As of May 19, 2026, the investment thesis for AI moving from the cloud to the edge has shifted from "competing on model parameters" toward competing on "entry points, distribution, hardware value content, power/privacy, and the ability to charge on an ongoing basis." Microsoft defines Copilot+ PC as a new generation of Windows PC built around a 40+ TOPS NPU. Apple anchors Apple Intelligence to hardware running A17 Pro / M1 and above plus the Private Cloud Compute privacy architecture, and opens local model calls to developers. Google is replacing Google Assistant with Gemini, while Amazon has turned Alexa+ into one of the few edge AI assistants that genuinely charges consumers a standalone subscription. At the same time, Meta and EssilorLuxottica have, for the first time, turned smart glasses into AI wearable hardware with proven sales.
Core Conclusions
On-device AI's place in the supply chain has been upgraded from "feature enhancement" to "the right to define the device." The real profit pool sits not just in the model, but in the operating system entry point, the default assistant, the chip platform, the developer API, and the multi-device account system. Microsoft, Apple, Google, Meta, and Amazon are all turning AI into part of the OS / default entry point.
The on-device AI scenarios that already generate real revenue are mostly not "charging separately for AI features," but rather "premium hardware shipments + SoC/NPU upgrades + a small amount of subscription revenue." The clearest revenue validation comes from chip upgrades and ASP defense in premium AI phones and AI PCs, Ray-Ban Meta smart glasses hardware sales, and subscriptions such as Alexa+ and Google AI Pro/Ultra.
The first AI consumer electronics to commercialize, and the ones with the highest level of validation, are not XR headsets or the AI Pin, but three lines: "AI phones, AI PCs, and AI glasses." Of these, the platform-driven installed base of AI phones and AI PCs is the most certain, while the incremental upside of AI glasses is the largest.
Many scenarios remain at the feature-demo, marketing-concept, or free-built-in stage: phone-side writing/photo editing/translation/summarization, most AI earbud and AI watch features, the vast majority of cross-app agents, personal memory / notification management, and the vast majority of "AI-native companion devices." The shutdown and sale of the Humane AI Pin is the most direct counterexample.
The commercialization of AI PCs in 2025-2026 is more a story of "a Windows 10 replacement cycle + NPU as an added selling point" than evidence of a strong willingness to pay separately at the device level. Canalys expects AI PC shipments to exceed 100 million units in 2025, roughly 40% of PC shipments; but channel surveys also show that enterprise customers currently care more about battery life, the Windows 11 migration, and security, and the share that genuinely lists AI capability as the top purchasing reason is not high.
AI phones look more like an "ASP protector for premium handsets" than a proven engine for a new mass replacement cycle. Most of the capabilities of Apple Intelligence, Galaxy AI, and Gemini on Android are still free built-ins; in the near term their contribution to revenue leans toward premium-handset retention, brand premium, and ecosystem lock-in, rather than direct subscriptions.
Smart glasses are the AI-native hardware category most deserving of attention right now. EssilorLuxottica has disclosed that Ray-Ban Meta has sold a cumulative 2 million units since launch; media outlets later cited internal Meta information putting 2024 sales above 1 million units, and cited Counterpoint putting Meta's smart-glasses market share at roughly 73% in the first half of 2025. This shows that "voice + camera + open-ear audio + fashion channels" has found product-market fit.
XR / spatial computing remains stuck between "product launch—user trial" and "limited shipments," far from profitability at scale. IDC expects AR/VR headset shipments to grow 41.4% in 2025, but Vision Pro's high price and thin content library have kept its sales well short of mass-market levels; Meta Quest's premium line is under similar pressure.
The profit pool is more likely to stay long-term with platforms and chips rather than ordinary OEMs. The reason is that AI features are being made free by the operating system: Apple opens local models to developers as a system capability; Microsoft is deeply embedding Windows AI Foundry, MCP, and Copilot into Windows; Google is making Gemini the new entry point for Android / Search / Chrome. Standalone apps and ordinary hardware brands are more easily squeezed.
"TOPS" is a necessary but insufficient metric. Microsoft set the Copilot+ PC threshold at 40+ TOPS, but the real experience is also constrained by LPDDR bandwidth, memory capacity, runtime, model compression, quantization, and scheduling. Academic research likewise shows that the bottleneck for local LLMs on LPDDR-class devices often lies in memory management and random access, not just peak compute.
Over the next 12-24 months, the clearest "shovel sellers" remain the SoC/NPU, memory, optics/acoustics, and foundry chains. The reason is straightforward: whether or not AI features are monetized, for a device to support local inference it must first add NPU, LPDDR/UFS, cameras, microphones, thermal headroom, and battery capacity. Tight memory supply has already begun, in 2026, to eat into the profit structure of PCs and phones.
On valuation, large-cap platform stocks have mostly already priced in AI expectations; where an expectations gap may genuinely exist is in "edge-side shovel sellers + the glasses supply chain + China/Taiwan midstream hardware capability" — for example Qualcomm, EssilorLuxottica, parts of the optics/acoustics/ODM/EMS space, and low-power edge AI chip companies. By contrast, platform stocks such as Apple, Microsoft, Meta, and Alphabet are more "high-quality assets" than "low-expectation assets."
