What are the top startups in the edge AI market?

Why is ranking the top edge AI startups so hard right now?

There is no clean edge AI startup leaderboard today because the companies with the most deployed devices, the fastest-growing orders and the strongest technology are different companies.

Syntiant has already put its AI technology into tens of millions of devices. DEEPX has a much smaller installed base, but its commercial orders have accelerated sharply since mass production began. Axelera AI has built a broader customer footprint and raised enough money to fund several generations of chips. Quadric is taking another route entirely: it licenses processor IP and already generated roughly $15 million to $20 million of licensing revenue last year, according to its CEO in a TechCrunch interview.

Then there is Hailo. A few years ago, Hailo would have been an obvious candidate for number one. It raised about $340 million, reached unicorn status, built a real customer base and developed one of the better-known edge AI chip families. Lately, however, the company ran into severe financing pressure and signed a definitive agreement to be acquired by Microchip. That changes how we should rank it as an independent startup.

So we have to look at several kinds of evidence together: whether chips are actually in production, whether evaluations are becoming orders, whether customers keep deploying the technology, whether the software stack is usable, and whether the company has enough money to survive long semiconductor cycles. Looking at funding alone gives the wrong answer.

What actually counts as an edge AI startup?

For this ranking, an edge AI startup is an independent private company whose core technology lets AI models run on a device or nearby local hardware instead of sending most inference work to a remote cloud.

That includes companies such as Axelera AI and DEEPX, which sell AI accelerators; SiMa.ai, which combines specialized silicon with a software platform; Quadric and Expedera, which license NPU processor IP; and Syntiant, which focuses heavily on very low-power on-device intelligence.

We also include several power levels. An industrial camera running object detection locally, a robot processing vision-language models in real time and a battery-powered sensor listening for specific sounds can all be edge AI products. Their workloads are very different, but local inference solves the same basic problems: latency, bandwidth, privacy, connectivity and power.

We exclude startups whose real business sits in cloud or data-center inference even when they occasionally describe their products as "edge." We also exclude robotics companies that buy edge compute but do not build the underlying AI computing technology.

Acquisitions change the list too. Qualcomm already acquired Edge Impulse. Microchip has signed an agreement to acquire Hailo, although that transaction has not yet closed. Both remain important technologies, but neither belongs in a ranking of the strongest independent edge AI startups we would back going forward.

Is edge AI finally turning into a real commercial market?

Yes, edge AI is now a real commercial market, although startup revenue still lags far behind the number of demos, development kits and announced partnerships.

The clearest change is what people are trying to run locally. Edge AI used to mean relatively narrow jobs such as wake-word detection, object classification or basic machine vision. These days, vendors are pushing transformers, small language models, vision-language models and robotic perception onto local hardware. NVIDIA's Jetson Thor, for example, reaches as much as 2,070 FP4 TFLOPS and targets robots and other physical AI systems rather than simple embedded vision.

Real deployment numbers are starting to appear underneath that technical shift. Syntiant recently said one customer program alone, Moultrie's outdoor-camera lineup, has passed one million deployed cameras using its vision AI models. DEEPX has moved from years of customer evaluations into dozens of paid purchase orders. Axelera AI is shipping production Metis hardware through a growing network of industrial system partners.

The gap between "customer interest" and actual semiconductor revenue is still huge. A company can announce hundreds of evaluations while shipping very few production chips. Automotive, industrial and robotics customers can spend years qualifying a new processor. We therefore give much more weight to purchase orders, production designs and deployed products than to demonstrations at trade shows.

Edge AI has crossed the commercial threshold. What remains unsettled is how much of the value will go to specialist startups once NVIDIA, Qualcomm, NXP, Microchip and other large chipmakers push harder into the same market.

Can edge AI startups actually take business from NVIDIA and Qualcomm?

Edge AI startups can take meaningful business from NVIDIA and Qualcomm, especially in power-sensitive applications, but they need a large enough advantage to justify leaving the incumbents' much stronger ecosystems.

