TradingKey - When the CEO of Nvidia (NVDA) discussed artificial intelligence at the G20, he did not focus on GPUs or the latest chips, but instead presented a more foundational framework: the AI economy is like a "five-layer cake," with energy at the bottom, followed sequentially by chips, land and data center infrastructure hosting the system, models, and ultimately data and applications that create commercial value.
This definition changes the perspective on AI investment. AI is not merely cloud software services, but a highly coupled chain of physical infrastructure: without stable, accessible, and scalable electricity, even the most advanced chips cannot be converted into computing power. Today, data center power supply is becoming a critical variable in whether AI capital expenditures can be implemented on schedule.
AI’s First Layer: Why Power Shifted From a Cost to a Constraint
Jensen Huang's concept of "power turning into compute" points to the simplest and most easily overlooked physical fact of AI computing: both training and inference fundamentally consume electrical power. As power per rack continues to rise, data centers are no longer just large-scale electricity consumers, but new incremental loads that are reshaping power grid planning, generation mix, and equipment supply and demand in local regions.
According to forecasts by the International Energy Agency (IEA), global data center electricity consumption is expected to nearly double to around 945 terawatt-hours by 2030. Forecasts from multiple institutions also point to the same trend: AI-optimized servers are significantly raising the electricity consumption intensity of data centers. The shift in demand is not limited to total volume, but lies even more in its concentration and delivery speed—projects in hot-spot regions often need to secure gigawatt-scale power within one or two years, whereas power transmission, new substations, and large transformers typically require several years to build and deliver.
This also explains the judgment emphasized by Berkshire Hathaway CEO Greg Abel that energy remains a "major constraint" on the rapid expansion of AI data centers. In his view, Berkshire sees not only a challenge but also a long-term opportunity, provided that new data center demand does not pass costs onto existing utility customers and that project locations can accept its impact on public resources such as water.
Power Is Not Just About 'Generation,' It's Also About Delivering Power to Racks
Over the past two years, the rotation path of the AI narrative in US stocks has been relatively clear: chips, servers, storage, networking, and liquid cooling have successively taken center stage in the market. The uniqueness of the power theme lies in the fact that it is not a cyclical shortage of a specific hardware component, but rather a systemic issue spanning power generation, grid connection, transmission and distribution, and rack-level power supply.
As grid connection queues lengthen, some hyperscalers have begun evaluating on-site power generation or near-site power supply. Gas turbines, fuel cells, long-term nuclear power purchase agreements (PPAs), and energy storage correspond to different delivery speeds, reliability, and cost structures. The core logic driving their returns lies in the real-world demand of data centers for reliable power and faster commissioning.
The generated electricity must also be transmitted to the campus and then converted and distributed to servers and cooling systems. Extended lead times for large transformers, switchgear, circuit breakers, and substation equipment have become hard constraints on project commissioning. Changes in orders and backlogs at equipment and engineering service providers such as Eaton and Quanta Services are therefore viewed by the market as leading indicators for monitoring AI power capital expenditures.
As AI rack power evolves from traditional levels to hundreds of kilowatts or even higher, power delivery efficiency and thermal management begin to jointly determine available computing power. High-voltage direct current (HVDC) power supply, UPS, energy storage, solid-state transformers (SST), and liquid cooling are not independent of each other: with every increase in compute density, power distribution capacity, redundancy design, and heat dissipation capabilities must be upgraded in lockstep. Market focus may shift further from 'whether electricity is available' to 'whether electricity can be stably delivered to the rack with sufficiently high efficiency.'
AI Power Chain Is Not the Next Memory Trade, But a Longer Build-Out Cycle
Comparing AI power to the previous memory trade helps clarify why capital is chasing supply bottlenecks, but the two are not entirely identical.
Supply-demand imbalances in memory products such as HBM can eventually be eased through capacity expansion and yield improvements. In contrast, power infrastructure involves permitting, land acquisition, transmission lines, equipment manufacturing, engineering construction, and regulatory coordination, making supply response inherently slower.
According to public industry data, disclosed project pipelines for North American data centers are massive, yet a time lag remains between grid interconnection and key equipment delivery. For companies, what truly determines the degree of benefit is not whether their name is associated with AI, but whether they can convert demand into verifiable orders, backlog, and margin expansion. Equipment manufacturers' orders and capacity utilization, power generators' long-term power purchase agreements (PPAs), and engineering service providers' active project backlogs generally offer more valuable guidance than pure thematic narrative.