The biggest catalysts: PC replacements driven by the end of Windows 10 support, the agentic upgrade of Android/Galaxy/Apple on premium handsets, Meta/Oakley/Ray-Ban smart glasses expanding into new categories, and whether Amazon/Google can genuinely scale consumer AI subscriptions. The biggest risks: the commoditization of AI features for free, consumers unwilling to pay a premium for AI, rising memory and battery costs, and tightening regulation of camera glasses / biometrics.
Supply-Chain Landscape and Stage of Commercialization
A Panoramic Map of On-Device AI and AI Consumer Electronics
Supply-Chain Position Sub-Segment Core Products/Services AI Demand Driver Revenue Model Commercialization Stage Representative Companies Benefit Intensity Investment Upside Chip / IP NPU/IP CPU/GPU/NPU IP, instruction sets, compiler tools Local inference, low power, heterogeneous scheduling IP licensing, royalties Already at scale Arm, Apple, Qualcomm, Intel, AMD 5 4 Chip Phone SoC Snapdragon 8 Elite, Dimensity 9400, A-series/Kirin AI phones, imaging, voice, agents Chip ASP, platform share Already at scale Qualcomm, MediaTek, Apple, Samsung, Huawei 5 4 Chip PC SoC Snapdragon X, Core Ultra, Ryzen AI Copilot+, local LLM, enterprise refresh Chip ASP, share gains At scale but early Qualcomm, Intel, AMD, Apple 5 5 Storage / Memory LPDDR/UFS/NAND/DRAM Higher memory capacity and bandwidth Local models, caching, imaging AI Higher per-device content Already at scale Micron, Samsung, SK hynix, Biwin, Longsys, Phison 4 5 Sensors / Acoustics Cameras/Microphones/Sensors Multi-camera, mic arrays, IMU, ambient sensing Multimodal, visual understanding, audio AI Volume and price both rising Already at scale Sony, Sunny Optical, Will Semiconductor, SmartSens, Goertek, Cirrus 4 4 Display / Optics OLED, waveguides, lenses, light engines AI glasses, AR/XR Lightweight display and see-through experience Module ASP Glasses/XR early stage EssilorLuxottica, LGD, Sony, Himax, Crystal Optech 4 5 Thermal / Battery Battery, graphite, VC, thermal management Long battery life and sustained inference Battery capacity / thermal upgrades Higher per-device value content Already at scale ATL, Sunwoda, Desay, Lingyi 3 4 Operating System AI OS Apple Intelligence, Windows AI, Gemini, HarmonyOS Default entry point, cross-app orchestration Ecosystem lock-in, device premium, service pull-through Launched, early monetization Apple, Microsoft, Google, Huawei, Samsung 5 5 Personal Assistant Assistant/Agent Siri, Copilot, Gemini, Meta AI, Alexa+ Voice, search, action execution Subscription, advertising, Prime/cloud bundling Early to growth stage Apple, MSFT, GOOGL, META, AMZN 5 5 Devices AI PC Copilot+ PC, Mac AI PC Office, meetings, creation, security Whole-device ASP, enterprise procurement Early growth Lenovo, HP, Dell, ASUS, Acer, Apple 4 4 Devices AI Phone S25, Pixel, iPhone 16/17, Huawei/Honor/Xiaomi Imaging, translation, writing, agents ASP, defending against churn At scale but weak monetization Apple, Samsung, Google, Xiaomi, Huawei, Honor, OPPO, vivo 5 4 Devices AI Glasses Ray-Ban Meta, Oakley Meta, Xreal/Rokid Hands-free entry, capture, translation, navigation Hardware sales, future services Shipment-validated Meta, EssilorLuxottica, Xreal, Rokid 5 5 Devices AI Earbuds Translation, transcription, assistant entry Always-on voice, audio DSP Accessory ASP, ecosystem defense Early stage Apple, Samsung, Sony, Bose, Nothing 2 3 Devices AI Watch / Health Wearables Smartwatches, health AI Sensor fusion, health coach Hardware + membership Mature hardware, weak AI monetization Apple, Garmin, Huawei, Samsung, Xiaomi 2 2 Devices XR / Spatial Computing Vision Pro, Quest, AR glasses Immersive computing, training, industrial use Hardware + content Early stage Apple, Meta, Sony, Xreal, Magic Leap 2 4 Devices Smart Home Echo/Nest/cameras/robots Home assistant, security, linked control Hardware + subscription + commerce entry Diverging Amazon, Google, Samsung, Xiaomi, Ezviz 3 3 Runtime / Toolchain On-Device Model Runtime Core ML, ONNX Runtime, AI Hub, AICore Model deployment, compression, quantization Platform take rate / developer stickiness Early but critical Apple, Microsoft, Qualcomm, Google 4 4 App Ecosystem AI Apps / Stores Documents, code, creation, meetings, search Putting edge AI to work Subscription / take rate / tool API Early stage Adobe, Microsoft ecosystem, Google Workspace, independent developers 3 3 ODM / EMS Foundry / ODM / Assembly AI PC/phone/glasses whole devices New-product introduction, more complex structures Orders, yield, scale Already at scale Foxconn, Quanta, Pegatron, Luxshare, Foxconn Industrial Internet, BYD Electronics, Huaqin, Longcheer 4 4 The factual basis behind this table includes: Microsoft's NPU threshold and typical configuration for Copilot+ PC; Apple Intelligence's compatible hardware and the local models opened to developers; Gemini replacing Assistant; Alexa+'s subscription price; and the sales validation of EssilorLuxottica/Meta smart glasses.
