Semiconductor: what are the top startups now?

Why are semiconductor startups attracting so much money right now?

Semiconductor startups are getting a huge boost from AI infrastructure spending, but the industry's record growth is hiding a surprisingly uneven market.

The Semiconductor Industry Association reported $791.7 billion in global chip sales in 2025, up 25.6% from the previous year. Gartner's August 2026 forecast then projected approximately $1.56 trillion in semiconductor revenue for 2026, compared with its earlier April estimate of $1.32 trillion. That's an upward revision of roughly $235 billion in just four months.

Yet most of that increase comes from memory. Gartner expects memory revenue to jump from $220 billion in 2025 to $837 billion in 2026, while nonmemory semiconductor revenue grows from $589 billion to $718 billion. Based on those projections, memory accounts for about 83% of the industry's additional revenue.

This changes how we should look at startups. Etched and SambaNova are developing AI processors, Ayar Labs and Lightmatter are working on optical connections, while Nearfield Instruments makes equipment used to measure increasingly complex chips. These companies benefit from the same AI investment cycle, but their markets are growing at different speeds.

For investors, the opportunity goes well beyond making faster GPUs. The expensive problems around moving data, connecting chips and manufacturing them accurately are creating room for an entirely new generation of semiconductor companies.

Which semiconductor startups are actually the biggest today?

Etched and SambaNova currently dominate semiconductor startup valuations, although Ayar Labs, Lightmatter and SiFive could prove just as important over the long term.

Etched reached a $21 billion valuation following its August 2026 financing. SambaNova secured an $11 billion valuation in July. Both companies build AI inference hardware, so their enormous valuations reflect expectations that specialized processors can capture part of the market currently dominated by Nvidia.

The next group looks quite different. Lightmatter, last valued at approximately $4.4 billion, and Ayar Labs, valued at $3.75 billion, work on optical connectivity. SiFive, valued at $3.65 billion, licenses RISC-V processor technology to chip designers. Nearfield Instruments, at $1.6 billion, develops semiconductor manufacturing equipment. Optical, RISC-V and manufacturing-equipment startups serve different corners of the industry, and the market maps and grids in our report on semiconductors show how those corners fit together.

Across the following 12 companies, the reported valuation anchors total approximately $59.7 billion. Etched and SambaNova account for roughly 54% of that amount, showing how concentrated investor enthusiasm has become around AI inference.

These figures come from different funding rounds, and a few are older than others. They also represent private financing valuations rather than continuously updated market prices.

Semiconductor startup	Valuation	Main business
Etched	$21.0B	AI inference systems
SambaNova	$11.0B	Enterprise AI accelerators
Lightmatter	$4.4B	Photonic interconnects
Ayar Labs	$3.75B	Optical I/O
SiFive	$3.65B	RISC-V processor IP
Groq	$3.5B	AI inference cloud
Rebellions	~$2.34B	AI inference accelerators
Tenstorrent	~$2.0B pre-money	AI processors and RISC-V
d-Matrix	$2.0B	AI inference hardware
Nearfield Instruments	$1.6B	Semiconductor metrology
SiMa.ai	$1.45B	Edge and physical AI
Eliyan	$1.0B	Chiplet interconnects

Does a high valuation really mean a semiconductor startup is winning?

The biggest semiconductor startup valuations currently tell us more about future expectations than about proven commercial success.

Etched is the clearest example. Its valuation rose from approximately $5 billion in December 2025 to $10.3 billion in July 2026, then to $21 billion in August. That's a 4.2-fold increase in less than eight months, during which the company delivered its first customer rack. A valuation that quadrupled in under eight months is the kind of fast shift we track as a market signal every quarter in our market report about semiconductors.

SambaNova, meanwhile, has spent years developing enterprise AI hardware and recently named JPMorganChase as a customer. Yet its $11 billion valuation is barely half of Etched's.

