Is the Edge AI Market growing now?
In our edge AI market deck, you will find everything you need to understand the market
SUMMARY
Yes. The Edge AI Market is growing now, and the strongest evidence is not a single market-size forecast but the fact that more AI compute is being pushed into phones, PCs, vehicles, cameras, robots and industrial machines at the same time.
The growth is stronger in AI penetration and semiconductor content than in total device volumes. Smartphones and PCs can ship fewer units overall while Edge AI still expands inside the devices that do get sold.
Consumer hardware is the messiest part of the story. GenAI-capable phones are taking share and NPUs are becoming standard in PCs, but neither category has produced the broad AI-driven replacement cycle manufacturers once hoped for.
Automotive looks much cleaner. Qualcomm's automotive revenue, design-win pipeline and long sequence of double-digit growth quarters show that local AI compute is becoming part of the core vehicle architecture rather than an optional feature.
Industrial Edge AI has also crossed an important line: production programs now matter more than demos. Ambarella's hundreds of deployed customer projects and long-term Hanwha agreement are much stronger evidence than prototype announcements.
Robotics is still small relative to phones or cars, but it is becoming a real standalone demand source for embedded AI chips. The reason is practical: robots cannot send every navigation, perception or control decision to a distant data center.
The software layer is finally catching up with the hardware. Apple and Google are exposing local models through normal developer frameworks, which lowers the engineering burden and makes on-device inference a product choice rather than a specialist project.
Cost savings are likely to matter most where inference is frequent, data-heavy or latency-sensitive. Video, vehicles and machines benefit more clearly from local processing than occasional complex reasoning, which still tends to belong in the cloud.
The biggest near-term constraint is memory. AI models need local memory capacity, and the current jump in DRAM and SSD costs is slowing the spread of capable Edge AI into cheaper phones and PCs just as the technology itself is becoming more efficient.
The market is therefore growing in a fairly specific way: more intelligence is moving into the device, but that does not mean the cloud is disappearing or that every AI-capable device is creating new consumer demand. Edge AI is winning first where local inference has an obvious operational advantage.
This market map, featured in our edge AI market deck, highlights top companies and startups in the edge AI market
What actually counts as Edge AI today?
Edge AI today means running AI inference on a phone, PC, vehicle, camera, robot, industrial machine or nearby edge system instead of sending every task to a distant cloud data center.
That definition sounds simple, but it changes the answer to whether the market is growing. A smartphone that runs summarization locally through its neural processor belongs in the Edge AI market. So does a car interpreting camera and radar data locally, or a factory camera detecting defects without uploading continuous video to the cloud. Nvidia Jetson systems, Qualcomm Snapdragon platforms, Ambarella's AI SoCs and Hailo accelerators are all directly exposed to this shift.
We should be much more careful with cloud AI accessed through an edge device. Using ChatGPT on a phone does not suddenly turn that workload into Edge AI if the model runs in a data center.
This is also why published Edge AI market-size estimates vary so much. Some research firms count semiconductors, others add software, servers, complete devices or even the economic value of applications built on top. We get a cleaner picture by looking at four observable things: AI-capable device shipments, local workloads, supplier revenue and production deployments.
Why is Edge AI becoming much more useful now?
Edge AI is becoming more useful now because local models can handle jobs that would have required cloud inference only a few years ago.
Apple's latest Foundation Models are a good example. The company's third-generation architecture includes multiple on-device models alongside larger models running through Private Cloud Compute. Apple's core local model remains around the 3-billion-parameter class, but it can now support much richer language features directly on Apple hardware.
Google has gone even further in distribution. In July 2026, Google said Gemini Nano was already running on more than 140 million devices. Gemini Nano 4 can be accessed by Android developers through ML Kit's Prompt API for tasks such as summarizing itineraries, extracting information from receipts and processing audio locally.
The jump is even clearer in robots and industrial machines. Nvidia's Jetson T4000 delivers up to 1,200 FP4 TFLOPS of AI compute with 64 GB of memory in a 40W to 70W power envelope. That level of embedded compute would have looked extreme for an edge device a few years ago. Boston Dynamics, NEURA Robotics, AGIBOT and other robotics companies are already integrating the Jetson Thor generation.
