How do companies in edge AI make money?
In our edge AI market deck, you will find everything you need to understand the market
SUMMARY
Companies in edge AI make money through six main business models: selling chips or modules, licensing semiconductor IP, charging for development software, charging to manage deployed devices, selling custom engineering, and packaging edge AI into complete vertical applications.
The best economics do not necessarily sit closest to the AI model. IP licensors and vertical software vendors can capture recurring royalties or subscriptions, while chip companies still carry manufacturing costs and heavy R&D even when demand is strong.
For chip vendors, growth is increasingly about putting more computing value into each machine, not simply shipping more units. Qualcomm’s recent automotive growth shows that higher revenue per vehicle can matter as much as volume.
Edge AI hardware also has an unusually delayed revenue curve. A design can spend a long time in evaluation and integration, then produce years of repeat shipments once the customer’s product reaches production.
That is why startup customer counts need to be read carefully. Hundreds of evaluations, partners or prospective customers can prove market interest without yet proving that a company has secured a handful of very large, durable production programs.
IP licensing has the cleanest structural margin profile because the customer manufactures the silicon. The trade-off is that royalties can be tiny per unit, so the model becomes exceptional only when the licensed technology reaches very large shipment volumes.
Running inference locally does not eliminate cloud revenue. It shifts the recurring bill away from raw inference and toward fleet management, updates, security, orchestration, observability and coordination across thousands of deployed machines.
Vertical applications can capture far more value than infrastructure because customers pay for an operational result rather than for TOPS, tokens or device connectivity. Samsara is the clearest example in the article: the hardware gets installed once, while software keeps expanding across the same customer base.
The deepest moat in edge AI comes from several layers locking together. Replacing a vendor can mean redesigning hardware, rebuilding model toolchains, migrating fleet-management systems and changing everyday workflows at the same time.
On-device generative AI is mostly strengthening existing models rather than creating a separate pricing category. Bigger local models raise silicon content, make device-management software more important and can increase the value of the final application; there is still little evidence of a dominant per-local-token business.
This market map, featured in our edge AI market deck, highlights top companies and startups in the edge AI market
What does an edge AI company actually sell?
Edge AI companies currently make money from six very different products: chips, chip IP, development software, device-management software, engineering work and complete applications sold to the end customer.
That distinction matters because “edge AI” tells us where the AI runs, while the invoice can look completely different from one company to another. Ambarella sells physical processors that run AI inside cameras, vehicles and robots. Ceva licenses processor designs that other companies put inside their own chips. Edge Impulse sells enterprise tooling around building and deploying models. balena and ZEDEDA charge companies to manage fleets of remote machines. Samsara bundles cameras, sensors and edge inference into recurring software subscriptions for fleets and industrial operators.
Even large public-company numbers can be misleading if we ignore this. Qualcomm’s latest quarter included $1.59 billion of automotive revenue and $1.83 billion of IoT revenue, but those categories include connectivity, digital cockpits, networking and other products alongside edge AI. NVIDIA has also used a broad “edge computing” category that includes PCs, robotics, automotive and other computing outside its data-center business.
A cleaner answer comes from following what the customer actually pays for. A camera maker may buy a $20 or $50 processor. A semiconductor company may pay an upfront IP license and royalties. A factory may pay a monthly fee for every managed device. A trucking company may pay recurring software fees for every vehicle.
| Edge AI business model | What the customer pays for | Typical revenue unit | Examples |
|---|---|---|---|
| AI chips and modules | Local computing power | Chip, SoC, card or module | Ambarella, Axelera AI, DEEPX |
| Semiconductor IP | Right to build AI technology into a chip | License + royalty | Arm, Ceva |
| Edge AI development software | Building and deploying models | Enterprise software contract | Edge Impulse |
| Edge fleet management | Running remote devices and workloads | Device, node or site | balena, ZEDEDA, AWS |
| Custom engineering | Customer-specific integration | Engineering contract | Chip and platform vendors |
| Vertical applications | A business result delivered with edge AI | Asset or software subscription | Samsara |
Is edge AI finally turning into real revenue?
Edge AI is producing substantial commercial revenue now, although established semiconductor companies are much further along than most pure-play startups.