We can put those valuations into perspective. At a hypothetical 10-times annual revenue multiple, Etched would need $2.1 billion in yearly sales to support a $21 billion valuation. Even at 20 times revenue, it would need more than $1 billion. Those are illustrative calculations, not estimates of its current business.

There's also a funding distinction worth making. A company raising $700 million does not suddenly generate $700 million in sales. That money can fund development, manufacturing and expansion, but it says relatively little about demand until customers start buying and using the products.

For our ranking, we therefore place more weight on working products, customer adoption and manufacturing readiness than on the biggest funding announcements.

Is Etched already the most promising AI chip startup?

Etched is currently the most exciting high-risk AI chip startup, with a $21 billion valuation, working hardware and its first major customer deployment.

In August 2026, Etched announced a $700 million financing led by Jane Street. The trading firm had tested its hardware, received its first rack in July and begun deploying the technology in its own data center. According to Etched and reporting by Reuters, the company also secured more than $1 billion in customer contracts from AI companies and cloud providers.

That's an impressive progression for a company that had been valued at approximately $5 billion only months earlier. Its technology focuses on making AI inference faster and less expensive through specialized chip and system architecture.

The Jane Street relationship is particularly interesting because the firm has experience building sophisticated computing infrastructure and evaluates technology against demanding workloads. Its involvement provides stronger technical validation than a funding round involving financial investors alone.

Still, Etched has only recently crossed from chip development into real customer deployment. We haven't seen evidence of large-scale manufacturing, billions in recognized annual revenue or sustained profitability.

Etched has earned its place among the leaders, but its valuation already assumes exceptional success. The next few large deployments will tell us far more about its future than another financing announcement.

Is SambaNova actually ahead of Etched in the AI chip race?

SambaNova is ahead of Etched in enterprise deployment experience, although Etched has captured much more investor excitement over a shorter period.

In July 2026, SambaNova announced the first close of a $1 billion Series F financing at an $11 billion post-money valuation. It also disclosed that JPMorganChase had selected its reconfigurable dataflow units for fast, on-premises AI inference.

SambaNova's processors are designed to run large AI models efficiently, supported by its own software and infrastructure. The company has spent several years developing this integrated approach, giving it a longer enterprise product history than Etched.

Its appeal is especially clear for banks and other regulated businesses that want tighter control over their data and computing infrastructure. Rather than relying entirely on external AI services, those customers can deploy dedicated systems internally.

Etched is chasing a potentially much larger infrastructure opportunity through highly specialized inference computing. SambaNova has already built a business around delivering complete enterprise systems.

For now, we give SambaNova the advantage in enterprise experience and Etched the advantage in recent investor backing and potential upside. Comparable audited revenue and profitability figures would be needed to say which business is actually performing better financially.

Is Groq still a semiconductor startup, or has it become something else?

Groq has increasingly become an AI inference cloud company, making it a different kind of semiconductor contender from Etched, SambaNova or Tenstorrent.

The major turning point came in December 2025, when Nvidia entered a non-exclusive technology licensing agreement involving Groq's inference technology. Nvidia subsequently incorporated related technology into its LPX platform.

Groq continued developing its inference services. It raised $650 million in June 2026 and another $350 million in August, bringing those two rounds to $1 billion. By August, the company reported operating 13 data centers and serving more than six million developers.

Its valuation history also needs careful interpretation. Groq was previously valued at approximately $6.9 billion, but its August 2026 valuation was $3.5 billion after the Nvidia agreement and related business changes. We cannot treat that difference as an ordinary decline in the value of an unchanged company.

The six-million-developer figure gives Groq a considerable distribution advantage, although developer registrations do not tell us how many users pay or how much computing they consume.

Groq remains one of the best-known independent inference platforms, but its current business is increasingly tied to operating AI infrastructure. We'd rank it alongside the leading AI computing providers while keeping that distinction in mind.

Could Tenstorrent really challenge Nvidia?