Edge AI has expanded well beyond traditional object detection. Local hardware can increasingly run multimodal models, language models and more autonomous software while staying inside the power limits of a phone, vehicle or machine.
As this chart shows, and as featured in our edge AI market deck, search interest in edge AI has increased sharply
Are AI smartphones actually growing now?
AI smartphones are growing now, even though the global smartphone market itself is shrinking badly.
Counterpoint currently expects GenAI-capable phones to represent about 45% of worldwide smartphone shipments in 2026, up from 36% in 2025. At the same time, it expects total smartphone shipments to fall 13.9% to roughly 1.08 billion units.
We can combine those figures. A 45% share of 1.08 billion implies roughly 486 million GenAI-capable smartphones this year. Counterpoint's figures imply about 1.25 billion total smartphone shipments in 2025, meaning the previous 36% share represented roughly 452 million GenAI phones.
So AI-phone penetration jumps by nine percentage points, while our estimate of actual AI-phone unit shipments rises by only around 8%. That is real growth, but much less spectacular than the penetration number suggests.
Counterpoint's latest semiconductor data tell a similar story from another angle. Global smartphone SoC shipments fell 15% year over year in the first half of 2026, while GenAI smartphone SoC shipments increased 24%. Manufacturers are clearly putting AI-capable chips into a much larger part of their premium and mid-high-end ranges.
Consumer behavior is the catch. Counterpoint also says GenAI is already standard in smartphones priced above roughly $400 wholesale but has yet to give buyers a strong enough reason to replace their phones early. For now, AI is changing what gets shipped faster than it is increasing the number of smartphones people buy.
| Smartphone measure | 2025 | 2026 |
|---|---|---|
| GenAI share of shipments | 36% | 45% |
| Total smartphone shipments | ~1.25B | ~1.08B |
| Implied GenAI phone shipments | ~452M | ~486M |
| Implied GenAI unit growth | ~8% |
If you want more recent data on this point, please see our latest edge AI market report.
Are AI PCs really taking off now?
AI PCs are still spreading, but the rollout has slowed enough that earlier forecasts now look too aggressive.
Gartner expected about 77.8 million AI PCs in 2025 and originally forecast 143.1 million in 2026, which would have taken AI PCs from 31% to almost 55% of worldwide PC shipments in a single year.
The memory shortage disrupted that trajectory. Gartner now expects surging DRAM and SSD prices to push the point when AI PCs reach 50% penetration out to 2028. It also estimates that PC memory will rise from around 16% of a machine's bill of materials in 2025 to 23% in 2026.
The broader PC market has already turned down. Counterpoint recorded roughly 65 million worldwide PC shipments in the second quarter of 2026, down 4% year over year and ending five consecutive quarters of growth. Gartner had already warned that the first quarter's apparent 4% growth was partly caused by vendors building inventory ahead of expected component-price increases.
Still, the underlying Edge AI transition keeps moving. PC processors increasingly include dedicated NPUs, Microsoft has made 40 TOPS of NPU performance one of the defining requirements for Copilot+ PCs, and vendors are gradually turning local AI compute into a normal part of the specification sheet.
The sharpest conclusion is that AI PCs are growing as a share of PCs faster than PCs are growing as a market. These days, the NPU transition looks increasingly inevitable, while the hoped-for AI-driven PC upgrade boom has been pushed further out.
This chart, featured in our edge AI market deck, illustrates yearly venture capital investment in edge AI startups
Is automotive the clearest Edge AI growth market?
Automotive is currently one of the strongest large-scale Edge AI markets we can actually measure.
Qualcomm's latest quarterly results make the case unusually clearly. Automotive semiconductor revenue reached $1.588 billion, up 61% from $984 million a year earlier. That was Qualcomm's 23rd consecutive quarter of double-digit year-over-year automotive growth.