Stage of Commercialization by Device, and Distinguishing "Product Launch" from "Adoption at Scale"
Device Current Stage Real Revenue Evidence Key Problems Conclusion AI PC Transitioning from shipment validation to revenue realization Canalys expects 2025 AI PC shipments >100 million units, ~40% of PCs; enterprises already piloting/deploying Willingness to pay is weak, more tied to Windows 11 / the replacement cycle Mid-term benefit is certain; near-term it looks more like a configuration upgrade AI Phone Revenue realization, but mostly hardware revenue Premium handsets keep ramping; Apple/Samsung/Google make AI a flagship selling point Most AI features are free; software monetization is weak More of a premium-handset defense and retention tool Smart Glasses Shipment / payment validation Ray-Ban Meta cumulative 2 million units; 2024 sales >1 million Privacy and social acceptance, battery life, display cost The AI-native hardware category with the most upside right now AI Earbuds Product launch / user trial Translation, transcription, and call enhancement exist, but lack standalone subscription evidence Phones/OS can easily build in the same capabilities An important entry point, but a weak monetization model AI Watch / Health Wearables Mature hardware + AI add-on features Hardware sales are large, but AI monetization is weak; overall smartwatch demand is diverging Medical regulatory boundaries, insufficient innovation More of a defensive category, not the largest source of edge-AI upside XR / Spatial Computing Launch—trial—limited shipments High-priced headsets sell weakly; IDC is bullish long-term but unprofitable near-term Content, price, weight Many long-dated optionalities, poor near-term returns Smart Home Revenue realization, AI subscriptions just starting Alexa+ is one of the few clear consumer AI subscriptions Fragmented home devices, legacy-device compatibility The key is whether the assistant can move from "free control" to "ongoing ARPU" Business Models, Profit Pools, and Scenario Analysis
How On-Device AI Products Actually Make Money
The most realistic ways on-device AI makes money today are fivefold. First is the hardware premium: flagship phones, Copilot+ PCs, and smart glasses use AI as a selling point to defend ASP. Second is the chip/NPU upgrade: Qualcomm, Intel, AMD, MediaTek, and Apple capture per-device value content through higher-tier SoCs. Third is the platform subscription: Alexa+ charges explicitly, and Google charges in tiers via AI Pro/Ultra. Fourth is enterprise seat licensing: the genuine, ongoing monetization of AI as a productivity tool currently happens more in enterprise software like Microsoft 365 Copilot than in consumer-facing edge OS. Fifth is ecosystem control: through the default assistant, App Store, developer API, and account system, vendors lock in multi-device usage and the right to monetize in the future.
Looking at strengths and weaknesses: the biggest problem with the hardware premium is that it is easily eroded by price competition and free features; the SoC upgrade has the highest certainty, because as long as devices keep racing to do local inference, chip value content materializes first; the subscription has the most upside, but its prerequisite is that users treat the default assistant as a high-frequency entry point; and ecosystem control has the strongest moat, but tends to concentrate in the hands of platform companies like Apple, Microsoft, Google, Meta, and Amazon.
Where the Profit Pool Will Stay
The near-term profit pool sits mainly in three layers. The first layer is the platform companies: Windows, iOS/macOS, Android/Gemini, Alexa, and Meta AI can decide the default entry point and the system API. The second layer is the shovel sellers: SoC/NPU, LPDDR/UFS, optics, acoustics, ODM/EMS. The third layer is the device brands: they can generally win ASP defense, but apart from a few brands, find it hard to turn AI into high-margin software revenue over the long run. Meta is one of the few exceptions in smart glasses, because it simultaneously controls the model, the platform, distribution, and the hardware definition.
In other words, Apple / Microsoft / Google / Meta / Amazon / Huawei look more like "entry-point rentiers"; Qualcomm / MediaTek / Intel / AMD / Apple silicon / Arm / Micron / SK hynix / TSMC look more like "shovel sellers"; and most PC OEMs and phone OEMs look more like "passive pass-throughs of the gains." This is why this round of on-device AI research cannot just count AI features, but must look at who controls default distribution and API scheduling.
Which Scenarios Already Generate Real Revenue, and Which Remain Conceptual
Scenarios that already generate real revenue include: First, the SoC/NPU installed base of flagship phones and AI PCs; once Microsoft standardized Copilot+ hardware, the NPUs of Qualcomm, Intel, and AMD became a clear BOM upgrade item. Second, Meta smart glasses, which already have clear sales and channel validation. Third, Alexa+ / Google AI Pro/Ultra subscriptions; although not strictly equivalent to "on-device subscriptions," their primary entry points are rapidly migrating to phones, PCs, and home devices. Fourth, enterprise AI PC procurement; though it has not yet formed a large-scale standalone AI budget line, piloting/deployment has begun.
Scenarios that remain conceptual, marketing-driven, or free built-ins include: First, automatic photo editing, summarization, writing, translation, and call cleanup on phones, which are mostly free. Second, AI earbuds and AI watches, which are mostly still ecosystem features rather than standalone ARPU. Third, standalone AI Pin / companion devices, which have proven fragile: Humane has failed, Rabbit has only limited sales, and they are far from establishing a sustainable business model. Fourth, XR spatial assistants and "AI OS" cross-app automatic execution, which remain early-stage in 2026.