Tenstorrent is one of Nvidia's most interesting architectural challengers, especially because it combines AI processors with RISC-V technology and more open development tools.

The company raised more than $693 million in December 2024 at an approximately $2 billion pre-money valuation. Its investors backed a strategy led by Jim Keller, a chip designer known for work on major processor architectures.

Tenstorrent develops AI accelerators, processor intellectual property and supporting software. It can therefore sell technology through several channels rather than relying entirely on one flagship chip. Selling IP, chips and software through several channels is a business-model choice; we compare which models make money and keep their margins in our report covering the semiconductor market.

That flexibility comes with a challenge. Nvidia's advantage extends well beyond its GPUs. Customers also rely on CUDA, optimized libraries, networking products and the experience of developers who already understand its systems.

Tenstorrent must give customers a compelling reason to take on the work of integrating alternative hardware. Better performance on a narrow benchmark won't necessarily persuade a company to replace infrastructure that already works.

We consider Tenstorrent a serious long-term contender because of its engineering strategy and breadth of technology. However, without public revenue and shipment figures comparable to those of established chip suppliers, calling it a commercial rival to Nvidia would be premature.

Could SiFive become more important than the companies building AI chips?

SiFive could eventually become more influential than several specialized AI chip startups because its processor designs can be used inside many different products.

In April 2026, SiFive announced a $400 million Series G round at a $3.65 billion valuation. Investors included Nvidia, Apollo Global Management, Point72 Turion and T. Rowe Price Investment Management.

SiFive develops processor intellectual property based on RISC-V, an open instruction-set architecture that lets chipmakers build customized processors. Companies can license SiFive designs instead of developing every processor component from scratch.

This gives SiFive a different business model from Etched or SambaNova. It can earn money by licensing technology to companies making data-center chips, embedded processors and specialized devices.

There is an important nuance: RISC-V is an open instruction set, but commercial processor implementations still require considerable engineering. SiFive sells precisely that expertise, including performance optimization, verification and integration support.

The company also benefits from a broader move toward custom silicon. As hyperscalers and device manufacturers create processors for specific workloads, they need flexible CPU designs that fit their requirements.

Among the startups we reviewed, SiFive has one of the strongest opportunities to serve several semiconductor markets without depending on one AI accelerator becoming dominant.

Are Ayar Labs and Lightmatter building something AI companies desperately need?

Ayar Labs and Lightmatter are developing some of the most important technologies in semiconductor infrastructure today: faster, more efficient optical connections between computing components.

As AI data centers grow, processors need to exchange enormous amounts of information. Traditional electrical connections become increasingly expensive in terms of power consumption and bandwidth, especially across larger systems.

Ayar Labs is trying to move high-speed optical communication closer to the chips themselves. In March 2026, it raised $500 million at a $3.75 billion valuation, followed by another $150 million in September. Its investors and partners include Nvidia, AMD, Intel, MediaTek, Alchip and Wiwynn.

That combination is significant. Several established chip and server-industry companies have committed strategic support, giving Ayar Labs multiple potential routes into commercial computing systems.

Lightmatter is taking a competing approach through photonic interconnects and advanced packaging. At the October 2026 Open Compute Project Global Summit, it demonstrated systems including a 204.8 Tbps optical switch tray and a 102.4 Tbps accelerator tray developed through industry collaborations.

Both companies also fit into the wider move toward chiplets and advanced packaging. Eliyan is developing high-speed connections between chiplets, while Silicon Box specializes in packaging technology.

The commercial importance of this area became clearer when Marvell completed its acquisition of Celestial AI in February 2026. Marvell subsequently projected that the acquired optical interconnect technology could reach a $500 million annualized revenue run rate in fiscal 2028 and approximately $1 billion a year later. Those remain forecasts rather than realized sales.

We currently favor Ayar Labs because of its financing, established partners and progress toward high-volume production. Lightmatter remains a close competitor, particularly in system integration. Neither has yet publicly demonstrated the shipment volumes needed to settle the race.