The longer trend is even more useful than the latest quarter. Qualcomm's automotive design-win pipeline reached $30 billion in 2022, stood around $45 billion earlier this year and has now expanded to $65 billion. The company simultaneously raised its fiscal 2029 automotive revenue target to $10 billion.
Recent customer wins show what is entering that pipeline. BMW selected Qualcomm as its lead compute-silicon provider for next-generation digital cockpit and automated-driving programs extending through the next decade. Stellantis is also adopting Snapdragon Digital Chassis platforms across cockpit, connectivity and driver-assistance systems. Qualcomm says its technology already powers more than 75 million vehicles with edge-AI capabilities.
We should not count every Qualcomm automotive dollar as pure Edge AI because the segment includes connectivity and conventional cockpit processing. Even with that qualification, the direction is unusually clear. More vehicle functions are moving onto centralized, high-performance compute systems, and AI inference is becoming part of the core architecture.
| Qualcomm automotive measure | Earlier level | Latest level |
|---|---|---|
| Quarterly automotive revenue | $984M YoY comparison | $1.588B |
| YoY revenue growth | +61% | |
| Consecutive double-digit growth quarters | 23 | |
| Design-win pipeline | $30B in 2022 | $65B |
| FY2029 revenue target | $8B previously | $10B |
If you want more recent data on this point, please see our latest edge AI market report.
Is industrial Edge AI moving past pilot projects?
Industrial Edge AI is already moving beyond pilots in machine vision, physical security and automation.
Ambarella gives us one of the cleanest datasets because Edge AI now represents most of its business. The company's fiscal 2026 revenue reached $390.7 million, up 37.2%, while management said about 80% of revenue came from Edge AI SoCs. Ambarella also reported that its Edge AI revenue itself grew roughly 50% during the year.
More interestingly, the company says more than 370 customer AI projects are already in production across at least 200 different model architectures. That is a much stronger measure of commercialization than the number of demos or development boards sold.
The Hanwha agreement takes this another step. In May 2026, Hanwha and Ambarella signed a relationship worth more than $800 million in potential revenue over more than ten years. The agreement spans video security, robotics, industrial automation and life sciences. Ambarella said its Edge AI platform already had an installed base exceeding 46 million units when the partnership was announced.
A decade-long sourcing and co-development agreement tells us something different from a startup pilot. Hanwha is planning multiple generations of products around this compute architecture. For industrial Edge AI, that is the kind of commitment worth paying attention to.
This chart, featured in our edge AI market deck, shows how Hailo is winning in edge AI
Are robots creating real demand for Edge AI chips?
Robots are creating real Edge AI chip demand now, although robotics is still a much smaller market than smartphones or automotive.
The technical reason is straightforward. A robot navigating a warehouse, manipulating an object or avoiding a person cannot depend on a round trip to a remote data center for every decision. Cameras, sensors and control models increasingly need substantial local compute.
Commercial evidence is starting to catch up with that technical logic. Nvidia's Jetson Thor generation is already being integrated by Boston Dynamics, NEURA Robotics, AGIBOT, LG Electronics and other robotics developers. The lower-cost Jetson T4000 brings 1,200 FP4 TFLOPS into a 70W maximum power envelope, more than four times the AI compute of the previous Jetson generation according to Nvidia.
Ambarella recently disclosed more than 15 robotics design wins, including aerial drones, representing more than $100 million of expected lifetime revenue. Its robotics pipeline now contains more than 30 customers.
Those numbers are still modest next to a multibillion-dollar automotive semiconductor business. They are large enough to show that robotics is starting to become a separate source of Edge AI demand rather than just another future use case in chip-company presentations.
Can developers finally build useful AI that runs locally?
Developers can now build genuinely useful on-device AI features without creating the entire inference stack themselves.
Google's Android tooling shows how quickly the developer experience has changed. Gemini Nano can be called through ML Kit's GenAI APIs for local summarization, information extraction, audio processing and other features. Google says the model is already available across more than 140 million devices.