NVIDIA starts with an enormous software advantage. CUDA, JetPack, pretrained models, development tools, system modules and a huge robotics community all reduce the risk of choosing Jetson. NVIDIA says its edge ecosystem now includes more than two million developers, more than 7,000 customers and roughly 1,000 hardware, software and sensor partners.

Qualcomm has been filling similar gaps. Its acquisition of Edge Impulse brought in an edge AI development platform that had already attracted more than 170,000 developers. The company is also pushing its Dragonwing processors further into industrial systems and robotics.

A startup therefore needs to save customers something important. Axelera's pitch is much more inference per watt and per dollar for specific workloads. DEEPX pushes ultra-low-power NPUs into systems where a GPU would be excessive. SiMa.ai targets complex physical AI below GPU-like power envelopes. Quadric lets chip companies put programmable AI processing directly inside their own SoCs.

There is plenty of room for specialist winners. The harder market is the middle ground, where a startup is only somewhat cheaper or somewhat more efficient. Most customers will accept weaker benchmark economics if NVIDIA or Qualcomm lets them ship faster and creates less supplier risk.

What customers care about	Why NVIDIA or Qualcomm is strong	Where a startup can still win
Software	Huge existing developer ecosystems	Make model deployment unusually easy
Power	Broad platforms often use more power	Build for a narrow power envelope
Cost	Scale helps incumbent pricing	Remove compute the customer does not need
Customization	Standard platforms cover many workloads	Fit the processor closely to one device class
Supplier risk	Large vendors are easier to trust	Offer a large enough technical or economic gain to justify switching

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

Is Axelera AI the top edge AI startup right now?

Axelera AI is our number-one edge AI startup today because it has the best mix of commercial reach, production hardware, financing and a credible next-generation roadmap.

The Dutch company raised more than $250 million in its latest financing and says it has secured more than $450 million across equity, grants and venture debt since it was founded. That is a lot of capital for an edge semiconductor startup, and here it has practical value: Axelera can support Metis while funding the larger Europa architecture and eventually moving further up the inference stack.

The customer evidence is also stronger than it was a year ago. The European Innovation Council said Axelera had deployed across more than 500 customers. More recently, ASRock Industrial added Metis accelerators to its edge AI systems, following other hardware relationships with companies such as Prodrive, SECO and duagon. Metis is becoming something customers can buy inside real industrial systems rather than only as an evaluation board.

The current Metis architecture delivers up to 214 INT8 TOPS, with Axelera listing typical power around 4 to 8 watts for some configurations. Those are vendor specifications and should be read as such, but the basic positioning is attractive: serious vision and local language-model inference without jumping straight to GPU-class power consumption.

The main missing number is revenue. Five hundred customers can include evaluations, pilots and small deployments, so we still cannot tell how much high-volume production sits underneath that figure. Even so, no other independent edge AI startup currently gives us a stronger overall combination of product maturity, customer breadth and financial runway.

Is DEEPX catching Axelera AI?

DEEPX is the fastest-rising edge AI startup right now, and its recent order growth makes it the clearest challenger to Axelera AI.

DEEPX spent roughly two to three years letting more than 400 companies evaluate its technology before its DX-M1 chip entered mass production. That long setup is finally producing measurable orders. The company initially reported 27 commercial orders across eight countries within seven months of mass production. More recent Korean reporting put cumulative purchase orders at 77, with 48 arriving during only the previous four months.

That means the pace changed dramatically. The first seven months worked out to roughly four new orders a month. The next four months ran closer to 12 a month. Order count alone does not tell us the size of each deployment, but a roughly threefold acceleration after mass production is exactly the kind of pattern we want to see from an emerging chip supplier.

The dollar amount is still much smaller than DEEPX's valuation suggests. Seoul Economic Daily recently reported more than $13 million in commercial purchase orders over the first year of mass production. That is early-stage semiconductor revenue, while the company's latest financing valued it at roughly $2.2 billion. Investors are clearly paying for what they expect DEEPX to become.