Is Nearfield Instruments one of the most underrated semiconductor startups?

Nearfield Instruments deserves a place among the top semiconductor startups because its measurement equipment tackles problems that chipmakers cannot simply solve by designing better processors.

The Dutch company raised $380 million in June 2026 at a $1.6 billion valuation. Its technology helps manufacturers measure extremely small and complex structures during chip production.

This is becoming harder as advanced semiconductor designs introduce gate-all-around transistors, high-NA EUV lithography and increasingly sophisticated three-dimensional packaging. A tiny manufacturing defect can make an expensive chip unusable, so accurate measurements are critical to improving production yields. Gate-all-around transistors and high-NA EUV sit at different stages of readiness, and we separate what works today from what is still being tested in our report on semiconductors.

Nearfield operates in a very different competitive environment from Etched or SambaNova. An AI processor vendor needs customers to choose its architecture. A metrology supplier can potentially sell equipment to manufacturers producing many different types of chips.

That makes Nearfield particularly attractive. If several competing processor designs succeed, manufacturers could still need increasingly advanced measurement tools.

We lack the detailed customer-level revenue data to rank Nearfield as a proven commercial leader. But among semiconductor equipment startups, its technology targets a problem that's likely to become more difficult as manufacturing advances.

Which semiconductor startups are leading in edge AI?

Axelera AI and SiMa.ai are two of the most promising edge AI semiconductor startups, with Axelera showing particularly strong evidence of customer adoption.

Edge AI means running models directly on devices such as industrial cameras, robots and manufacturing equipment, rather than sending every request to a distant data center.

Axelera, based in the Netherlands, announced more than $250 million in financing in February 2026 and reported reaching its 500th global customer. Its Metis and Europa platforms are designed for efficient AI inference on local hardware.

SiMa.ai focuses on related markets including robotics, drones and automotive systems. It raised $150 million in September 2026 at a $1.45 billion valuation, bringing cumulative financing to $500 million. The company also disclosed that revenue quadrupled between 2024 and 2025, although it didn't reveal the absolute figures.

There is a useful difference between these two disclosures. Axelera has given us a concrete customer count, while SiMa.ai has provided a growth rate. We cannot establish which company generates more revenue from those numbers alone.

SiMa.ai is also developing a new hardware generation targeting approximately 1,000 dense TOPS, with availability planned for the first half of 2028. That is a future product target rather than currently available performance.

For now, Axelera looks particularly convincing on customer reach, while SiMa.ai deserves attention for its hardware-software approach to physical AI.

Is South Korea's Rebellions becoming a serious AI chip competitor?

Rebellions is emerging as one of Asia's strongest independent AI semiconductor startups, helped by substantial financing and a move toward complete inference systems.

The South Korean company raised $400 million in pre-IPO financing in March 2026, following a $250 million Series C in September 2025. Those two rounds account for approximately 76% of its reported $850 million cumulative funding.

Rebellions reached an estimated $2.34 billion valuation in March and introduced RebelRack and RebelPOD infrastructure offerings designed for larger-scale AI inference deployments.

That product direction puts Rebellions closer to SambaNova and Etched than to companies selling individual chips. Customers buying AI infrastructure also need supporting software, networking and systems that can operate reliably.

South Korea offers a useful industrial environment, with major electronics manufacturers, memory suppliers and technology companies nearby. Nevertheless, local supply-chain expertise alone won't guarantee adoption abroad.

The next test for Rebellions is whether it can build repeatable sales beyond its initial market and prove that its systems remain competitive against Nvidia-based infrastructure.

We would rank Rebellions among the most promising Asian AI processor companies today, although its international commercial reach is still developing.

Are Chinese AI chip companies already ahead of Western startups?

Several Chinese AI chip developers have stronger reported revenue than many Western private startups, although their businesses operate under very different market conditions.