Ambarella is one of the clearest examples. Edge AI SoCs generated 80% of its $390.7 million fiscal-2026 revenue, which implies roughly $313 million from those processors. The company has now disclosed about $1 billion in cumulative edge AI revenue, more than 370 customer AI projects already in production and an installed base above 46 million edge AI chips. Its latest reported quarter added another $100.4 million of total revenue, up 16.9% year over year.
Qualcomm shows the same move at a much larger scale. Its latest quarter produced $1.59 billion in automotive revenue, up 61% year over year, and $1.83 billion in IoT revenue, up 9%. Together those businesses reached $3.42 billion for the quarter and grew 28%. Qualcomm has now recorded 23 straight quarters of double-digit year-over-year automotive growth and raised its fiscal-2029 targets to $10 billion of automotive revenue and more than $14 billion of IoT revenue.
The IP layer is moving too. Ceva’s latest results showed licensing and related revenue rising 21% to $18.2 million, its highest level in three years. It signed ten licensing agreements during the quarter, including a deal for its NeuPro-M NPU with a large global AI and computing platform company. Ceva-powered device shipments reached 567 million units during the quarter.
Those three examples sit in very different parts of the stack, but the broad point is hard to miss. Edge AI has moved well beyond a market built mainly on evaluation kits and demonstrations. The harder question now is which companies can turn deployment growth into attractive economics.
As this chart shows, and as featured in our edge AI market deck, search interest in edge AI has increased sharply
Do edge AI chip companies make more money by selling more chips or pricier chips?
For the stronger edge AI chip businesses, selling more valuable silicon inside each device is becoming at least as important as increasing unit volumes.
Qualcomm’s latest automotive numbers let us measure this unusually well. Automotive revenue rose by $604 million year over year. According to Qualcomm’s quarterly filing, $381 million of that increase came from higher revenue per unit because of better product mix and higher average selling prices. Another $223 million came from higher shipments as new vehicles launched with Snapdragon cockpit and driving products.
That means roughly 63% of the increase came from earning more per vehicle, while about 37% came from selling into more vehicles.
The nine-month numbers tell almost the same story. Qualcomm attributed $560 million of automotive growth to higher revenue per unit and $551 million to additional shipments. Over a longer period, the split is practically 50/50.
Ambarella has described the same mechanism in its own results. Its recent growth came partly from more unit shipments and partly from customers moving toward higher-priced AI inference processors. The company now expects newer products aimed at more complicated AI workloads to carry average selling prices well above its current portfolio.
A carmaker does not need to sell twice as many cars for its chip supplier to double the computing value inside each one. A basic vision processor can gradually give way to a more expensive SoC handling perception, sensor fusion, language models and other workloads.
For chip vendors, edge AI is becoming a silicon-content story as much as a unit-volume story. The best products make each physical machine worth more revenue.
If you want more recent data on this point, please see our latest edge AI market report.
Can edge AI chip startups actually reach production scale?
Edge AI chip startups are reaching real customers, but most of them still have much more proof of adoption than proof of large recurring production revenue.
Axelera AI is probably the clearest example of breadth. When the company raised more than $250 million earlier this year, it said it was already shipping to its 500th global customer across manufacturing, robotics, security, agriculture, retail, defense and other physical-AI markets. More recently, it announced another industrial deployment partnership with ASRock Industrial and released new software intended to make applications easier to build on its hardware.
Five hundred customers sounds impressive, but the number includes customers at different stages of deployment. One company buying accelerator cards for evaluation is economically very different from an OEM building the processor into tens of thousands of machines. Axelera has clearly solved the problem of getting its hardware in front of users. We still have less public evidence about the size and duration of those production programs.
DEEPX is showing the next step in that transition. Its hardware partner network in Taiwan grew from roughly 15 companies to more than 30 in about a year. AAEON said products built with DEEPX processors received initial pre-orders after certification, while Avnet Silica identified more than 30 prospective customers in areas such as machine vision, smart cities, autonomous mobile robots and factories. DEEPX is also developing robotics computing with Hyundai Motor Group’s Robotics LAB.
The pattern is fairly consistent across younger edge chip vendors. They can spread across many markets because the same low-power inference processor can go into a camera, robot, industrial PC or drone. That diversification helps them find demand, but every new hardware format also adds integration work, software support and another long path from evaluation to production.
