Overview
The AI investment theme has already produced obvious mega-cap winners, but this page is designed for the less obvious enabling layer: companies selling the tools, silicon, photonics, test systems, connectivity, embedded compute and edge-AI platforms needed to make AI work outside the cloud and inside real devices.
The screen deliberately separates cloud AI, where scale advantages dominate, from edge and physical AI, where microcap and small-cap suppliers can still own important bottlenecks. The strongest recurring overlap with the Robotics and Human Augmentation pages is edge inference: robots, wearables, medical devices, drones and industrial sensors all need low-power intelligence near the point of action.
ScalingTheme maturity
Very HighCapex intensity
StrongCross-theme relevance
SelectiveMicrocap investability
Stock Table
Working AI infrastructure watchlist. This is not a buy list; it is a research table for tracking enabling suppliers around AI infrastructure and edge intelligence.
| Rank | Company | Ticker | Role in AI stack | Cap / style | Research view |
|---|
| 1 | CEVA | CEVA | Licensable edge-AI NPU, DSP, connectivity and sensor-fusion IP | Borderline microcap / small-cap | Best quality edge-AI IP bottleneck; royalty/licensing model with high gross margin and Physical AI relevance. |
| 2 | Ambiq Micro | AMBQ | Ultra-low-power AI SoCs for wearables, hearables, healthcare and edge devices | Growth micro/small-cap | Excellent thematic fit for always-on AI; valuation and losses mean position sizing matters. |
| 3 | Lantronix | LTRX | Edge AI, Industrial IoT, onboard compute and connectivity for unmanned systems | Microcap | Attractive edge/physical-AI infrastructure angle; drone/UAS momentum is a real catalyst. |
| 4 | Aehr Test Systems | AEHR | Wafer-level and packaged-part burn-in/test for AI processors, HPC and photonics | Small-cap | AI processor reliability/testing bottleneck; strong future visibility but cyclical and recently loss-making. |
| 5 | QuickLogic | QUIK | eFPGA IP, ruggedized FPGAs, endpoint AI and sensor/voice processing | Microcap | Interesting government/eFPGA/edge-AI option; valuation and contract lumpiness are the key constraints. |
| 6 | POET Technologies | POET / PTK.V | Photonic integrated circuits, optical engines and optical interposer for AI/data centres | Speculative photonics small-cap | Potentially powerful AI interconnect story, but extremely speculative relative to current revenue. |
| 7 | ACM Research | ACMR | Wafer processing, cleaning and advanced packaging equipment | Small-cap / mid-cap boundary | Not AI pure-play, but advanced packaging and wafer-processing exposure matters for AI chips. |
| 8 | BrainChip | BRN.AX | Neuromorphic edge-AI IP | Speculative micro/small-cap | Thematically pure but still too early; valuation-to-revenue risk is high. |
| 9 | Kopin | KOPN | Microdisplays, optics and near-eye systems for defence, enterprise and wearable AI interfaces | Small-cap | AI interface adjacency rather than core AI; better suited to HMI/spatial computing watchlist. |
| 10 | Cerence | CRNC | Automotive conversational AI and in-car HMI software | Small-cap | Direct AI software revenue, but auto-cycle, customer concentration and competition matter. |
Value Chain Map
AI is best mapped as a stack. The lower layers are harder to replace and usually more capital intensive; the higher layers can scale faster but face heavier competition and lower switching costs.