Google also explains the economics quite openly. A feature running locally can scale from a few users to millions without creating a matching cloud-inference bill. The same architecture works offline and keeps sensitive data on the device.
Apple has opened its on-device Foundation Models to third-party developers as well, with structured generation and tool-calling capabilities built into its software stack. Qualcomm, Nvidia, Ambarella, Hailo and newer suppliers such as Axelera are also spending heavily on SDKs, compilers and model tooling.
Hardware used to be one of the biggest obstacles to Edge AI. Software compatibility was another. The latter is improving quickly enough that developers can increasingly choose local inference because it fits the product, rather than because they have the specialized engineering team required to make it work.
This chart, featured in our edge AI market deck, illustrates yearly funding for edge AI startups
Do consumers actually care about on-device AI yet?
Consumers are not yet showing enough demand for on-device AI to create a broad phone or PC upgrade boom.
The smartphone data are especially useful here. Counterpoint expects GenAI-capable smartphones to move from 36% to 45% of shipments while simultaneously saying that AI still has not given consumers a compelling enough reason to upgrade. Total smartphone volumes are heading sharply lower even as AI penetration climbs.
PCs show something similar. Buyers increasingly receive an NPU as part of a normal laptop upgrade, but memory inflation and higher system prices are extending replacement cycles rather than shortening them. Gartner now expects the 50% AI-PC penetration milestone later than it did previously.
This does not make the hardware transition meaningless. Consumers did not explicitly demand dedicated image processors, GPS chips or 5G modems before those components became standard either.
Still, we should separate adoption from pull. Manufacturers are pushing Edge AI capability into consumer hardware faster than consumers are demanding new devices because of it. The demand case looks much stronger in vehicles, cameras and machines where local inference immediately solves a latency, bandwidth, privacy or reliability problem.
If you want more recent data on this point, please see our latest edge AI market report.
Does running AI at the edge save enough money to matter?
Edge AI can save a lot of money when an application generates frequent inference requests or huge amounts of sensor data.
Google gives a simple example in its own Android developer guidance. Once a feature runs through Gemini Nano on the user's device, usage can grow from a small audience to millions without the developer paying for a cloud-model request every time someone invokes it.
Video makes the calculation even more compelling. A security system with thousands of cameras can process frames locally and send only detections or relevant clips upstream. Uploading every high-resolution stream continuously creates bandwidth and cloud-compute costs before the AI has produced any value.
Vehicles, drones and robots add reliability to the equation. Losing cellular connectivity cannot be allowed to disable collision avoidance, navigation or basic machine control. Local inference has operational value even when cloud inference would be affordable.
There is a cost on the hardware side, of course. More capable NPUs require silicon area, memory and power. Counterpoint says smartphone memory prices jumped more than 300% year over year in the second quarter of 2026, and memory costs have now overtaken application-processor costs across smartphone segments.
Edge AI economics are strongest for workloads that happen often, need fast responses, produce lots of raw data or cannot tolerate connectivity failures. Occasional complex reasoning still makes much more sense in the cloud.
This chart, featured in our edge AI market deck, compares the main business model options for edge AI accelerator companies
How much AI work is really moving from the cloud to devices?
A meaningful share of everyday AI inference is moving toward devices, while the hardest reasoning workloads remain concentrated in the cloud.
Apple's current architecture makes this split easy to see. The company now maintains several Foundation Models for on-device use and larger server models on Private Cloud Compute. Different requests go to different compute environments depending on how much capability they need.
Google is building the same idea directly into Android. Its Hybrid Inference API lets an application prefer local execution, prefer cloud execution, require one of the two, or fall back between them. A developer can run a simple review-generation task on Gemini Nano and route more demanding requests toward a cloud Gemini model.
This approach is likely to become normal because model capability is improving on both sides at once. Cloud systems keep getting larger and better at difficult reasoning, while smaller models are becoming good enough for summarization, transcription, extraction, vision and personalization.
The useful Edge AI metric is not how quickly the cloud disappears. We should watch how much frequent inference can move closer to the user without an unacceptable drop in quality. That share is clearly rising.