The customer pipeline makes that bet more interesting. Hyundai Motor Group's Robotics LAB signed a three-year collaboration around physical AI computing for robots. AAEON also entered a three-year mass-production cooperation covering industrial computers, single-board computers and edge gateways. DEEPX says it has built relationships with 27 semiconductor distributors and more than 50 hardware and module partners.

We still put Axelera first because its overall platform and commercial footprint look broader. But DEEPX is closing the gap quickly, and another year of order growth at anything close to the recent pace could change the ranking.

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

Is SiMa.ai becoming the edge AI chip for robots and physical AI?

SiMa.ai is currently one of the strongest edge AI startups for heavier physical AI workloads, especially when robots and industrial machines need to run multimodal models locally.

SiMa.ai has raised roughly $355 million and recently added Micron as a strategic investor. Its second-generation Modalix chip is already in production, with modules, development kits and PCIe products available rather than sitting on a future roadmap.

The workload mix is what makes SiMa.ai particularly interesting. Modalix is built for CNNs, transformers, language models and vision-language models under relatively tight power limits. SiMa.ai has also worked with Advantech on a standard half-height PCIe card for industrial computers, giving customers a familiar way to add local AI acceleration without redesigning an entire machine.

Its software strategy is getting more aggressive too. SiMa.ai recently introduced Palette Neat, an agentic development environment intended to make physical AI applications easier to build and deploy. Whether developers adopt it broadly is still unknown, but SiMa.ai understands the problem correctly: a better chip will struggle if every new model requires painful custom integration.

The reason we stop at third place is commercial visibility. We can see production hardware, strategic investors and relationships with companies such as Advantech, Cisco, TRUMPF and Emerson, but SiMa.ai does not disclose the kind of customer count Axelera provides or the purchase-order progression DEEPX now provides.

We have high confidence in SiMa.ai's technical position and less confidence in its current commercial scale. That distinction is enough to keep it behind the first two for now.

Has Quadric quietly built the strongest edge AI chip-IP business?

Quadric has quietly become one of the most commercially convincing edge AI startups because it is already generating eight-figure licensing revenue without having to manufacture its own chips.

The numbers have improved quickly. Quadric's CEO told TechCrunch that licensing revenue reached roughly $15 million to $20 million last year, up from about $4 million the previous year, and that the company is targeting as much as $35 million this year. That puts Quadric in a different category from startups whose main commercial evidence is still development kits and design evaluations.

Funding has followed the revenue. Quadric first closed a $30 million Series C, then extended the round to $46 million with IFC, part of the World Bank Group, leading the second close. The company now says it has raised about $90 million in total.

Quadric licenses its Chimera general-purpose NPU architecture to companies that want AI processing inside their own chips. DENSO is an important customer because automotive design wins can eventually translate into very large unit volumes. TIER IV has also licensed Quadric technology for autonomous-driving development, while the company's customers extend into AI PCs, office equipment and edge systems.

The business model gives Quadric an advantage that is easy to overlook. A merchant semiconductor startup has to finance chip fabrication, inventory, modules and distribution. Quadric can collect licensing fees first and royalties later if customers ship silicon containing its IP.

That does not make the business easy. Customer chips take a long time to reach production, and today's licensing revenue is still small compared with established semiconductor-IP companies. But among private edge AI companies, Quadric now has unusually concrete evidence that customers will actually pay for its technology.

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

Is Syntiant really an edge AI leader if most of its revenue comes from sensors?

Syntiant is a proven edge AI company, but calling Syntiant a pure edge AI startup today badly misrepresents what the business has become.

Syntiant's AI deployment record is excellent. Its latest SEC filing says its processors and AI models have been deployed in tens of millions of devices. The company passed 20 million cumulative NDP devices and AI software models as early as 2021, and the recent Moultrie program has now exceeded one million outdoor cameras using Syntiant vision AI.