So yes, pure-play edge AI semiconductor companies can build real businesses. As of now, the market is still waiting to see which startups can turn hundreds of customer relationships into a few very large and durable production programs.
This chart, featured in our edge AI market deck, illustrates yearly venture capital investment in edge AI startups
Why can one edge AI design win be worth years of revenue?
A large edge AI design win can lock in years of chip shipments because changing the processor after a product has been designed around it is expensive, slow and sometimes commercially unrealistic.
Ambarella gave us an unusually concrete example this year. Hanwha entered a decade-long agreement with the company covering edge AI across security, robotics, industrial automation and other businesses. Ambarella said the partnership represents more than $800 million in potential revenue.
For perspective, that potential contract value is more than twice Ambarella’s entire fiscal-2026 revenue. The money will obviously arrive over many years and depends on deployments actually materializing, but it shows why semiconductor companies obsess over design wins long before the revenue appears.
Qualcomm has built the same model on a much larger scale. Its automotive design-win pipeline has reached $65 billion. More recently, BMW selected Qualcomm as its lead compute-silicon provider for next-generation digital cockpits and automated-driving systems through the next decade. The commercial relationship follows years of engineering work, including the launch of Snapdragon Ride Pilot in BMW’s iX3.
Long development cycles explain the economics. Edge processors can require a year or more of evaluation, hardware design, software work, model optimization and validation. Automotive programs can take considerably longer. Custom engineering is often part of that process because customers need reference boards, drivers, compilers and application-specific work before they can ship.
As seen above, Ambarella already has hundreds of AI projects in production. That installed base is valuable partly because each finished design is harder to displace than an ordinary software subscription.
A successful edge AI chip company therefore has a strange revenue curve: little money during evaluation, a long wait after the design decision, and then recurring unit revenue each time the customer manufactures the product.
Is licensing edge AI IP a better business than selling chips?
Edge AI IP licensing currently produces much better gross margins than selling physical chips, provided the technology reaches enough devices to make small royalties add up.
Arm shows how powerful that model can become. In its latest quarter, Arm generated $574 million from licenses and other revenue, up 23%, plus $715 million of royalty revenue, up 22%. Its GAAP gross margin was 97.2%.
Arm’s business extends far beyond edge AI, so those numbers cannot be treated as pure edge-AI revenue. The economics are still highly relevant. A customer pays Arm while designing a processor, then Arm can collect royalties every time that processor ships. Arm never has to manufacture the physical chip.
Ceva gives us a smaller and more directly edge-focused example. Its latest quarterly revenue reached $29 million, with $18.2 million from licensing and related revenue and $10.8 million from royalties. GAAP gross margin was 87%. The company has now had its technology incorporated into more than 20 billion devices.
Compare those margins with the 58.4% GAAP gross margin Ambarella reported in its latest quarter. Ambarella still has an attractive semiconductor gross margin, but physical products carry manufacturing costs that an IP licensor largely avoids.
The catch is volume. Selling one sophisticated accelerator can bring in tens or hundreds of dollars immediately. A royalty per chip may be tiny. The IP model becomes extraordinary when the design reaches millions or billions of units.
Ceva’s latest licensing deal is interesting for exactly that reason. The customer is a large AI and computing platform company building custom silicon around Ceva’s NPU IP. One successful license can eventually open a royalty stream across an entire product family.
Licensing has the better structural economics. Chip sales can generate more dollars from each individual device, while IP companies win through extreme reuse.
If you want more recent data on this point, please see our latest edge AI market report.
This chart, featured in our edge AI market deck, shows how Hailo is winning in edge AI
Can edge AI software make SaaS money if the AI runs locally?
Edge AI software can absolutely generate recurring SaaS revenue because companies still need to build, deploy, update and manage the AI long after local inference starts running.
Edge Impulse is a good example of where the paid layer is moving. The company made its former Professional developer capabilities free, including production-ready licensing for individual developers. That looks strange if we assume the goal is to charge every person building an edge model.
The logic becomes clearer at enterprise scale. Edge Impulse now uses free developer access to grow adoption while enterprise customers pay for team workflows, security, deployment controls, support and larger production environments. Its developer community grew from roughly 160,000 at the beginning of 2025 to more than 250,000 by year-end.