Are Edge AI chip companies actually making money?
Edge AI suppliers are generating more real revenue, but the gap between established winners and speculative chip startups is already becoming obvious.
Ambarella is the clearest pure-play example. Fiscal 2026 revenue rose 37.2% to $390.7 million, with roughly 80% coming from Edge AI SoCs. The company says it has now generated about $1 billion of cumulative Edge AI revenue and has hundreds of customer programs in production.
Qualcomm operates at a different scale. Its automotive business alone generated $1.588 billion in the latest quarter, while IoT contributed another $1.83 billion. Qualcomm's IoT category is broader than Edge AI, but it includes personal AI, industrial systems, networking and robotics, so part of that growth directly reflects the move toward local AI compute.
Specialists are also starting to show customer depth. Axelera AI says it is working with more than 500 customers and raised more than $250 million earlier this year, taking total financing since its founding above $450 million. Microchip, meanwhile, agreed to acquire Hailo, which brings more than 100 current customers and a developer community exceeding 10,000 users.
The commercial pattern is getting clearer. Edge AI hardware companies with production customers, usable software and access to established distribution channels are finding real demand. An impressive benchmark chip on its own is nowhere near enough anymore.
| Company | Evidence today | What we learn |
|---|---|---|
| Ambarella | $390.7M FY revenue, +37.2%; ~80% Edge AI | Edge AI can already support meaningful semiconductor revenue |
| Qualcomm | $1.588B quarterly automotive revenue, +61% | Large embedded platforms are scaling fast |
| Axelera AI | 500+ customers; $250M+ new financing | Specialist accelerators are reaching broader deployment |
| Hailo | 100+ customers; 10,000+ developers; Microchip acquisition agreed | Strategic buyers are consolidating proven Edge AI technology |
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 revenue distribution by customer segment in the edge AI market
Is Edge AI still stuck in expensive devices?
Edge AI is moving toward cheaper hardware, but high memory costs are currently slowing that expansion.
Counterpoint says GenAI functionality has become standard in smartphones above roughly $400 wholesale. Moving further down the price ladder requires enough DRAM to store model weights and enough local compute to run them at an acceptable speed.
The current component environment makes that harder. Smartphone memory prices rose more than 300% year over year in the second quarter, while Counterpoint expects entry-level smartphone SoC shipments to decline more than 30% for the full year. The research firm also expects the memory shortage to delay wider on-device GenAI availability in mainstream phones.
PCs face the same pressure. Gartner estimates that combined DRAM and SSD prices could rise around 130% over the year, pushing average PC prices higher and delaying AI-PC adoption. It even expects the sub-$500 new-PC segment to disappear by 2028 if the economics remain this difficult.
There is a strange race happening now: small models are becoming more efficient while the memory needed to run them is becoming more expensive.
For the moment, that keeps the most capable Edge AI concentrated in premium phones, PCs, vehicles and higher-value industrial hardware. Once memory supply normalizes, the technical improvements already happening in smaller models should make much broader deployment easier.
What could slow the Edge AI market from here?
The biggest threats to Edge AI growth right now are expensive memory, weak consumer upgrade demand and local models that still cannot match cloud systems on difficult tasks.
Memory is the most immediate problem. It is already cutting total smartphone and PC shipments, pushing prices higher and making manufacturers more cautious about adding expensive local-AI specifications to cheaper products.
Utilization is the second issue. Hundreds of millions of devices can contain NPUs without users running enough valuable local AI to justify the extra silicon. The smartphone market currently shows that gap quite clearly: AI-capable hardware is spreading much faster than AI is changing replacement behavior.
Model quality is another ceiling. Small local models have improved enormously, but frontier cloud systems still win on many difficult reasoning, research and agentic workloads. Developers will keep routing those requests to data centers until local hardware closes more of the gap.
Industrial adoption adds a different kind of friction because automotive, factory and robotics programs can take years to qualify and reach volume production. A design win announced today may contribute revenue over much of the next decade.