The financial picture is more surprising. Syntiant acquired Knowles' consumer MEMS microphone business, which transformed the company. Its SEC filings show total revenue jumping from $13.6 million in 2024 to $271.8 million in 2025. Most of that increase came from the acquired sensor operation.

Only about $12.3 million of 2025 revenue came from the AI segment. That was up roughly 40% from $8.8 million a year earlier, so the AI business is growing, but it represented only around 4.5% of company revenue.

Syntiant has since filed for an IPO under the proposed Nasdaq ticker SYTN. The offering has not yet been completed, which means we still treat it as private for this ranking.

We rank Syntiant highly because millions of deployed AI devices are harder evidence than almost any startup can show. We rank it below Quadric because the company's current economics are overwhelmingly driven by sensors, while Quadric is building a cleaner edge AI revenue story.

Syntiant metric	What we found
2024 total revenue	$13.6M
2025 total revenue	$271.8M
2024 AI revenue	$8.8M
2025 AI revenue	$12.3M
AI share of 2025 revenue	About 4.5%
AI deployment	Tens of millions of devices

Is EdgeCortix becoming the edge AI specialist for defense and space?

EdgeCortix currently has the strongest specialist position in defense and space among the independent edge AI chip startups we reviewed.

Its SAKURA-II accelerator recently went through a much tougher validation process than a normal industrial proof of concept. EdgeCortix integrated the chip into a U.S. Air Force mission system and flew it during a large-force exercise. The U.S. Defense Innovation Unit subsequently issued a Success Memorandum after the prototype work met its technical objectives.

Space adds another unusual layer. NASA heavy-ion testing found no destructive events in SAKURA-II and relatively few transient radiation effects, supporting its use in environments ranging from low Earth orbit toward lunar missions. The Defense Innovation Unit now lists EdgeCortix's high-performance edge compute solution in its product catalog.

The company has also built enough financial backing to pursue those markets seriously. EdgeCortix says cumulative funding has passed $110 million, alongside Japanese government support and a separate credit facility.

Defense programs move slowly, so technical validation should not be confused with large recurring revenue. Still, a processor that survives radiation testing and performs inference inside an actual airborne mission system has cleared a much higher technical bar than a trade-show demo.

That gives EdgeCortix a narrower market than Axelera or DEEPX but potentially deeper defensibility. We currently see it as the most interesting specialist edge AI startup in environments where power, connectivity and reliability are all constrained at once.

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

Is Kneron still one of the top edge AI startups?

Kneron is still a top edge AI startup, but it has slipped behind several faster-moving competitors because we can see plenty of product activity and much less fresh commercial data.

The company has raised more than $200 million over its lifetime and built one of the broader full-stack offerings in the market. Kneron makes its own NPUs, development hardware and software while targeting security cameras, vehicles, industrial systems and increasingly local generative AI.

Its latest KL1140 generation pushes further into larger on-device models. Kneron says four KL1140 chips can be linked to support models with as many as 120 billion parameters, while its KNEO hardware extends the architecture into local enterprise inference. It also says its developer platform has accumulated roughly 28,000 developers.

Kneron has stayed active lately. At COMPUTEX, it showed local AI agents, enterprise inference, security products and automotive systems, and one of its 3D facial-recognition modules won a Best Choice Award. Customers and partners named by the company over time include Toyota, Quanta and Hanwha.

The weak point is simple: we cannot see enough recent revenue, order or shipment data. Kneron's last large publicly disclosed equity round dates back several years, and the company has produced far fewer hard commercial numbers than DEEPX, Quadric or Syntiant.

That does not make Kneron weak. It means our confidence in its exact position is lower. We keep it in the top ten because the technology and product breadth are real, but we would need fresh production numbers to move it back toward the top five.

Is Expedera stronger than its funding numbers make it look?

Expedera is stronger than its modest funding total suggests because its edge AI processor IP has already reached more than ten million devices.

Expedera has raised a little over $47 million, far below Axelera, SiMa.ai or DEEPX. The comparison is misleading, though, because Expedera licenses NPU IP instead of manufacturing accelerator chips itself.