Qualcomm’s acquisition of Edge Impulse also shows why development software can be worth more strategically than its standalone subscription revenue. Qualcomm now owns a development platform that can funnel hundreds of thousands of developers toward its own Dragonwing edge processors. The company has also acquired Arduino, Foundries.io, Focus AI and Augentix as it builds a broader industrial and embedded stack.
Further downstream, the pricing unit often becomes the physical device. balena’s current Production plan starts at $1,439 per month and includes 110 devices, with additional devices generally costing $2 per device per month. AWS IoT Greengrass charges $0.16 for every active core device each month. ZEDEDA sells its edge orchestration platform through an enterprise subscription and pay-as-you-go model.
Those prices say something about competition too. Basic runtime management can become extremely cheap. Companies that want to charge several dollars per device need to provide more than connectivity: remote updates, security, observability, application deployment, model management or enterprise support.
| Edge software | Current monetization unit | What keeps the revenue recurring |
|---|---|---|
| Edge Impulse | Enterprise contract | Model development, deployment and team workflows |
| AWS IoT Greengrass | Active core device | Cloud-connected edge runtime |
| balena | Managed device | Fleet management, updates and support |
| ZEDEDA | Enterprise usage | Infrastructure, applications, models and edge orchestration |
Does moving AI to the edge kill cloud revenue?
Moving AI onto devices cuts some cloud-compute spending, while creating a different recurring cloud bill around managing those devices.
The easiest example is video. A camera running computer vision locally can analyse thousands of frames without uploading every frame to a remote GPU. Bandwidth and inference costs fall dramatically. The camera can then send only an alert, a short video clip or metadata when something important happens.
AWS IoT Greengrass shows how cloud providers adapted their pricing to this architecture. Applications can continue running locally even when the device loses internet connectivity, yet AWS charges when Greengrass core devices connect to its cloud service. Customers may then pay additional AWS IoT Core, storage or data-transfer charges.
The same pattern appears in ZEDEDA’s latest product positioning. Its platform now manages infrastructure, applications, AI models and autonomous agents across remote edge nodes. The intelligence runs close to the machine, while orchestration stays centralized.
For a factory with 5,000 intelligent cameras or a retailer with thousands of stores, that management layer becomes a real SaaS business. Models change. Security patches need to be pushed. Devices fail. New versions need staged rollouts. Someone needs to know what is running on every machine.
Cloud AI monetization revolves heavily around the amount of computation consumed. Edge AI monetization tends to follow the number of deployed assets and the complexity of operating them.
Hybrid architectures therefore look commercially dominant: local compute handles latency-sensitive and data-heavy inference, while cloud software handles training, fleet management, analytics and coordination.
This chart, featured in our edge AI market deck, illustrates yearly funding for edge AI startups
Do vertical edge AI companies make better money than infrastructure vendors?
Vertical edge AI companies can capture much more revenue per customer because they charge for the business problem being solved rather than for the underlying AI compute.
Samsara is probably the best public example. Its cameras use embedded AI to detect events such as distracted driving, missing seat belts and dangerous following distances. The company then packages those capabilities into a much broader connected-operations platform sold to fleets, construction companies and other physical businesses.
Samsara’s latest quarter reached $478.8 million of revenue, up 31% year over year. Annual recurring revenue reached $1.99 billion, up 30%. Customers spending more than $100,000 annually now account for about $1.2 billion of ARR, while ARR from customers spending more than $1 million grew 62%.
Those numbers are striking beside a pure hardware vendor. Samsara has already built an annual recurring revenue base several times larger than Ambarella’s entire annual revenue, even though both companies benefit from computer vision running at the edge.
Their gross margins show the difference too. Samsara reported a 75% GAAP gross margin in its latest quarter. Ambarella reported 58.4%. Samsara still has hardware and connectivity costs, yet the customer ultimately pays for an ongoing operational application.
There is another useful clue in Samsara’s latest numbers: emerging products generated more than 20% of net new annual contract value for the second consecutive quarter. Once Samsara has a customer’s vehicles, cameras, data and workflows connected, it can sell additional applications into the same installed base.
That is one of the best business models we can find in edge AI today. The hardware creates a foothold, local AI makes the product valuable, and recurring software captures the economics over many years.
If you want more recent data on this point, please see our latest edge AI market report.