None of these problems looks capable of stopping Edge AI outright. They do make the likely growth curve rougher than the hype suggests, especially in low-cost consumer devices.
This chart, featured in our edge AI market deck, shows how on-device AI assistant technology has evolved over time
So, is the Edge AI market growing now?
Yes. The Edge AI market is growing today, and the evidence is strong enough that we can say so without relying on optimistic market-size forecasts.
The clearest proof comes from several markets moving independently in the same direction. GenAI smartphone chips are growing while overall smartphone chip shipments fall. AI PCs continue taking a larger share of new systems despite a much weaker PC market. Qualcomm has now recorded 23 straight quarters of double-digit automotive growth and expanded its automotive design-win pipeline to $65 billion. Industrial Edge AI has reached hundreds of production programs at Ambarella, while Nvidia's latest embedded processors are being integrated into a growing set of robots and autonomous machines.
Consumer demand is the weakest part of the story. AI has not created the phone or PC upgrade cycle manufacturers hoped for, and the memory shortage is making that even harder. We would therefore avoid calling Edge AI a device-volume boom.
The bigger change is happening inside the devices that continue to ship. More compute, more model execution and more economic value are moving onto phones, cars, cameras, robots and machines. Automotive and industrial systems currently provide the cleanest commercial evidence because local inference solves problems that cloud-only architectures handle poorly.
Our final judgment is clear: the Edge AI market is genuinely growing now. The growth is strongest in compute content, AI penetration and high-value embedded deployments rather than total consumer-device volumes. That distinction explains why Edge AI suppliers can report rising deployments and revenue during a period when smartphone and PC shipments are actually falling.
The market is also becoming broader. Edge AI once meant mostly computer vision and narrow embedded inference. These days, local systems can handle language, multimodal perception and increasingly autonomous workloads. Cloud AI will continue handling the hardest requests, while more routine, private, latency-sensitive and physically interactive inference runs close to the user or machine.
That shift is already large enough to call Edge AI a growing market today, rather than a future market waiting for the technology to catch up.
If you want more recent data on this point, please see our latest edge AI market report.
OUR METHODOLOGY
This analysis tests whether the Edge AI market is growing now by breaking the question into the parts that can actually be observed: AI-capable device shipments, penetration inside existing device categories, local model capability, supplier revenue, production deployments and the amount of inference that can realistically move away from the cloud.
We use a narrow definition of Edge AI. A workload counts when AI inference runs on a phone, PC, vehicle, camera, robot, industrial machine or nearby edge system. Cloud AI accessed through an edge device does not count simply because the user happens to be holding a phone or working on a PC.
For each part of the market, we prioritized recent commercial evidence over broad market-size forecasts. Shipment data, reported revenue, production programs, design wins, installed bases, developer availability and real product integrations carry more weight here than demos, benchmarks or long-range projections.
We also compare AI-specific growth with the underlying device market when that distinction changes the interpretation. This is especially important in smartphones and PCs, where AI-capable hardware can gain share rapidly even while total unit shipments decline.
No single metric determines the conclusion. We aggregate the evidence across consumer devices, automotive, industrial systems, robotics, developer tooling and supplier economics, while keeping the weaker parts of the story in view, especially soft consumer replacement demand and unusually high memory costs.
Key sources used for this analysis include Apple Machine Learning Research on its third-generation Foundation Models, Apple Developer documentation on on-device Foundation Models, Google's Android Developers Blog on Gemini Nano and on-device inference, Google ML Kit documentation on GenAI APIs, Counterpoint Research on GenAI smartphone penetration, Counterpoint Research on smartphone SoC shipments and memory costs, Gartner on memory costs and AI-PC adoption, Qualcomm's Q3 FY2026 results, Ambarella's fiscal 2026 results, Ambarella and Hanwha's long-term Edge AI agreement, Nvidia on Jetson T4000 and robotics integrations, Axelera AI on financing and customer expansion, and Microchip on its agreement to acquire Hailo.
In our edge AI market deck, we identify pain points entrepreneurs should prioritize
Related blog posts
- The full range of business models in the edge AI market