Three Scenarios
Dimension Conservative Base Case Aggressive Core assumptions AI features keep becoming free; enterprise refresh driven mainly by Windows 10 end-of-life; consumers cautious about glasses AI PCs become standard for premium/commercial; AI phones stay concentrated in flagships; smart glasses enter mainstream wearables Agentic experiences genuinely work; glasses become a second entry point; home-assistant subscriptions scale AI PC penetration Low-to-medium Medium-to-high High AI phone penetration Medium High penetration at the high end, diffusion into mid-range High Smart glasses shipments High growth off a low base High growth Explosive growth On-device AI payment rate Low Medium Medium-to-high ASP change Phone/PC ASP edges up, glasses rely on hardware Premium-device ASP stays strong Glasses/accessories/services grow together Main beneficiaries SoC, memory, some premium OEMs Platforms, SoC/NPU, optics, acoustics, ODM Platforms, glasses chain, shovel sellers, home assistants Main risks Commoditization for free, rising memory prices, experience falling short Regulation and social acceptance Privacy regulation, platform wars, supply constraints The base case has the strongest factual support: on one hand, Canalys's shipment expectations for AI PCs are already high; on the other, Apple/Samsung/Google are all turning AI into a default flagship capability; add to this Meta smart glasses with proven sales, and the "platform + premium device + shovel seller" triangle has already formed. Whether the aggressive scenario holds depends on whether cross-app agents, voice/visual interaction, and AI subscription ARPU genuinely work.
Value Content, Cost Structure, and Segment Breakdown
Value Content and Cost Structure of AI PCs, AI Phones, and AI Glasses
Device Main Incremental Value Content Easiest-to-Upgrade Components Components Likely to Be Squeezed Key Bottlenecks AI PC NPU, LPDDR, SSD, camera/microphone, thermal, battery, system software SoC/NPU, memory, battery life, mic array Ordinary OEM assembly margin Memory cost, developer ecosystem, enterprise procurement ROI AI Phone Flagship SoC, memory, imaging, system AI, cloud collaboration SoC/NPU, camera, storage Mid-low-end ASP, single-point app monetization Large-model power draw, regional compliance, commoditization for free AI Glasses Camera, microphone, acoustics, battery, lenses/channels, low-power SoC Optics, acoustics, structure, brand channels Standalone cameras / some wearable accessories Battery life, privacy, thermal management, social acceptance In the unit economics of an AI PC, what genuinely determines the on-device AI experience is not "one more AI button" but the NPU + sufficient memory/bandwidth + the software runtime. Microsoft explicitly requires Copilot+ PCs to have a 40+ TOPS NPU; the Qualcomm X Elite delivers 45 TOPS; Intel's Lunar Lake NPU reaches up to 48 TOPS; and Apple builds local models directly into the system framework, labeled to developers as offline / no cost per request. This means value content will concentrate in the SoC, memory, and OS runtime.
The value-content upgrade in AI phones leans more toward the flagship SoC + memory + imaging. Samsung explicitly positions the Galaxy S25 as an "AI companion"; Apple gives Apple Intelligence only to A17 Pro / A18 / M1+; Google extends Gemini upward as the Android assistant and subscription system. This means edge AI first strengthens premium handsets, rather than reaching every price tier.
The value content in AI glasses concentrates more in optics/lens channels, low-power platforms, cameras, acoustics, and industrial design. The success of Meta/Ray-Ban shows that in this category, "fashion and channels" are almost as important as the SoC — which is exactly why EssilorLuxottica deserves to be studied separately.
Deep Segment Breakdown Matrix
The table below compresses the 30 sub-segments requested by the user into five dimensions — "revenue conversion, commercialization stage, moat, risk, attractiveness" — and gives conclusive judgments. The scores are this report's analytical ratings based on public information, out of 5.
Segment How Revenue Converts Current Stage Main Moat Main Risk Attractiveness AI PC Chip installed base, whole-device ASP, enterprise refresh Shipment validation Windows ecosystem, NPU, enterprise channels Insufficient demand education 4.0 Copilot+ PC Premium Windows standard premium Early growth Microsoft's right to define the standard Weak sense of value 4.2 Mac On-Device AI Premium Mac ASP, ecosystem defense Revenue realization Apple silicon + OS integration Weak incremental monetization 4.3 AI Phone Flagship ASP, defending premium share At scale OS, SoC, brand Commoditization for free 4.2 Premium AI Smartphone ASP and retention both stable At scale Brand + chip + imaging Intensifying China competition 4.4 AI Operating System Entry point, developer distribution, ecosystem tax Platform reinforcement OS and account system Regulation 5.0 Personal AI Assistant Subscription, advertising, procurement