Dominant AI chip leader for the compute phase

Wall Street Observer | The AI Decade Super Cycle: A Three-Stage Narrative, The 5×5×5 Investment Roadmap for Compute-Power-Physical AI The market is obsessed with chasing the next Nvidia. Most traders' perspectives are limited to short-term stock price surges, ignoring that AI is a structural capital expenditure revolution spanning 15 years. Elon Musk's aggressive expansion of the Colossus 2 million-GPU cluster is merely the prologue to this grand wave. When we extend the timeline to five-year stages, AI evolution can be divided into: Phase 1: AI Physical Compute (2026‑2030), Phase 2: AI Capability Realization (2031‑2035), and Phase 3: Physical AI Economy (2036‑2040). These three stages progress sequentially, with the focus of capital expenditure shifting downward step by step, moving from silicon wafers to electronics, and finally permeating every corner of the physical world. Core Logic: Compute is just the entry ticket, power is the mid-term bottleneck, and physical integration is the ultimate value. Capital will migrate along these bottlenecks in sequence; positioning ahead of bottlenecks is the core method for capturing excess returns. Phase 1 | AI Physical Compute 2026‑2030 (Next 5 Years): Frantically Piling Up Cloud Compute Infrastructure Theme Interpretation: AI Physical-Compute. We are currently in the middle-to-late stage of this cycle. The main theme of the market is building ultra-large-scale GPU clusters at any cost to solve the underlying problem of "whether it can run." Massive capital flows into chips, servers, optical networks, storage, and data centers. The track is already highly crowded, with leaders like Nvidia fully priced in, but there are still undervalued opportunities in niche supporting segments. The ultimate contradiction of this phase will slowly shift from GPU shortages to hard constraints on room space, cooling, and power supply, laying the groundwork for the second phase. Core Target Analysis • $AMZN: Ultra-large-scale cloud provider, AWS continues to take on global AI compute outsourcing needs, investing massive capital expenditures in self-built clusters, making it the most direct buyer of compute demand. • NVDA, AVGO: Dual giants dominate AI acceleration chips; Nvidia GPUs monopolize the training market, while Broadcom custom ASICs deeply penetrate cloud providers' private compute clusters. • $SNPS: EDA software, the industrial mother machine for chip design; no new generation of AI chips can do without its toolchain. • $TSM: TSMC, the core carrier of global advanced process manufacturing, the lifeline for GB200 and GB300 capacity. • $ASML: EUV lithography machine oligarch, the irreplaceable source of equipment for advanced chip manufacturing, with nearly insurmountable barriers. • $MU: Micron, the core domestic supplier of HBM; the thirst of large model training for high-bandwidth memory continues to break production capacity expectations. • ANET, LITE, COHR, CIEN: High-speed networking and optical communications. High-speed interconnection within clusters determines the actual efficiency of ten-thousand-card clusters; optical modules are the neural network of compute clusters. • $ARM: CPU architecture leader; in the era of AI Agents, the demand for coordinated scheduling between CPUs and GPUs explodes, opening up long-term ceilings for architecture licensing. • IREN, EQIX: AI data center operators, providing facility carriers and taking on cloud providers' outsourced compute deployments. • $VRT: Liquid cooling systems, a rigid demand for high-density GPU clusters; the higher the compute density, the more severe the cooling pressure. • DELL, SMCI: Server OEMs, directly benefiting from the global surge in compute cluster orders. • $PSTG: Enterprise-grade storage, responsible for reading, writing, and archiving massive training datasets. • $CRWV: GPU cloud service provider, offering elastic compute leasing for small and medium-sized enterprises. • $FIX: Data center engineering construction, undertaking civil works and electromechanical construction for server rooms. • $CRWD: Cybersecurity, as compute clusters expand, there is a rigid growth in demand for data security and cluster protection. Phase Risks: Rapid oversupply of compute capacity leads to industry price wars; cloud providers' capital expenditure growth rates temporarily decline. Phase 2 | AI Capability Realization 2031‑2035 (Second 5 Years): From Chip Shortages to Power Shortages Theme Interpretation: AI Power-Enable. When millions of GPUs are deployed, the harsh reality emerges: chips can be expanded quickly, but power infrastructure has a long construction cycle. Transformers, high-voltage grids, gas turbines, nuclear power, and mineral resources have construction cycles ranging from 5 to 10 years, unable to be delivered quickly like chips. In this phase, power replaces silicon wafers as the biggest choke point for AI expansion. The industry's contradiction shifts from "buying GPUs" to "matching stable, cheap, uninterrupted power supplies for data centers." AI no longer competes solely on chip quantity, but on inference capability per unit of energy consumption. Grid infrastructure, new power generation, energy storage, and critical metals face epic capital expenditure increases. Core Target Analysis • $PWR: Power engineering giant, responsible for large-scale grid upgrades and substation construction, the core contractor for data center supporting grid upgrades. • $ETN: Electrical equipment leader, transformer and high-voltage distribution equipment, the core supplier of hardware for powering compute parks. • $AEP: Large US utility company, responsible for grid interconnection, dispatching base power for large compute clusters. • $VST: Independent power producer, providing customized gas power generation for compute parks. • $GEV: General Electric gas turbines, large-power units ensure 7x24 uninterrupted power supply, adapting to the high stability requirements of data centers. • $BE: On-site distributed generation, providing self-sufficient energy solutions for data centers to reduce reliance on the public grid. • $EOSE: Energy storage systems, shaving peaks and filling valleys, smoothing out drastic electricity usage fluctuations in compute parks. • $OKLO: Advanced modular nuclear power, small reactors are expected to become stable baseload energy sources for remote compute bases. • $CCJ: Uranium mining leader, the most upstream resource for nuclear power expansion. • $FCX: Copper resources, grids, electrical equipment, and servers all heavily rely on copper; energy infrastructure expansion drives long-term copper demand. • $ALB: Lithium mining, upstream in the energy storage battery industry chain. • $MP: Domestic rare earths, strategic resources indispensable for motors and power semiconductors. Phase Risks: Delays in nuclear project approvals, shifts in energy policy, and剧烈 fluctuations in commodity cycles.