While the market is still debating whether the bottleneck of AI compute lies in chips, memory, or advanced packaging, the real "hard constraint" is surfacing—electricity. Over the past two years, US stock investors have grown used to hunting for opportunities around GPUs, HBM, and optical modules. But entering the second half of 2026, an increasingly clear consensus has emerged: whether AI data centers can be built and run at all depends first on having enough power. The second half of the compute race is shifting from "making chips" to "finding power," giving rise to a new theme overlooked by many investors—the power equipment supply chain.
Why Power Scarcity Has Become AI Compute's Toughest Constraint
To understand this power crunch, first grasp the scale of the demand shift. Traditional data centers rely mainly on general-purpose servers, with per-rack power in the single digits to low teens of kilowatts; AI training and inference clusters, by contrast, deploy high-density GPU servers en masse, pushing per-rack power to dozens or even hundreds of kilowatts, while entire campuses are planned in gigawatts—one large AI campus consumes nearly as much electricity as a mid-sized city. With demand surging and grid construction inherently long-cycle, the supply-demand gap is hard to close in the short term.
The process is even trickier. In the US, a large data center typically endures long queues and approvals from site selection and grid interconnection to formal power delivery. For tech giants, chips can be grabbed at a premium, but grid capacity upgrades allow no "queue jumping," making power the longest waiting line after land, capital, and chips. That is why tech giants are bypassing the grid to sign long-term power purchase agreements directly with generators—Microsoft's small modular reactor nuclear PPA, which we covered earlier, is a microcosm of this logic.
Gas Turbines: A Textbook Seller's Market
With nuclear approval cycles long and renewable output unstable, gas-fired power has become the fastest realistic option to fill the gap—relatively short construction periods, stable output, and deployment right next to data centers. As a result, order books at the world's major heavy-duty gas turbine makers have filled up fast; the industry widely reports new-order delivery schedules measured in years, with some large units stretching to several years.
From an investment perspective, this is a classic "capacity-constrained" business: core components of heavy-duty gas turbines—castings, blades, combustion chambers—carry high manufacturing barriers, and expanding capacity requires massive capex and years of construction, so no one can ramp up output at scale in the short term. That means extremely long order-book visibility, letting leaders "profit while customers queue." GE Vernova, Siemens Energy, and Mitsubishi Heavy Industries form a global heavy-duty gas turbine oligopoly; GE Vernova's continued market re-rating since its spin-off is a direct reflection of this logic.
Transformers and Grid Equipment: The Overlooked Invisible Bottleneck
If gas turbines solve "where the power comes from," transformers and grid equipment solve "how it gets delivered." Every link of an AI data center—high-voltage transmission, substation step-down, on-campus distribution—depends on transformers and switchgear. Global supply of large power transformers has long been tight, with delivery cycles likewise measured in years; slow capacity expansion upstream in grain-oriented silicon steel, copper, and other materials further stretches the supply curve.
Players in this segment are more dispersed: Hitachi Energy and Siemens Energy dominate the high-end large transformer market, while Eaton, Schneider Electric, and others each hold positions in medium- and low-voltage distribution; transformer and switchgear exporters listed in Hong Kong and A-shares are likewise being repeatedly repriced on this global power shortage logic. For US stock investors, the appeal of power equipment lies in this: demand is global and the client base is diverse, spanning tech giants and utilities; its cyclicality is milder than that of pure AI concept stocks, yet its growth slope is just as impressive.
From Grid to Rack: The Last Mile Is Equally Valuable
The job is not done once power reaches the campus. The "last mile" inside data centers—power distribution, uninterruptible power supplies, busways, and thermal management—are also beneficiaries of the high-density compute era. As rack power density rises, traditional power architectures are being forced toward higher voltages and fewer conversion stages, and the industry has even begun discussing solid-state transformers to rebuild data center power chains. Vertiv, Eaton, and peers have consequently been added to the core roster of the AI compute supply chain, with their orders increasingly correlated with data center capex.
Investment Logic and Risks: How to Track This Theme
- Watch backlog and delivery cycles: The coverage multiple of backlog to annual revenue and shifts in delivery timelines are the most direct indicators of industry momentum.
- Watch the pace of capacity expansion: Capacity is double-edged—expanding too slowly misses demand, while expanding too fast sows the seeds of a supply-demand reversal; management's capex guidance deserves more attention than single-quarter profits.
- Watch PPAs and power prices: The scale and price levels of tech giants' long-term power purchase agreements directly determine returns on generation and grid assets—an upstream signal of the whole chain's momentum.
- Beware of demand falling short: The biggest risk to this theme is AI compute demand itself. If large-model commercialization lags expectations and data center construction slows, the ultra-long order cycles of power equipment could turn into a liability.
Looking back from the vantage point of Q4 2026, the AI compute investment map has gone through multiple rounds of expansion: from GPUs to HBM, from optical interconnects to liquid cooling, from advanced packaging to nuclear power. Now it is the turn of gas turbines and grid equipment—devoid of the spotlight chips enjoy, they hold the "master switch" of the entire AI buildout. For investors, rather than agonizing over when the compute bubble will burst, it is better to see a plain fact: as long as AI keeps expanding, power will never be enough, and the "shovel sellers' shovel sellers" often offer the highest certainty of the cycle.
