valued at $3.5 billion, down from $6.9 billion in September 2025. Groq says that is a new valuation for a different company, not a down round. The deal is contested: the Justice Department is probing it, the New York Times reported in September 2026, and two former engineers sued Groq’s board in October, saying shareholders were shortchanged. Either way, the money built technology that Nvidia now also uses.

SambaNova has raised more than $2.4 billion, before any second close of its Series F. That funding produced several chip generations and serious customers, but little disclosed financial output. Its CEO said in July 2026 that an IPO, likely in the U.S., is under strong consideration, which would finally force disclosure.

Tenstorrent announced a $1.5 billion Series E at a $6.2 billion valuation in October 2026. A Form D filed in September showed about $1.29 billion sold from a $1.42 billion offering, so the final figure depends on later closings. It is now among the best-funded private players, without disclosed revenue.

Taalas stands out for frugality: a team of 24 spent around $30 million on its first product, the company says, before AMD agreed to buy it.

Etched is the outlier. Its signed contracts exceed $1 billion, and it says it has raised more than $1.9 billion, while the rounds it has disclosed add up to about $1.8 billion. Contracts mean something only after Etched delivers racks, collects cash and earns acceptable margins. FuriosaAI reached volume production on $246 million, but its 800 billion won Series D has stalled: a 400 billion won state pledge approved in May 2026 was still unpaid in late September because private investors had not matched it, ChosunBiz reported. d-Matrix reached production on $450 million. Both may prove more capital-efficient, but private revenue keeps us from saying so with confidence.

Is Etched now the strongest private AI chip challenger?

Not yet. Etched has had the strongest run of any private company, but SambaNova still has more proven systems in customers’ hands.

Etched raised $800 million across several financings, then $300 million at a $10.3 billion valuation led by Sequoia in July 2026, with SK Hynix among the investors. In August it raised $700 million at $21 billion, led by Jane Street, with Kleiner Perkins, Sequoia, Andreessen Horowitz and Tiger Global taking part. The valuation doubled in four weeks and is now the highest among private AI chip startups.

The substance behind it is real but thin. Etched has working silicon, more than $1 billion in contracts and one delivered rack. It has not named its other customers, published independent benchmarks or disclosed shipment timing. A higher valuation reflects investor belief in the specialized approach, not proof of a lead. Unnamed customers and undisclosed shipment timing are why we separate who leads on shipments from who only promises in our report on who's ahead in AI chips.

SambaNova had a strong run too: SN50 launched, Intel deepened its ties, the company raised a $1 billion first close at $11 billion and signed JPMorganChase. Delivering SN50 at volume is its next test. Tenstorrent announced the year’s largest private round in the field.

FuriosaAI built quieter but concrete momentum through volume production, a Samsung SDS service, RNGD systems at an Equinix site in Lisbon and a May 2026 Broadcom partnership on rack-scale clusters. d-Matrix launched a demo cloud and rack-scale clusters. Rebellions is targeting a KOSPI listing in the first half of 2027, its CEO said in July 2026. Positron says it raised an $875 million Series C at a $5 billion valuation, though its site gives no date.

Groq is the company losing ground. Its cloud and developer base keep growing, but it lost its founding leadership to Nvidia, halved its valuation and now runs Nvidia hardware. It is stronger as an operator than as a chip challenger.

How reliable are the AI chip startup claims?

The evidence behind Cerebras is far stronger than the rest of the field, so its lead carries much more confidence than the lower rankings.

Its revenue, margins, cash, order book, customer concentration and losses appear in SEC filings. Major counterparties have confirmed the largest relationships. Reports on its first post-IPO quarter even differ slightly, with figures of $191.3 million and $193.4 million in circulation, but both show revenue nearly doubling. Future capacity can still be delayed, yet the current business is measurable.

Groq’s developer count, data-center footprint and megawatts come from the company. SambaNova’s funding and named customers are credible; deployment size and SN50 volumes remain unknown. Tenstorrent’s announced $1.5 billion round is larger than what its September filing showed as sold.