| Layer | What matters | Example companies | Investment note |
|---|
| Power and cooling | Grid capacity, substations, power electronics, liquid cooling, thermal management | Mostly larger industrials; to be expanded in Energy page | Second-order AI infrastructure bottleneck; huge but often not microcap-friendly. |
| Semiconductor manufacturing | Wafer processing, advanced packaging, test, burn-in, reliability | Aehr, ACM Research | Critical for AI accelerators, ASICs, memory, photonics and high-reliability compute. |
| Interconnect and photonics | Optical engines, silicon photonics, co-packaged optics, transceivers | POET, Aehr via photonics test | AI scaling increasingly depends on data movement, not just compute. |
| Edge AI silicon and IP | NPUs, DSPs, sensor fusion, ultra-low-power SoCs, eFPGA | CEVA, Ambiq, QuickLogic, BrainChip | Best microcap hunting ground because AI moves into devices, robots and wearables. |
| Embedded compute and physical AI | Industrial IoT, drones, onboard compute, rugged gateways | Lantronix, SECO, Kontron as larger reference | Connects AI with robotics, defence, smart infrastructure and industrial automation. |
| Application and workflow layer | Agents, copilots, enterprise automation, vertical AI, in-car AI | Cerence; larger software names not listed here | More obvious and competitive; stronger revenue but potentially less hidden. |
Sub-Themes
- Edge AI: inference in wearables, robots, drones, vehicles, industrial sensors and medical devices.
- Physical AI: AI moving from screens into machines that sense and act in the real world.
- AI semiconductors: accelerators, custom ASICs, advanced packaging, test and burn-in.
- AI data-centre infrastructure: power, cooling, optical networking, racks, deployment and reliability.
- Agentic software: autonomous workflows, enterprise copilots, tool-using agents and AI orchestration.
- AI interfaces: voice, eye tracking, haptics, wearables, in-car assistants and multimodal user interfaces.
Market Forces
- Capex arms race: hyperscalers and AI labs are spending heavily on compute, networking and power.
- Custom silicon: AI ASICs create demand for specialised design, test, burn-in and packaging suppliers.
- Data movement bottleneck: photonics and interconnect become more important as clusters scale.
- Inference moves to the edge: robots, wearables and industrial systems cannot rely entirely on cloud AI.
- Commoditisation risk: model capability diffuses quickly; infrastructure bottlenecks may be more defensible.
- Power constraint: AI growth increasingly depends on energy availability and thermal management.
Technology Deep Dive
The AI stack is becoming a physical infrastructure problem. Model progress still matters, but investable bottlenecks increasingly sit in compute density, memory bandwidth, networking, photonics, power, cooling, semiconductor reliability and efficient inference at the edge.
| Bottleneck | Why it matters | Public-market angle |
|---|
| On-device inference | Latency, privacy, battery life and reliability require AI to run locally in many devices. | CEVA, Ambiq, QuickLogic, BrainChip. |
| AI processor reliability | High-power AI accelerators need burn-in and test to reduce failure risk in data-centre environments. | Aehr Test Systems. |
| Optical interconnect | Large AI clusters are limited by how fast data can move between chips, boards, racks and data centres. | POET, silicon-photonics suppliers, photonics test equipment. |
| Advanced packaging | AI chips depend on high-bandwidth memory, chiplets and complex packaging workflows. | ACM Research, semiconductor equipment ecosystem. |
| Embedded physical AI | Drones, robots and industrial systems need rugged, connected edge compute. | Lantronix and SECO-style embedded platforms. |
| AI interfaces | AI systems need new ways to receive human input and return useful output: voice, gaze, touch and spatial displays. | Cerence, Tobii, Kopin, Interlink, Immersion. |
Company Profiles
1. CEVA · CEVA
Edge-AI IP, NPU/DSP licensing and connectivity · core AI infrastructure watchlist
CEVA licenses silicon and software IP for smart-edge devices. Its NeuPro NPU, SensPro DSP, wireless connectivity and sensor-fusion IP make it a picks-and-shovels supplier for Physical AI: devices where connectivity, sensing and inference converge.
- Why it matters: AI IP licensing can scale through customer silicon without CEVA needing to manufacture chips.
- Recent evidence: 2025 included 10 NeuPro NPU agreements, AI contributing more than 20% of licensing revenue, and 2.1 billion CEVA-powered devices shipped.
- Main risks: licensing-cycle lumpiness, royalty ramp timing and valuation.