Do customers actually pay for “edge AI”?
Customers rarely budget for edge AI as an abstract technology; they pay for a cheaper, safer, faster or more autonomous physical system.
A carmaker buys a Qualcomm processor because it needs a digital cockpit, driver assistance or increasingly centralized vehicle computing. A security-camera company buys an Ambarella SoC because it wants the camera to identify people, vehicles or unusual activity without streaming everything to a server. A manufacturer using an Axelera accelerator wants computer vision to catch defects or safety violations on-site.
The willingness to pay is tied to the application.
That is also why counting “edge AI revenue” gets messy. Imagine a $500 industrial computer containing a $70 accelerator. The chip supplier sees $70 of edge AI revenue, while the system maker sells a $500 product whose differentiation may depend heavily on that accelerator. Add a software subscription on top and the same physical deployment supports several companies with completely different revenue models.
Pricing power gets stronger as the offer moves away from raw compute. Selling a processor based mainly on TOPS per watt invites direct comparison with competing processors. Selling a factory-inspection system can be priced against the cost of defects, downtime and manual inspection.
Across the market, chip vendors are adding software, software companies are adding fleet management, and vertical companies are bundling everything into an operational product.
The closer the vendor gets to a problem measured in accidents, downtime, labor hours or defective products, the less its pricing depends on the raw cost of running the model.
This chart, featured in our edge AI market deck, compares the main business model options for edge AI accelerator companies
What makes edge AI customers hard to lose?
Edge AI customers become sticky when changing vendors would force them to redo hardware, software and operating workflows at the same time.
Hardware creates the first layer. Once an OEM has designed a processor into a camera, car or robot, moving to another chip may require a new board design, different drivers, fresh thermal testing and another validation cycle.
The software toolchain adds another layer. Edge AI models need to be compiled, quantized and optimized around particular hardware. Developers then build application logic, monitoring and deployment workflows around those tools.
Qualcomm has been aggressively filling these gaps. Its recent acquisitions now include Edge Impulse for AI development, Foundries.io for device software, Arduino for the developer ecosystem, Augentix for vision processors and other businesses around industrial AI. The strategy gives Qualcomm more opportunities to stay involved from prototype to deployed product instead of competing only on the processor specification.
Fleet management makes switching harder again. A company managing thousands of devices through balena or ZEDEDA may depend on that platform for remote access, updates, security policies, application deployment and device health. Replacing the platform means migrating an operating system, processes and production infrastructure.
Vertical applications create the deepest lock-in because the edge system becomes part of everyday work. Samsara customers can use the same installed hardware and data across safety, telematics, maintenance and other applications. Once several departments depend on the system, a competitor has to replace much more than an AI camera.
Benchmark leadership alone looks weak as a long-term moat. The stronger edge AI companies are building an installed base that becomes progressively more painful to replace.
If you want more recent data on this point, please see our latest edge AI market report.
Will on-device generative AI create a new business model?
On-device generative AI is currently increasing the value of existing edge AI business models much faster than it is creating a new one.
The first effect is straightforward: bigger models require more memory, more compute and more expensive processors. Ambarella is already developing edge chips that can support models with tens of billions of parameters and has said its newer products aimed at advanced AI workloads should carry much higher average selling prices.
DEEPX recently demonstrated small language models running fully on its DX-M1 edge processor. Axelera is also pushing beyond conventional computer vision through products designed for generative AI and multi-model inference. These companies are trying to make local language and multimodal models fit inside the power and thermal limits of physical machines.
The second effect appears in software. ZEDEDA now explicitly talks about managing models and autonomous edge agents across remote infrastructure. An edge agent deployed across a factory or retail network still needs version control, security, observability and remote updates. More capable AI makes those management layers more important because the software can take actions rather than simply classify an image.
We see little evidence today that a dominant “per local token” business is emerging. Local inference is attractive partly because customers want predictable cost, privacy, low latency and less dependence on a remote compute bill.
So generative AI should mostly increase dollars per chip, software value per managed device and the price of the final application. It strengthens the models that already work.
This chart, featured in our edge AI market deck, shows revenue distribution by customer segment in the edge AI market
So which edge AI business models actually work best today?
The strongest edge AI businesses today combine recurring revenue or royalties with deep technical integration, while pure hardware remains attractive when the vendor can keep increasing the amount of silicon value inside each device.