entry Early to growth stage Default entry point, data, execution capability Free substitutes 4.6 On-Device Large Models Improving local capability and privacy Toolization stage Compression/quantization/distillation Unstable experience 3.8 On-Device Model Runtime Platform stickiness, developer lock-in Critical middle layer Compilers, APIs, hardware adaptation Open-source dilution 4.3 Phone SoC and NPU Chip ASP, design wins At scale Efficiency, ISP, baseband Android cycle 4.7 PC SoC and NPU Driven by AI PC penetration At scale but early x86/ARM compatibility, NPU Ecosystem and memory cost 4.6 Arm IP Royalties and rising penetration At scale Architecture IP monopoly Partial impact from in-house CPUs 4.4 On-Device Memory and Storage Higher per-device capacity and bandwidth At scale Supply oligopoly Strong price cycle 4.1 Camera Modules Spec upgrades driven by multimodal At scale Optics/algorithms/customer certification Phone cycle 4.0 Microphones and Acoustic Components Array upgrades for glasses/earbuds/PCs At scale Tuning, power, packaging Commoditization 3.8 AI ISP and Sensor Fusion Imaging and visual understanding At scale ISP-algorithm-chip co-design Concentration at the high end 4.1 AI Smart Glasses Hardware sales, future services Shipment validation Fashion channels, low power, multimodal Privacy regulation 4.8 AR Glasses Industrial/navigation/display Early stage Optical waveguides, display, split compute Cost and weight 3.7 AI Earbuds Accessory premium, assistant entry Early stage Audio DSP, voice interaction Strong phone substitutability 3.2 AI Watch and Health Wearables Hardware and health membership Mature hardware Sensors, algorithms, insurance/regulation Insufficient innovation 3.1 XR and Spatial Computing Hardware + content + enterprise solutions Early stage Content ecosystem, display/compute High price, low frequency 3.0 AI Smart Home Subscription + commerce entry + hardware Diverging Home entry point and device base Legacy-device compatibility 3.8 AI Gaming Devices AI upscaling, NPCs, content Exploratory GPU/platform ecosystem Monetization model not yet formed 2.9 AI Hardware Foundry/ODM/EMS New-product introduction, order upgrades At scale Customers/yield/scale Customer concentration 4.0 AI Device Thermal and Battery Higher per-device value content At scale Materials, reliability Price competition 3.7 On-Device AI Security Enterprise procurement and platform necessity Early uptrend TEE/Secure Enclave/governance Hard to monetize standalone 4.0 Privacy Computing and Local Inference Platform/enterprise necessity Rising stage OS-and-hardware security co-design High development cost 4.2 AI App Stores and Developer Ecosystem Take rate and control Platform reinforcement OS, payments, distribution Regulation 4.7 AI Companion Hardware New entry-point hypothesis Very early / high churn Design and distribution Absorbed by phones 2.2 Enterprise On-Device AI Device Management Software licensing + security management Early growth Enterprise MDM/IT systems ROI needs validation 4.1 The core basis for this matrix is: the official product definitions of Microsoft/Apple/Google/Amazon/Meta/Samsung; the relevant market views of Canalys, IDC, and Counterpoint; and the public changes in financing, sales, or going-concern status at companies such as Humane, Rabbit, Hailo, Axelera, and Ambiq.
Tiering of Investment Targets, the Scoring Model, and Key Companies
Company Tiering and Investment Priority
Tier A: Core direct beneficiaries of on-device AI Apple, Microsoft, Alphabet, Meta, Qualcomm, Samsung Electronics, TSMC, Micron, EssilorLuxottica. They either control the entry point, control the core chip/foundry/channel, or already have clear sales validation.
Tier B: Clear beneficiaries, but with valuation, cycle, or competition risk Intel, AMD, Lenovo, HP, Dell, Xiaomi, SK hynix, Goertek, Luxshare, Sunny Optical. They generally benefit from the AI PC / AI phone / AI glasses chains, but are more exposed to the PC/phone cycle, memory prices, and customer concentration.
Tier C: AI mainly improves hardware competitiveness, with weak near-term financial upside Sony, Garmin, Logitech, Dolby, Sonos, LG Electronics, Nintendo, and others. AI looks more like feature completion or defense than a restructuring of the profit pool.
Tier D: Strong narrative, weak actual validation — pseudo-beneficiaries Most AI earbuds, AI companions, AI pins, standalone device brands without ecosystem support, and some component companies merely riding the "AI phone / AI PC" concept. Humane is the most typical case.
Tier E: Potentially hit by platform AI/OS built-ins Traditional single-point translation/photo-editing/summarization/notes apps, weak-brand low-end Android phones, undifferentiated PCs, some smart speakers / traditional voice assistants, and some POV cameras and peripheral brands. Apple, Samsung, Google, and Microsoft have already built large numbers of AI features into the system layer.
Scoring Model
We suggest splitting company scoring into two models.