- Research rating: highest-quality edge-AI picks-and-shovels name.
2. Ambiq Micro · AMBQ
Ultra-low-power AI SoCs for edge devices
Ambiq is a direct way to track the move from cloud-only AI to always-on AI in wearables, hearables, healthcare, industrial sensors and small battery-powered devices. Its SPOT technology focuses on ultra-low-power processing, and Atomiq adds an NPU to the stack.
- Why it matters: the edge-AI problem is not just compute power; it is useful AI within tiny power and thermal budgets.
- Recent evidence: FY2025 net sales were $72.5m, Q4 2025 non-GAAP gross margin was 45.5%, and Q1 2026 net sales guidance was $21m–$22m.
- Main risks: still loss-making, post-IPO valuation, semiconductor cyclicality and customer concentration.
- Research rating: high thematic fit but valuation-sensitive.
3. Lantronix · LTRX
Edge AI, Industrial IoT and drone/UAS compute
Lantronix supplies edge AI and Industrial IoT solutions for unmanned systems, critical infrastructure and enterprise networks. It is a practical physical-AI supplier rather than a model company.
- Why it matters: drones and unmanned systems need onboard compute, connectivity and compliant embedded platforms.
- Recent evidence: fiscal Q3 2026 revenue was $30.2m, non-GAAP EPS was $0.04, Embedded IoT Solutions grew 22%, and management increased FY26 drone revenue expectations to $10m–$14m.
- Main risks: small scale, customer concentration, defence/UAS programme timing and hardware margins.
- Research rating: physical-AI microcap watchlist.
4. Aehr Test Systems · AEHR
AI processor, HPC and photonics test/burn-in
Aehr is a semiconductor reliability and test bottleneck. As AI processors, ASICs, photonics and high-performance devices become more power-dense, the need for wafer-level and packaged-part burn-in rises.
- Why it matters: AI infrastructure needs reliable chips, not just faster chips.
- Recent evidence: Aehr said it had improved visibility from AI processor and data-centre semiconductor test and burn-in, with expected second-half FY2026 bookings of $60m–$80m and AI processor customer activity in wafer-level and packaged-part burn-in.
- Main risks: revenue volatility, customer concentration, cyclical semiconductor spending and recent losses.
- Research rating: high-upside infrastructure supplier, but not low-risk.
5. QuickLogic · QUIK
eFPGA IP, endpoint AI, rugged FPGA and sensor/voice processing
QuickLogic sits in the programmable logic and endpoint-AI niche. Its relevance is not broad model leadership, but eFPGA IP, ruggedized FPGAs and sensor/voice processing at the edge.
- Why it matters: eFPGA and endpoint-AI IP can matter in custom silicon, defence, industrial and embedded applications.
- Recent evidence: FY2025 results highlighted eFPGA IP, ruggedized FPGAs and Endpoint AI; 2025 announcements included Intel 18A-related eFPGA work, government programme funding and design wins.
- Main risks: contract lumpiness, high valuation versus current revenue, government timing and debt.
- Research rating: speculative but relevant edge-AI/IP watchlist.
6. POET Technologies · POET / PTK.V
Photonic integrated circuits, optical engines and AI/data-centre interconnect
POET is one of the more dramatic AI-photonics stories. Its optical interposer and photonic products are aimed at AI and data-centre markets, where data movement and optical connectivity are increasingly strategic.
- Why it matters: AI scaling increasingly depends on interconnect and photonics, not only GPU count.
- Recent evidence: the company reported Q4 2025 results and said it had moved from development to execution, supported by more than $225m of financing in Q4 and a further $150m in January 2026.
- Main risks: extremely speculative relative to revenue, financing history, execution risk and valuation volatility.
- Research rating: speculative photonics optionality, not core quality list.
7. ACM Research · ACMR
Wafer processing and advanced packaging equipment
ACM Research is not an AI pure-play, but its wafer-processing and advanced-packaging equipment sits in the semiconductor manufacturing chain supporting AI chips and complex packaging.