IP licensing has the cleanest economics. Arm and Ceva can sell a design once and then collect royalties across large production volumes. Gross margins around 90% or higher are possible because customers manufacture the chips themselves.
Vertical subscription software has the strongest evidence of recurring commercial scale. Samsara is approaching $2 billion of ARR while growing around 30%, using edge AI as part of a larger product that customers keep paying for every year.
Edge AI chips are clearly working too. Qualcomm’s automotive business is now above $1.5 billion per quarter, with more of its recent growth coming from higher revenue per vehicle than from shipment volume alone. As pointed out above, Ambarella has already generated roughly $1 billion of cumulative edge AI revenue and recently secured an agreement with Hanwha carrying more than $800 million of potential revenue over a decade.
Edge-management software can also create good recurring economics because every additional deployed machine becomes another billable node. Current prices range from cents per month for basic cloud-connected runtimes to several dollars per device for more complete fleet-management platforms.
Development tools look more difficult as standalone businesses. Edge Impulse’s decision to make its former paid developer plan free is telling. Developer tooling still has considerable strategic value, but the money is increasingly attached to enterprise deployment, silicon adoption or the larger platform around the tool.
Custom engineering works best as a door into something bigger. Getting paid to help an automaker or industrial company build its first product is useful. The real prize comes when that work leads to years of chip shipments, royalties or subscriptions.
So how do companies in edge AI make money? They sell the processor, license the processor design, charge to build and operate the AI, or sell the business outcome produced by the AI. Today, the further a company can move toward recurring software or royalties without losing its technical position inside the device, the better the economics tend to become.
| Edge AI model | How the company gets paid | Our view today |
|---|---|---|
| Vertical AI + subscription | Recurring fee for the operational application | Strongest combination of growth and recurring revenue |
| IP licensing + royalties | Upfront licenses plus payments on every chip shipped | Best structural margins at scale |
| Edge AI chips | Revenue per chip, increasingly driven by higher content per device | Proven and growing, but capital and R&D intensive |
| Fleet / device software | Recurring fee per machine, node or deployment | Attractive when the platform controls more than basic connectivity |
| Development software | Enterprise tooling and production workflows | Strategically valuable, harder to monetize alone |
| Custom engineering | Project fees leading into larger deployments | Useful when it unlocks later product revenue |
If you want more recent data on this point, please see our latest edge AI market report.
OUR METHODOLOGY
This analysis asks how edge AI companies actually turn deployments into revenue. Because the category spans semiconductors, IP licensing, development tools, fleet infrastructure and complete applications, we broke the question into narrower commercial models and tested each one against observable evidence.
We prioritized recent company filings, earnings releases, disclosed pricing, production programs, customer agreements and product announcements. Revenue and growth matter, but we also looked behind them at shipment volumes, revenue per unit, gross margins, royalties, recurring revenue, design wins, production deployments, customer expansion and pricing units.
We did not treat every adoption metric as equivalent. Customer counts can show that a product is getting tested without proving production scale; a design-win pipeline can show future commercial potential without being current revenue; and a high gross margin is attractive only if the model can reach enough volume.
Broad corporate figures were kept separate from revenue explicitly identified as edge AI. Qualcomm and Arm both operate well beyond this market, so their numbers are used here when they reveal a specific economic mechanism, such as higher automotive revenue per unit or the margin structure of licensing and royalties.
We then compared the models across recurring or repeatable revenue, margin structure, demonstrated commercial scale, pricing power, depth of technical integration and the durability of the customer relationship. The final ranking is a synthesis of those factors rather than a mechanical score.
Key sources used in the analysis include Ambarella’s fiscal-2026 results, Ambarella’s fiscal-2026 Form 10-K, the Ambarella-Hanwha long-term agreement, Qualcomm’s 2026 Investor Day disclosure, Qualcomm’s financial results, Arm’s Q1 FY2027 filing, AWS IoT Greengrass pricing, AWS IoT Greengrass documentation, balena pricing, ZEDEDA Edge Inference Services, DEEPX’s DX-M1 materials, and Edge Impulse’s edge-AI material.
This chart, featured in our edge AI market deck, shows how on-device AI assistant technology has evolved over time
Related blog posts