Positive Scoring Model Direct on-device AI revenue exposure 20% Chip/ecosystem/operating-system moat 20% Shipments, customers, and scale validation 15% Supply-chain capability and cost control 15% Financial quality and margins 10% Market space and growth upside 10% Valuation reasonableness 10%
Reverse Risk Model Insufficient consumer payment and adoption 20% Risk of AI features becoming free 20% Consumer-electronics cycle and inventory risk 20% Hardware cost and memory-price risk 15% Risk of being built into the OS 15% Overvaluation 10%
Master Table and Priority Research List
Company Market Segment On-Device AI Benefit Path Key Evidence Tier Overall Judgment Apple US Platform/Phone/PC Apple Intelligence + Apple silicon + multi-device entry A17 Pro/M1+ support, local model framework, PCC A Platform winner; monetization still to be proven Microsoft US OS/PC platform Copilot+ PC standard-setting, Windows AI Foundry 40+ TOPS Copilot+ standard, MCP/Foundry A Strongest PC entry point, but consumer monetization still weak Alphabet US OS/Search/Assistant Gemini replacing Assistant, AI Pro/Ultra, Android entry Assistant giving way to Gemini, clear paid tiers A Strong assistant/search entry; device monetization a mid-term positive Meta US Glasses/Assistant AI glasses + Meta AI + developer ecosystem Ray-Ban cumulative 2 million units, sales still expanding A Glasses are the strongest incremental hardware line Amazon US Home Assistant Alexa+ subscription, Prime bundling, Echo entry Alexa+ $19.99/month, free with Prime A One of the few clear consumer AI subscription cases Qualcomm US Phone/PC SoC Flagship phone SoC + PC NPU + Edge AI toolchain X Elite 45 TOPS; PC/glasses platform positioning A An on-device AI shovel seller; an expectations gap still exists Samsung Electronics Korea Phone/Memory/Platform Galaxy AI + memory + devices S25 as an AI companion, AI features stay free A Dual-engine drive from devices and components TSMC Taiwan/ADR Foundry Core foundry for AI phones/PCs/glasses Apple/Qualcomm/AMD/Intel advanced-node beneficiary A One of the most stable underlying shovel sellers over the medium-to-long term Micron US Memory AI devices raise DRAM/NAND content Memory shortage already eating into PC/phone cost structure A High upside but a strong cycle EssilorLuxottica Europe Glasses channel/brand Ray-Ban Meta, Oakley Meta, prescription channels Cumulative 2 million smart-glasses units sold A A core channel asset for AI glasses Intel US PC SoC Core Ultra / Lunar Lake / enterprise PC 40-48 NPU TOPS, Copilot+ compatible B Clear beneficiary, but heavy share/competition pressure AMD US PC SoC Ryzen AI series, enterprise PC 50 TOPS-class NPU product line already pushed to OEMs B Big opportunity, but installed-base share needs ongoing validation Arm US IP The architecture base for on-device AI CPU/NPU Edge/Physical AI strategy strengthening B Underlying benefit is certain; valuation usually not cheap Lenovo Hong Kong AI PC OEM Enterprise and consumer AI PC shipments Channels judge AI PC will become mainstream B One of the strongest PC OEMs, but the profit pool is not in the device maker HP US AI PC / AI Asset Integration AI PC, Humane asset integration Acquired most of Humane's assets and CosmOS B Has an AI narrative, but execution remains to be seen Dell US AI PC / Enterprise Hardware AI PC + enterprise server synergy Launched local AI laptops, enterprise AI solutions B Deep enterprise customers, but the PC side is cycle-constrained Xiaomi Hong Kong Phone/Wearables/IoT Premium AI phone + smart home synergy Clear China market-share recovery B Strong ecosystem, but weaker platform moat than OS-native giants SK hynix Korea Memory Dual benefit from devices and servers AI driving market cap and supply shift B Big beneficiary, but more deeply tied to HBM/the cycle Goertek A-shares Acoustics/Assembly Glasses/wearables/earbuds chain Benefits from smart glasses and acoustics logic B Strong build-to-order capability; customer mix needs tracking Luxshare A-shares Assembly/Acoustics/Connectivity Apple-chain wearables and component upgrades Edge AI raises complexity B Manufacturing leader; alpha depends on the ramp of new categories Sunny Optical Hong Kong Optics Phone/glasses camera upgrades Multimodal driving high-spec modules B Optics is core to the glasses chain Will Semiconductor A-shares CIS Beneficiary of visual AI Device visual understanding raises sensor value B Depends on share and ASP Ezviz A-shares Smart Home Home cameras / home AI Home assistant and security AI linkage C Clear application but high platform dependence Garmin US Wearables Health/sports device AI Mature hardware; AI is more of an add-on value C High financial quality, average AI upside Sonos US Audio Home audio entry point May be squeezed by platform voice/assistants E Higher risk of being built into platforms Note: The latest complete financial figures for some China-Taiwan / Japan / Korea / Europe component companies were not verified company-by-company this time. To expand this table into a due-diligence base table "including the last three years' revenue growth, gross margin, EV/EBITDA, P/S, and FCF yield," we recommend further completing each item using exchange filings and device-level data sources. This uncertainty does not affect this report's ranking conclusions on "benefit paths / impact paths." The relevant judgments are based on all of the official product materials, market data, and public news cited in this report.
Listed Companies Most Worth Digging Into Further
Apple On the on-device AI front, Apple's core is not "the number of AI features" but the closed loop of a device threshold + a system entry point + a developer framework + Private Cloud Compute. Apple Intelligence only supports devices running A17 Pro / M1 and above, which naturally creates replacement and premium-handset lock-in; meanwhile Apple opens on-device foundation models to developers, emphasizing offline, privacy-centric, no-charge-per-request use. The biggest near-term problem is that AI features remain mostly free, and monetization shows up more in ASP and the ecosystem than in extra software revenue.
Microsoft Microsoft is the most central platform winner in AI PCs, because the right to define the Copilot+ PC lies with Windows. The 40+ TOPS NPU, system-level AI experience, Windows AI Foundry, and the MCP interface push Microsoft from "application-layer plugins" to "OS-layer orchestration." But its risk is equally clear: if users see Copilot+ as an ordinary upgrade rather than a must-have, the PC-side AI value leans more toward OEM replacements than new Microsoft monetization.
Alphabet Google's strongest weapon in on-device AI is not the Pixel but Android + Gemini + Search/Chrome distribution. Google has explicitly made Gemini replace Assistant, and built subscription tiers through AI Pro/Ultra. For investment research, the key to Alphabet is not how many AI phones it sells, but whether it can convert the device-AI entry point into search, a defense of browser share, and new subscription ARPU.