Moore Threads, MetaX and Biren Technology have moved into public markets, so they no longer belong in a strict ranking of private semiconductor startups. They remain useful benchmarks for evaluating the sector. Whether listed chipmakers belong beside startups is a definition call, and we set out what counts in our market report about semiconductors.

A September 2026 report published through Shanghai's government information portal identified Moore Threads and MetaX as STAR Market-listed companies and Biren as listed in Hong Kong.

Moore Threads reported approximately RMB 1.74 billion in first-half 2026 revenue, equivalent to hundreds of millions of US dollars. That provides a far clearer measure of commercial scale than the customer contracts or funding valuations disclosed by several private Western competitors.

China's market also has unusual conditions. Restrictions on exports of advanced US technology, domestic procurement preferences and supply-chain independence goals can support Chinese chipmakers even when their products are less competitive internationally.

Strong domestic revenue therefore doesn't automatically prove that these processors could win the same customers in an unrestricted global market.

Still, overlooking China's chip companies because they've already gone public would be a mistake. Their financial disclosures show that a number of emerging chipmakers have progressed much further commercially than the private-company rankings alone would suggest.

Do semiconductor startups really need billions of dollars to compete?

Frontier AI chip startups now need enormous amounts of capital, but companies developing processor designs or semiconductor equipment can follow very different financial paths.

Recent financing demonstrates the scale involved. Etched raised $700 million in August 2026, SambaNova announced the first close of a $1 billion financing in July, Ayar Labs secured $650 million across 2026, SiFive raised $400 million in April and Nearfield Instruments raised $380 million in June.

Taken together, those five financing programs represent approximately $3.13 billion.

Building an AI accelerator involves design teams, fabrication, packaging, software development, testing and often complete server systems. A company must spend heavily before it knows whether customers will deploy its product at scale.

SiFive has a different cost structure because it licenses processor intellectual property. Nearfield Instruments needs sophisticated manufacturing and testing operations, but its customers purchase specialized equipment rather than complete computing platforms.

The capital burden also affects competition. A processor startup that takes several years to reach volume production may find itself competing with a newer generation of Nvidia or AMD products by the time its own systems are ready.

Large funding rounds help companies survive that development cycle. They also reveal just how expensive an unsuccessful attempt to enter the AI chip market can become.

Does Nvidia still leave room for smaller semiconductor startups?

Nvidia still leaves substantial room for semiconductor startups, especially companies offering specialized computing or technology that improves existing AI infrastructure.

Etched and d-Matrix are pursuing inference hardware designed to improve the economics of running AI models. SiFive offers customizable processor intellectual property, while Ayar Labs and Lightmatter target the connections between computing components.

The competitive pressure is very different across those businesses.

An inference accelerator company must persuade customers that its performance, price and software support justify moving away from Nvidia hardware. That's difficult because Nvidia's CUDA ecosystem, networking technology and developer tools are already deeply established.

A company making optical connections can potentially supply technology for systems that continue using Nvidia processors.

Groq's licensing agreement with Nvidia also shows how the relationship can evolve. A startup can develop a competing architecture and later commercialize part of its technology through the industry leader.

We see particularly strong opportunities among companies whose products help large AI systems work better, regardless of which processor manufacturer supplies the main chips.

Which semiconductor startups have actually proven customers will pay?

Etched, SambaNova, Groq, Axelera AI and SiMa.ai have all disclosed meaningful commercial activity, but their public figures measure very different things.

Etched announced more than $1 billion in customer contracts and delivered a first rack to Jane Street. SambaNova named JPMorganChase as an enterprise customer. Groq reported 13 operating data centers and more than six million developers. Axelera disclosed 500 customers, while SiMa.ai reported a fourfold increase in annual revenue.

A useful benchmark is Cerebras, which has already become a public company. Its second-quarter 2026 results reported $180.1 million in GAAP revenue, up 74% year over year. GAAP cloud and other services revenue reached $126 million, growing 281%.

