Data Center Electricity Demand Growth
+100%
▲ '22→'26 IEA Forecast
Edge AI Market CAGR
38%
▲ 2025~2030E
Agentic AI Market Size 2030E
45B$
▲ '26 $8.5B → '30 $45B
Ratio of Companies with Complete AI Security Strategies
6%
Gartner 2026 · Serious Gap
The 2024~2025 rally was predictable. GPU shortage → HBM shortage → Optical communication shortage. The bottleneck was clear, and the market reflected the price in that order. The current COHR at $250 and LITE at $677 already contain that story.
The problem is the next bottleneck. The AI ecosystem is now transitioning from overinvestment in "Training" infrastructure to the "Inference Diffusion" phase. Unlike training, inference creates completely different bottlenecks: power, edge, security, and automation. Most investors are still reading this transition only as an extension of the GPU theme.
🌿 Growing — Stock price already reflecting / 2nd order opportunity exists
THEME 01 · POWER INFRA
Power Is the New GPU
— AI Power Infrastructure
Power Grid · SMR · Transformer
Why Now: In 2026, data center power consumption in the United States exceeded 4% of the entire country. In Northern Virginia (the world's largest DC cluster), the wait for new facility power connections is 3~5 years. The situation where "you can buy a GPU but have nowhere to plug it in" has become a reality.
What the Market Missed: Nuclear power and independent power producers like Vistra (VST) and Constellation (CEG) have already risen by 20~80%. The real opportunity is in 2nd-tier beneficiaries: power grid connection contractors (Quanta Services·PWR), transformer manufacturers (SPX Technologies·SPXC), and cooling instrumentation (Watts Water·WTS). And the longest-term play is SMR — NRC streamlining of permits is underway, Oklo (OKLO) listing completed.
Risk: SMR commercialization may be delayed until 2028~2030. Short-term play is construction companies like PWR·SPXC.
PWR
SPXC
OKLO
CEG
VST
🌱 Emerging — Still low analyst coverage
THEME 02 · EDGE INFERENCE
Inference Leaves the Cloud
— Edge AI Inference
On-Device · AI PC · Industrial Edge
Why Now: Gartner forecasts that AI PCs will account for 55% of all PCs in 2026. The number of use cases where cloud inference costs ($0.001~0.01/query) are not economical compared to on-device (0 won) is exploding. Qualcomm CEO declared at Davos that "the edge winner is the ultimate AI winner."
What the Market Missed: ARM Holdings (ARM) is well known, but CEVA (CEVA) — mobile NPU·DSP IP licensor, entering the industrial edge with a strategic partnership with Microchip MCU — is still off the radar with a market cap of $2.6B. Arteris (AIP) — in-chip data routing IP, an essential component for increasing SoC complexity. Lattice Semiconductor (LSCC) — low-power FPGA, a key player in industrial and edge AI.
Risk: On-device LLM quality gap, high-performance inference still relies on the cloud.
CEVA
AIP
LSCC
ARM
🌱 Emerging — Buzz started after CES 2026
THEME 03 · PHYSICAL AI
AI Gets a Body
— Physical AI and Robotics
Humanoid · Cobot · Industrial Vision
Why Now: Jensen Huang declared at CES 2026 that "the ChatGPT moment for physical AI has arrived." Hyundai Atlas announced factory deployment, Figure AI valuation from $2.6B→$39B in 7 months. NVIDIA GR00T·Isaac Sim is playing the role of GPT-3 for robot learning data.
What the Market Missed: Opportunity lies in robot component suppliers rather than humanoid itself. Cognex (CGNX) — industrial machine vision, occupying 70% of robot eyes. Currently adjusting from $35→$28 (bottom of sales cycle). Zebra Technologies (ZBRA) — industrial IoT·RFID, essential for connecting automated factories. Teradyne (TER) — collaborative robot (UR) subsidiary + semiconductor tester double beneficiary.
Risk: Delay in mass deployment if 99%+ reliability is not met. 2~3 year pipeline.
CGNX
TER
ZBRA
ROK
🌱 Emerging — Essential infrastructure in the agent era
THEME 04 · AI SECURITY & GOVERNANCE
As Agents Increase,
A Defense Is Needed
AI Agent Security · Data Governance · Sovereign AI
Why Now: AI agents are expected to be installed in 40% of enterprise applications in 2026 (Gartner). However, only 6% of companies have mature AI security strategies. AI agents access enterprise systems at a ratio of 82:1 compared to existing people — creating a new attack surface. Even Anthropic revealed cases of state-sponsored hacking groups automating cyberattacks using Claude.
What the Market Missed: CrowdStrike·Palo Alto are already overvalued. An earlier play is SentinelOne (S) — AI-based autonomous detection and response, still in the early stages of monetization, resulting in a valuation discount. Zscaler (ZS) — zero-trust network, an essential architecture in the agent era. Informatica (INFA) — data governance·master data, the biggest beneficiary of strengthened AI governance regulations.
Risk: Regulatory Timeline uncertainty. Depends on the speed of AI governance legislation.
S
ZS
INFA
PANW