Company Name | Ticker | Segment | Core Positioning & Competitive Advantage |
GE Vernova | GEV | Gas Turbines, Power Generation Equipment & Grid Technology | One of the broadest equipment providers in the AI power chain: offering both fast-add power sources like gas turbines and covering transformers, HVDC, and grid equipment. |
Bloom Energy | BE | On-Site Power Generation / Fuel Cells | A representative company in solid oxide fuel cell (SOFC) on-site power generation. Its key selling point is relatively fast deployment and continuous power delivery, suitable for data center campuses seeking to shorten "time to power"; however, project execution and customer concentration require ongoing validation. |
Constellation Energy | CEG | Nuclear Baseload / Independent Power Producer | One of the leading nuclear power operators in the U.S., with the advantage of providing 24/7 low-carbon baseload power. Its long-term power purchase agreements (PPAs) with tech giants make it a key generator asset operator in the "AI load growth – long-term power contract" thesis. |
Vistra | VST | Nuclear, Gas & Integrated Power Generation | Owns diversified generation assets including nuclear and gas, operating across multiple competitive U.S. power markets. If AI data centers drive up regional load and forward power prices, its asset portfolio and long-term supply arrangements stand to benefit, though it remains sensitive to power price and regulatory changes. |
Talen Energy | TLN | Direct Nuclear Supply / Independent Power Producer | Anchored by generation assets in the PJM region, market focus centers on its nuclear resources and long-term power supply arrangements with large data centers. |
Eaton | ETN | Transmission & Distribution Equipment, UPS & Power Management | Covers key nodes from "grid connection to campus distribution to rack-level power supply" across switchgear, circuit breakers, busways, transformers, and UPS. Its technology stance is relatively neutral; whether power comes from gas, nuclear, or renewables, its distribution and protection equipment remains essential. |
Vertiv | VRT | Data Center Power & Thermal Management | A pure-play provider of internal data center infrastructure, covering UPS, PDUs, switchgear, and liquid cooling. As AI rack power density rises, power conversion efficiency and thermal management become simultaneous hard requirements, making its orders more sensitive to hyperscale buildout pace. |
nVent Electric | NVT | Busways, Enclosures & Power Distribution Components | Focuses on high-power distribution, enclosures, and connection solutions, benefiting from internal data center distribution upgrades and modular construction. Compared to full-stack giants, its positioning targets the "last mile" of power distribution; order elasticity and product market share are key indicators to watch. |
POWL | Medium/High-Voltage Switchgear & Custom Power Distribution | Provides customized medium- and high-voltage switchgear and power distribution systems adapted to high-load connections for large industrial facilities and data centers. Strengths lie in engineering capabilities and high-spec project delivery; attention should be paid to converting order backlog into revenue and profit. | |
Caterpillar | CAT | Diesel / Gas Generator Sets & Bridging Power | Deep supply capabilities in large generator sets, suitable for data center backup power, bridging power, and select distributed power scenarios. Its AI upside stems mostly from incremental orders in its electric power systems business, rather than company-wide performance being directly tied to data centers. |
Cummins | CMI | Diesel / Gas Backup Power | A major supplier of diesel and gas generator sets for critical infrastructure, while also expanding into hybrid power and fuel cells. Data centers' demand for highly reliable backup power provides direct application scenarios for its engine and power systems segments. |
Generac | GNRC | Backup Power, Microgrids & Energy Storage Integration | A leading North American backup power provider expanding into commercial and data center high-capacity units and microgrid solutions. The core thesis centers on enhancing power resilience for critical loads, though data center revenue share and order fulfillment require separate tracking. |
Quanta Services | PWR | Transmission, Substation & Grid Interconnection Engineering | A leading North American grid infrastructure engineering firm, responsible for constructing transmission lines, substations, and large-load grid connections. Moving an AI campus from "site acquisition" to "actual energization" relies heavily on such engineering capabilities; its benefits skew toward grid capex and construction cycles. |
For the market, the next step should not merely be hunting for "AI power concepts," but identifying which segment of supply is hardest to expand, which companies possess genuine orders and delivery capabilities, and who ultimately bears the cost of incremental load. Whether the re-rating of power infrastructure persists depends on whether these questions can be answered one by one.
Find out more