Meta Meta is the most important "underestimated hardware variable" in on-device AI for 2025-2026. Ray-Ban Meta has already reached cumulative sales of 2 million units, showing that smart glasses are, for the first time, more than just a demo. More importantly, Meta simultaneously owns the model, the social graph, content distribution, and a glasses-hardware roadmap. The weakness is that Reality Labs is still posting large losses, and glasses regulation and privacy controversy may amplify with scale.
Amazon Amazon's on-device AI logic rests on Alexa+ being one of the few edge AI assistants that genuinely "charges consumers." Free for Prime members, $19.99 per month for non-Prime — meaning Amazon is turning home devices into the landing container for an AI subscription. If Alexa+ can genuinely raise the usage frequency of shopping, content, home control, and household management, the market will re-rate Amazon's profit potential at the home entry point.
Qualcomm Qualcomm is the most typical "on-device AI shovel seller." On phones, it leads in flagship Android SoCs and NPUs; on PCs, Snapdragon X slotted directly into the first wave of Copilot+ PCs; in glasses and at the edge, it is filling out its developer and embedded ecosystem through acquisitions such as Foundries.io, Edge Impulse, and Arduino. Relative to platform stocks, Qualcomm's on-device AI expectations gap remains larger.
Intel Intel reclaimed a voice in the AI PC arena through Lunar Lake / Core Ultra 200V, with an NPU of up to 48 TOPS that finally meets Microsoft's Copilot+ threshold. But Intel's challenge is that it must simultaneously face Qualcomm's offensive on ARM laptops and AMD's head-on competition in high-performance x86 APUs. In other words, Intel is a beneficiary, but not the most certain beneficiary.
AMD AMD's Ryzen AI series offers strong competitiveness in AI PCs; public information shows that Ryzen AI 300 has reached 50 TOPS-class NPU capability and has been adopted by multiple OEMs. AMD's advantage is its ability to build CPU/GPU/NPU into a higher-performance heterogeneous combination, well suited to creation and performance localization; its weakness is that its enterprise PC channels and default-platform control are still weaker than Microsoft's/Intel's.
Arm Arm is not the "highest traffic" name in on-device AI, but is often one of "the most stable underlying beneficiaries." Whether in phones, PCs, glasses, or IoT edge computing, Arm keeps benefiting at the architecture and royalty level. In 2026 Arm further strengthened its Physical AI organization, showing that its bet on robotics and edge computing is deepening. The risk lies mainly in valuation and the tug-of-war with customers' in-house CPUs.
TSMC TSMC is not an on-device AI concept stock but the de facto core infrastructure of on-device AI. The advanced on-device chips of Apple, Qualcomm, AMD, and Intel can almost never do without TSMC's advanced process and packaging roadmap. As long as local inference keeps penetrating premium devices, demand for advanced nodes is unlikely to fall. The catch is that its valuation reflects the entire AI semiconductor cycle more than on-device AI alone.
Micron For Micron, on-device AI is not a story but a physical demand upgrade in memory capacity and bandwidth. Both AI PCs and AI phones push up the LPDDR/SSD configuration threshold; but near-term the company is also constrained as AI infrastructure pulls away capacity, raises costs, and squeezes device demand, making it both a beneficiary and a cycle amplifier. It belongs in the "high upside, high volatility" basket.
Samsung Electronics Samsung's uniqueness is that it is simultaneously an AI phone brand, a major memory maker, a display/component manufacturer, and one of the entry points of the Android flagship camp. The Galaxy S25 uses "AI companion" as its main selling point, while Galaxy AI's basic features stay free — meaning Samsung leans toward using AI to defend hardware competitiveness. Its advantage is vertical integration; its risk is the overlay of device competition and the memory cycle.
Lenovo Lenovo is the global PC leader, and AI PC penetration itself will bring shipment opportunities; but whether the profit pool improves meaningfully depends on whether it can sell genuinely differentiated AI devices and management solutions into the enterprise market, rather than just "ordinary PCs with an NPU." On research priority, Lenovo is worth tracking, but leans more toward beta than the strongest alpha.
Xiaomi Xiaomi's opportunity lies in the synergy of phone + IoT + wearables + China channels. It is not an underlying OS leader, but it has a large enough device installed base in China and emerging markets to make AI a unified experience across phones and home devices. The risk is that its system-layer moat is weaker than Apple's/Google's, and the AI differentiation of its premium phones must face competition from Huawei, Honor, vivo, and OPPO.
EssilorLuxottica Viewed purely through "sales validation" and "distribution moat," EssilorLuxottica is one of the companies most easily underestimated by traditional tech investors in 2026's on-device AI. It controls channels and brands such as LensCrafters, Sunglass Hut, and Ray-Ban — exactly the thing that is hardest to replicate as AI glasses move from demo to mass market. It is not a model company, yet it may be the most central consumer-channel asset for the next round of AI device adoption.