Its Origin architecture has production licenses across consumer devices, smartphones, automotive systems and other SoCs. The company says its technology has been deployed in well over ten million devices, giving Expedera much more real-world volume than many better-funded chip startups.

The product is still evolving. Origin Evolution recently won the Edge AI and Vision Product of the Year award for processor IP and adds specific support for the transformer, attention and feed-forward operations increasingly required by language and vision-language models. Expedera also supports common frameworks such as PyTorch, ONNX, TensorFlow and Hugging Face workflows.

We still rank Quadric higher because Quadric gives us actual licensing-revenue numbers and unusually strong recent growth. Expedera discloses less about revenue and individual design wins.

Even so, Expedera is a good reminder that funding tables can hide the strongest edge AI businesses. Getting processor IP into more than ten million devices with roughly $47 million of capital is a much more interesting achievement than raising hundreds of millions before reaching production.

Is MemryX big enough to stay among the top edge AI startups?

MemryX has enough production progress to remain in our top ten, although it is currently a smaller commercial company than the leaders above it.

Its MX3 accelerator has been available as production silicon since 2024, and MemryX raised $44 million in a Series B after reaching production quality. The chip is designed primarily for continuous vision inference and typically uses roughly 0.5 to 3 watts per chip, according to the company's current technical documentation.

MemryX has spent the past year building the distribution layer around that silicon. Its partner network now includes industrial hardware vendors, software companies and distributors, while a recent agreement with Saudi systems integrator Ebttikar is aimed at production-ready computer-vision deployments across Saudi Arabia and potentially the broader Middle East.

This is a useful stage of the semiconductor journey: MemryX is past tape-out and past basic product availability, but we still need much larger shipment or revenue numbers before calling it a category leader.

The company stays in our top ten because production silicon at this power level has clear applications in multi-camera systems and industrial vision. If its expanding partner network starts producing repeat volume orders, MemryX could move several places higher.

Is EnCharge AI ready for production or still mostly a technology bet?

EnCharge AI is still mostly a technology bet today, although its analog in-memory architecture gives it one of the highest technical ceilings in the edge AI market.

The company has raised more than $144 million, including a $100 million Series B led by Tiger Global. Its EN100 accelerator is designed for laptops, workstations and edge systems and uses analog in-memory computing to reduce the constant movement of model weights between memory and compute.

That idea goes directly after one of local AI's biggest problems. Large models spend huge amounts of energy moving data around, so performing more of the computation close to where the weights are stored could materially improve efficiency.

EnCharge says EN100 can deliver more than 200 TOPS within client and edge power limits. The architecture is different enough from standard digital NPUs that successful production could make EnCharge a major competitor rather than another incremental accelerator vendor.

Commercial evidence remains the problem. EnCharge has announced the product and moved toward commercialization, but we have not found shipment, revenue or production-order figures comparable with DEEPX, Quadric, Syntiant or even MemryX.

So we put EnCharge at the bottom of the current top ten. The upside is unusually large, but today we are still betting on the architecture working commercially rather than measuring a business that has already scaled.

What could knock today's edge AI leaders off the top?

The biggest risk for today's edge AI startups is that edge AI grows exactly as expected while larger semiconductor companies capture most of the market.

The competitive pressure is already visible. NVIDIA keeps increasing Jetson performance. Qualcomm bought Edge Impulse and is expanding its industrial and robotics processors. Microchip's pending Hailo acquisition would give it a mature family of edge AI accelerators, vision processors and software immediately.

Startups also have to keep redesigning around fast-changing models. An architecture built mainly for CNN vision workloads can lose relevance when customers suddenly want transformers, language models and vision-language-action models. We can see Axelera, SiMa.ai, DEEPX, Kneron, Quadric and Expedera all pushing their newer products toward these workloads.

Then there is the semiconductor cash problem. Customers can evaluate a chip for years before placing a serious production order, while the startup has to keep paying engineers and funding new silicon. Hailo showed how dangerous that timing mismatch can become even for a company with a respected product and real customers.