- Why it matters: AI chip supply depends on process complexity, packaging and manufacturing throughput.
- Recent evidence: ACM reported record 2025 annual revenue of $901m, up 15% year-on-year, and described product progress across cleaning, furnace and panel-level plating solutions.
- Main risks: China exposure, export controls, semiconductor capex cycles and not being AI-specific.
- Research rating: semiconductor infrastructure watchlist rather than microcap core.
Future Scenarios
Bull case: AI demand broadens from cloud GPUs into edge devices, robots, wearables, drones, medical devices and industrial systems. Bottleneck suppliers in low-power inference, photonics, test and embedded compute see sustained order growth.
Base case: AI capex remains strong but uneven. Obvious mega-cap winners dominate, while microcap opportunities work only where there is a clear bottleneck, design win or revenue inflection.
Bear case: AI infrastructure capex pauses, model economics disappoint, power bottlenecks slow deployment, and speculative suppliers with high valuations rerate sharply before revenue arrives.
Signals to Watch
- Edge-AI design wins converting into shipped silicon and royalty revenue.
- AI processor burn-in/test bookings and backlog at Aehr.
- Drone/UAS revenue growth and OEM engagement at Lantronix.
- Ambiq Atomiq and neuralSPOT adoption in wearables, hearables and healthcare devices.
- POET purchase orders, manufacturing scale-up and optical-engine customer validation.
- AI data-centre power, cooling and networking constraints.
- Enterprise agent adoption moving from demos to paid workflow automation.
Metrics That Matter
- Inference cost per useful task: more important than model benchmark scores for real adoption.
- Power per inference: essential for edge AI and wearable AI.
- Design wins and royalty conversion: crucial for CEVA, Ambiq, QuickLogic and POET.
- Bookings / backlog: critical for semiconductor equipment and data-centre suppliers.
- Gross margin: separates IP and specialist suppliers from commodity hardware.
- Customer concentration: a single hyperscaler, OEM or government programme can distort the thesis.
- Capex sensitivity: AI infrastructure suppliers can fall hard if spending pauses.
Risk Map
- Valuation risk: many AI-adjacent companies trade on future narratives rather than current revenue.
- Capex cyclicality: AI infrastructure orders can arrive in waves and then pause.
- Commoditisation: model and application layers may face faster competition than expected.
- Power limits: data-centre expansion may be constrained by grid access and cooling.
- Geopolitics: export controls and China exposure matter across semiconductor equipment.
- Customer concentration: small suppliers can be dependent on one or two major programmes.
- Execution risk: photonics, edge AI and semiconductor test stories need real production ramps.
Convergence
- AI + Robotics: Physical AI requires edge inference, motion, perception and sensor fusion.
- AI + Human Augmentation: wearables, neurotech and medical interfaces need low-power AI and cognitive software.
- AI + Energy: data-centre growth drives power, cooling, grid and efficiency bottlenecks.
- AI + Next-Gen Computing: photonics, neuromorphic computing, quantum, advanced packaging and custom ASICs.
- AI + Cybersecurity: agents increase identity, access, model-security and trust requirements.
- AI + Mobility: autonomy, drones, in-car AI and sensor-fusion systems.
Research Library
Source links used for this first filled-out AI page:
Summary
The AI page should not become another list of obvious mega-cap AI winners. Its value is in mapping the second- and third-order suppliers: edge AI IP, ultra-low-power silicon, test and burn-in, embedded compute, photonics, advanced packaging and AI interface infrastructure.
Current working conclusion: CEVA is the highest-quality edge-AI IP name; Ambiq is the strongest pure low-power edge-AI silicon story but valuation-sensitive; Lantronix is an interesting physical-AI microcap; Aehr is an AI processor reliability/test bottleneck; QuickLogic and POET are more speculative IP/photonics options; ACM Research belongs on the broader semiconductor infrastructure watchlist.