Unlisted Opportunities, the Disrupted, and Final Conclusions
Important Unlisted Companies and Primary-Market Directions
Company Field Core Thesis Latest Public Signal Watch Point OpenAI io AI-native hardware Jony Ive design, OpenAI models combined with devices OpenAI acquired io for $6.5 billion May define a new form of personal device, but timing and form remain unclear Xreal AR/Smart Glasses Consumer-grade display glasses Strong potential to bind with big-tech ecosystems The real degree of AI integration and profitability needs validation Rokid Smart Glasses China AR/AI glasses Active domestic ecosystem Shipments and channels still need validation Rabbit AI companion Small AI device r1 cumulative sales ~130,000 units Shows demand exists, but far from proving a long-term business model Limitless Personal memory/wearable Meeting / personal-memory entry point High market attention Easily hit by privacy and OS built-ins Friend / Tab / Bee, etc. AI companion Companionship and personal assistance Active concept Very high churn rate Hailo On-device AI chip PC/automotive/edge generative-AI acceleration 2024 financing at $1.2 billion valuation, 300+ customers Has upside if PC/edge AI scales Axelera AI Edge/Industrial AI chip European edge inference and industrial deployment 2026 financing of $250 million, cumulative >$450 million Strong in industrial edge, weaker on the consumer side Ambiq Ultra-low-power edge AI Wearable SoC 42 million units shipped in 2024, 40% supporting AI The wearable and health-device chain is worth tracking Kneron / SiMa.ai / Useful Sensors / Edge Impulse Runtime/chip/toolchain On-device model deployment and low power Each advancing ecosystem partnerships Watch design wins rather than concepts Traditional Devices and Apps Restructured or Disrupted by AI
The first to be disrupted are single-point feature apps. As Apple, Google, Samsung, and Microsoft build writing, translation, summarization, photo editing, and search enhancement into the OS, standalone tools will lose pricing power. Next are undifferentiated devices: low-end phones and PCs with no AI features, no brand strength, and no channel strength will increasingly be left with nothing but price competition. After that come traditional voice assistants and some smart-speaker brands, because the real value is converging toward "an assistant that executes tasks + a default platform," not the speaker hardware alone. Finally, standalone AI hardware, if it cannot form a high-frequency entry point, tends to be absorbed by the phone. Humane's outcome is already a warning.
Risk Analysis
The biggest risk for on-device AI is not that the technology cannot be realized, but that commercialization cannot be cashed in. The risk for AI PCs is that users treat them as ordinary replacements; the risk for AI phones is that the premium-handset spillover is unclear and features stay free; the risk for AI glasses is privacy and social acceptance; the risk for wearable health AI is the regulatory boundary; and the risk shared across the industry is rising memory/storage/battery costs and platform commoditization for free. In addition, the EU AI Act has advanced GPAI and high-risk rules on its original timeline, and is more sensitive to scenarios such as emotion recognition, manipulative behavior, and biometrics, which is a real constraint on always-on camera / microphone devices.
Final Conclusions
The importance of on-device AI within the AI supply chain has shifted from "a device feature upgrade" to "the next battle for the entry point." The five sub-segments most worth watching are: AI operating systems and personal-assistant entry points, AI smart glasses, phone/PC SoCs and NPUs, on-device memory/storage, and on-device runtimes/developer ecosystems.
The ten listed companies most worth digging into further: Apple, Microsoft, Alphabet, Meta, Amazon, Qualcomm, Samsung Electronics, TSMC, Micron, EssilorLuxottica. They respectively represent the platform entry point, assistant monetization, AI glasses, core chips, component supply, and the channel moat.
The ten unlisted / pre-IPO companies or directions most worth tracking: OpenAI io, Xreal, Rokid, Rabbit, Limitless, Hailo, Axelera AI, Ambiq, Kneron, and Edge Impulse / similar runtime toolchains. Among these, what truly deserves higher weight is not "whether they can build a flashy device," but who can turn AI into a default entry point, ongoing usage, and sustainable monetization.
The five points most easily misunderstood by the market: First, TOPS does not equal real experience; Second, AI phones do not mean a new replacement cycle has been established; Third, the core driver of AI PCs is still mixed up with the Windows 10 replacement; Fourth, the moat of AI glasses lies not only in the chip but also in lenses, channels, and social acceptance; Fifth, the biggest long-term winner in on-device AI is more likely the platform than the ordinary OEM.
The metrics most worth tracking over the next 6-12 months: AI PC share of PC shipments, flagship AI phone shipments and ASP, Ray-Ban/Oakley/China AI glasses sales, NPU installed base, average LPDDR/UFS content per device, Alexa+/Gemini paying users, and developer adoption of OS-level agents/cross-app actions.
For narrower follow-up research directions, we recommend prioritizing six: AI PCs, AI phones, AI smart glasses, on-device NPUs, AI operating systems/personal assistants, and the on-device AI supply chain. This is the closest current intersection of "revenue upside + profit upside + valuation gap."
Open Questions and Data Boundaries
This report prioritizes high-confidence, cross-verifiable public information. To upgrade it into an investment memo, three more categories of data are still needed: First, the latest rolling valuation multiples for each listed company, especially forward P/E, EV/EBITDA, and FCF yield; Second, the latest 2025-2026 financial figures for Asian supply-chain companies, including gross margin and a breakdown of AI-related revenue; Third, finer shipment and design-win data for smart glasses / AI earbuds / runtime companies. At this stage, a considerable portion of these items remains insufficiently disclosed by the companies, so this report uniformly labels this part as "needs further validation" rather than filling it with unverified figures.
This report is based on public information and does not constitute investment advice. Markets carry risk; invest with caution.
Full report
Sign in to read the full report
Sign up free to unlock the full text, the Baillie growth scorecard, and full-text search.
Log in / Sign up free