Funding therefore helps, but we care more about what happens after the money arrives. Axelera needs its broad customer base to turn into large production programs. DEEPX needs its rapidly rising order count to become much larger revenue. SiMa.ai needs more visible commercial scale. EdgeCortix needs defense validation to turn into repeat procurement. EnCharge needs to show that analog in-memory computing works reliably in products customers will buy.

Startup risk	Companies most exposed	What we would watch
Long production cycles	Most chip sellers	Evaluations converting into repeat orders
NVIDIA and Qualcomm improve faster	Axelera, DEEPX, SiMa.ai, MemryX	Performance and cost advantage on real workloads
Model architectures change	Every accelerator startup	Support for transformers, LLMs, VLMs and physical AI
Cash runs out before scale	Capital-intensive chip companies	Revenue growth versus new financing needs
Specialist market stays small	EdgeCortix, ultra-low-power specialists	Number of production programs rather than demos

So, what are the top startups in the edge AI market right now?

Axelera AI is the top edge AI startup in our ranking today, with DEEPX and SiMa.ai forming the strongest group immediately behind it.

Axelera wins on balance. It already has production products, broad commercial reach, strong industrial partnerships, a deep funding base and a roadmap that can carry the company from compact edge inference into much heavier workloads.

DEEPX is the company moving upward fastest. Its purchase-order pace has accelerated sharply since mass production, and the Hyundai and AAEON relationships give it credible paths into robots and industrial systems. We would not be surprised to see DEEPX take first place if that order growth turns into much larger revenue.

SiMa.ai takes third because physical AI is moving toward exactly the type of workloads its Modalix platform targets: multimodal models, local language models, robotics and industrial reasoning under tight power constraints. Commercial disclosure is the main thing holding it back.

Quadric moves into fourth place in our updated ranking. Roughly $15 million to $20 million of licensing revenue last year, rapid growth and another financing close give us harder commercial evidence than we had before. Syntiant follows in fifth. Its deployment record remains exceptional, but its SEC filing shows that the company investors may soon buy through an IPO is economically dominated by sensors rather than AI.

EdgeCortix is sixth because its recent U.S. Air Force, DIU and NASA work gives it an unusually credible niche in defense and space. Kneron remains seventh: broad technology, a real ecosystem, but insufficient fresh commercial disclosure to rank higher. Expedera comes eighth because its processor IP has already reached millions of devices on surprisingly little capital.

MemryX takes ninth on the strength of production silicon and an expanding commercial ecosystem. EnCharge AI rounds out the top ten. We think its technology could ultimately prove more important than its current position suggests, but production evidence has to catch up with the technical promise.

Innatera sits just outside the ranking. Its neuromorphic Pulsar processor is worth watching for extremely power-constrained sensor intelligence, a part of edge AI where the winners may look very different from the companies building accelerators for robots and industrial computers.

The edge AI market still has no independent startup that can match NVIDIA or Qualcomm across hardware, software and distribution. We do, however, see a much clearer first tier than even a year ago. Axelera AI currently has the strongest all-around position, DEEPX has the most aggressive commercial trajectory, and SiMa.ai looks particularly well placed if physical AI becomes the next major source of edge inference demand.

Rank	Startup	Our current read
1	Axelera AI	Best overall mix of production, customers, funding and roadmap
2	DEEPX	Fastest recent commercial acceleration
3	SiMa.ai	Strong position in heavier physical AI workloads
4	Quadric	Most convincing recent edge AI processor-IP revenue story
5	Syntiant	Exceptional deployment history, but AI is now a small part of total revenue
6	EdgeCortix	Strongest specialist position in defense and space
7	Kneron	Broad platform with weaker fresh commercial disclosure
8	Expedera	Large deployed IP footprint relative to capital raised
9	MemryX	Production silicon with growing industrial distribution
10	EnCharge AI	High-upside architecture that still needs production proof

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

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

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