Etched has shown working silicon, financing, signed contracts and one delivery. Because its first customer is also its lead investor, and other customers are unnamed, we treat the orders as serious demand rather than guaranteed sales. Its total funding figure is slightly higher than the sum of the rounds it has disclosed.

Taalas’s headline benchmark comes from the company, runs one hardwired model and accepts some quality loss. Fractile’s performance and cost figures describe a chip that has not shipped.

Independent evidence is improving. Artificial Analysis measures providers such as SambaNova, and academic work has compared Cerebras, Groq and SambaNova under shared workloads while measuring power. We are highly confident about Cerebras, reasonably confident about SambaNova, Etched and Groq, and more cautious further down the list.

Which AI chip startups are actually ahead?

Cerebras is ahead overall, and the gap is too large to call the race close. Behind it, SambaNova now holds second place, Etched third and Groq fourth.

We give the most weight to paid adoption, working products, accessible performance, delivery capacity and recent growth. Cerebras leads most of those at once, backed by a $25.4 billion order book. Customer concentration, thin GAAP margins and the difficulty of building hundreds of megawatts are serious risks, but none cancels the lead visible in its results.

SambaNova ranks second. It has mature systems, a credible new chip, one of the best enterprise customers in the field and an $11 billion valuation. It is now the strongest independent private chip platform. Visible SN50 deliveries and revenue would make that lead clearer.

Etched takes third because it turned working silicon into a delivered rack and more than $1 billion in contracts. Its ranking rests on those contracts, not on its $21 billion valuation. Volume delivery will decide whether it climbs or falls.

Groq drops to fourth. Its cloud still has real operating scale and the broadest developer use, but it now runs Nvidia systems, lost its founding leadership and was revalued at half its 2025 price. Tenstorrent follows in fifth, with a broad product strategy, open software, real deployments and fresh capital.

FuriosaAI holds sixth on the strength of its 4,000-chip volume batch, complete servers and cloud access. Its delayed Series D is the main risk to that place. Rebellions and d-Matrix sit close behind.

Hailo and Taalas leave the ranking because they are no longer independent. Positron now has much more capital, Axelera AI, now the leading independent edge specialist, has a new architecture and SiMa.ai has growing drone partnerships, but all three show limited disclosed scale. Fractile finishes last because its funding has yet to become shipping silicon.

The ranking can change. Etched could take second if more racks reach more customers. SambaNova could widen its lead by showing SN50 revenue. Groq could climb back if its cloud grows faster than its rivals’ chips. Cerebras would lose ground if its buildout slips badly or customer concentration starts hurting revenue.

As of now, Cerebras has already built the commercial company that the rest of the AI chip startup field is still trying to become.

Rank	Startup	Why it is ahead
1	Cerebras	Clear leader in reported revenue, growth, mature systems, a $25.4B order book and delivery capacity
2	SambaNova	Strongest independent private enterprise platform, backed by major funding and a credible new chip
3	Etched	Working silicon, a first delivered rack and more than $1B in contracts, with volume production still to prove
4	Groq	Largest visible private inference cloud and developer base, now partly running Nvidia systems
5	Tenstorrent	Broad hardware and IP strategy, open software, real deployments and a $1.5B Series E
6	FuriosaAI	Volume production, complete systems and fast geographic expansion
7	Rebellions	Strong financing, rack-scale products, Korean telecom workloads and a planned 2027 listing
8	d-Matrix	Focused inference architecture in production, now moving to rack-scale clusters
9	Positron	Shipping first-generation systems and much larger funding, with next-generation chips due in 2027
10	Axelera AI	Well-funded edge platform with a new Europa architecture and Dell and Supermicro as partners
11	SiMa.ai	Commercial physical-AI products with attractive low-power performance, but limited disclosed scale
12	Fractile	Strong financing and an ambitious architecture, with no shipping chip yet
–	Hailo	Acquired by Microchip in September 2026; no longer ranked as an independent startup
–	Taalas	AMD agreed in August 2026 to acquire it; no longer ranked as an independent startup