Cerebras also reported a 14% GAAP gross margin. Its results demonstrate that even a semiconductor company generating substantial revenue can face difficult economics while expanding its infrastructure.

That comparison exposes the limitations of private-company announcements. A billion dollars in contracts may be recognized over several years. A customer count doesn't reveal the average contract size. Developer registrations say little about paying usage.

Here is what the main disclosed figures actually establish.

Company	Commercial evidence	What remains unclear
Cerebras	$180.1M quarterly GAAP revenue	Public company; profitability remains challenging
Etched	Over $1B in customer contracts	Recognized annual revenue
SambaNova	JPMorganChase customer selection	Revenue and contract size
Groq	13 data centers; over 6M developers	Paying usage and cloud margins
Axelera AI	500 reported customers	Average revenue per customer
SiMa.ai	Revenue grew 4x year over year	Absolute annual revenue
Rebellions	RebelRack and RebelPOD offerings	Shipment volumes and sales
Lightmatter	Optical system demonstrations	Commercial production volumes
Ayar Labs	Manufacturing and strategic partnerships	Large-scale recurring revenue

Which semiconductor technologies look like the best long-term bets?

Optical interconnects, advanced packaging, RISC-V processor designs and precision manufacturing equipment look especially attractive because they solve problems that several generations of chips will face.

Specialized AI accelerators still offer enormous potential. Etched, SambaNova, Tenstorrent and d-Matrix could build large businesses if their systems deliver sufficiently better computing economics.

However, accelerator developers face constant competition from Nvidia, AMD, hyperscaler-designed chips and one another. Their products must keep improving as models and computing requirements change. Constant pressure from the giants is one of the risks we rate in our report covering the semiconductor market.

Ayar Labs and Lightmatter have a different opportunity. The more processors data centers connect, the more important high-bandwidth, low-power communication becomes. Nearfield Instruments benefits from increasingly complicated chip manufacturing, while SiFive can license processor technology across several types of devices.

These markets have their own risks. Optical adoption may take longer than expected, large equipment suppliers can develop rival measurement technologies, and RISC-V providers still compete over performance and software compatibility.

Even with those uncertainties, we prefer semiconductor technologies that can serve multiple winning chip architectures. They offer more ways to grow without requiring one processor design to dominate the industry.

So, what are the top semiconductor startups right now?

Ayar Labs currently leads our overall semiconductor startup ranking, followed by Etched, Lightmatter, SambaNova and SiFive. Tenstorrent, Nearfield Instruments, Rebellions, d-Matrix, Axelera AI and SiMa.ai also belong among the strongest contenders.

Ayar Labs takes first place because its optical technology tackles a growing data-center bottleneck, its financing has reached $650 million during 2026, and its strategic supporters include Nvidia, AMD, Intel, MediaTek and other established industry players. It has a credible route into systems built by several major semiconductor companies, although large-scale commercial shipments remain to be demonstrated.

Etched comes second, but it is arguably the most ambitious company in the group. Its $21 billion valuation, first Jane Street deployment and more than $1 billion in customer contracts make it an unusually well-funded challenger to today's AI inference leaders. The remaining question is whether production and revenue can catch up with expectations.

Lightmatter earns third place through its photonic interconnect and packaging technology. SambaNova follows because of its enterprise inference experience and JPMorganChase customer relationship, while SiFive stands out for licensing processor designs that can be used across many chip categories.

Nearfield Instruments deserves special attention. Its manufacturing measurement technology may be less visible than an AI processor, but it addresses a problem that chipmakers face regardless of whose architecture eventually wins.

Our conclusion is that Ayar Labs offers the strongest combination of strategic relevance, industry backing and long-term opportunity today. Etched is the bigger speculative bet, while SiFive and Nearfield Instruments offer particularly interesting alternatives to the direct AI accelerator race. The companies that solve the difficult problems surrounding AI chips could ultimately become just as important as the companies designing the chips themselves.