As we enter Q4 2026, the spotlight in the AI computing race still shines on GPU stockpiling, HBM capacity, and Big Tech capex figures. But a quiet shift deserves more attention than “which chipmaker just beat expectations again”: the industry's biggest constraint is gradually moving from chip supply to the kilowatt-hours on the meter. Electricity—old infrastructure long overlooked in the compute narrative—is becoming the hardest bottleneck on the road to AI data center expansion, and it is quietly redrawing the investment map for US AI stocks. Put another way, every dollar of AI capex now has to be split in two: half for chips, the other half for the power systems that keep those chips running.
1. From Chips to Kilowatt-Hours: How Did Power Become the Compute Bottleneck?
To understand the power shortage, start with the leap in per-device consumption. In the V100 and A100 era, a training server drew only a few kilowatts, well within a single rack's capacity; by the H100 generation, per-card power was pushed to about 700 watts, and NVIDIA's Blackwell generation approaches the kilowatt scale—a full GB200 NVL72 rack can consume on the order of 100 kW, more than ten times a traditional general-purpose rack. The more chips you pack in, the scarcer power becomes; more crucially, the supporting power and cooling infrastructure must be rebuilt from scratch, and liquid cooling has rapidly shifted from an option to a standard feature.
What's worse is the mismatch in timing. A hyperscale data center typically takes only one to two years from groundbreaking to operation, but new grid interconnection queues in the US often run for years, and delivery lead times for high-voltage transformers and switchgear are just as long. Multiple industry surveys show that order backlogs for transformers and large gas turbines now stretch years out, and “money in hand, land secured, but no power” has become the norm for many projects. The International Energy Agency (IEA) has also specifically noted that global data center electricity consumption will still grow sharply by 2030, and the contribution of US data centers to electricity demand growth may even exceed that of energy-intensive industries such as aluminum, steel, cement, and chemicals combined.
2. Big Tech's Power Rescue: Nuclear, Gas Turbines, and Self-Generation
Faced with the power gap, the cloud giants chose the most direct approach—locking in power upstream.
Nuclear: A Decade-Long Bet
As early as 2024, Microsoft signed a 20-year nuclear power purchase agreement with Constellation Energy to restart a Three Mile Island unit dedicated to data centers; Google signed a series of agreements with small modular reactor (SMR) company Kairos Power to lock in future nuclear capacity ahead of time; Amazon bought a data center campus next to the Susquehanna nuclear plant directly from Talen Energy, literally “building server rooms beside the power source”; and Meta issued a nuclear procurement request for proposals, seeking gigawatt-scale new capacity. A basic consensus has taken shape in the industry: whoever locks in stable, clean power first holds the initiative in the next round of compute expansion.
Gas Turbines and Self-Built Power: A Faster Stopgap
Beyond nuclear, natural gas has become the faster bridge. Gigawatt-scale gas turbines were snapped up long ago, and some giants are simply building “behind-the-meter” captive power plants next to their data center sites to bypass the grid queue; solutions such as renewables paired with storage and fuel cells are being piloted in parallel. In other words, the contest in the AI computing race is quietly shifting from “who grabs more GPUs” to “who can get more reliable electricity.”
3. The US Stock Investment Map Amid the Power Shortage: Where Will the Money Flow?
For ordinary investors, the opportunities created by the power shortage fall roughly into four threads:
- Power and cooling equipment: Integrated “power distribution plus cooling” solution providers such as Vertiv are among the most direct beneficiaries of AI data center expansion, with orders continuously driven by hyperscale customers; the transformer and switchgear businesses of Eaton and peers are similarly oversubscribed, with lead times stretching ever longer;
- Generation assets: Nuclear and gas generation assets represented by Constellation Energy and Vistra are being repriced on the value of long-term power purchase agreements (PPAs); Talen, which transformed itself through the “nuclear plant plus data center” model, is the landmark case of this round of revaluation;
- Power engineering and the grid: Transmission and substation contractors such as Quanta Services and gas turbine manufacturers such as GE Vernova are positioned right on the obligatory path of grid expansion;
- Backup and distributed power: Diesel and gas backup generator sets from Caterpillar and Cummins, along with Bloom Energy's fuel cells, have also been pushed onto the stage by “power anxiety.”
One caveat: this supply chain has rallied significantly over the past two years, and some names' valuations already embed quite optimistic growth expectations. Whether the power story can deliver over the long run still needs to be validated by actual order wins and earnings conversion.
4. The Other Side of the Party: Overbuilding Concerns Cannot Be Ignored
Beneath the heat of the power shortage, dissenting voices exist too. Some institutions worry that if the commercialization of AI applications falls short of expectations, today's rapidly launched data center and power projects could face declining utilization and asset impairments in the future; the controversy over “circular financing”—suppliers taking stakes in customers who then use the money to buy the suppliers' products—has repeatedly rattled market sentiment. Moreover, hard constraints such as power permitting, interconnection approvals, and community opposition will not disappear overnight; any change in the pace of any single link could affect how quickly Big Tech capex lands and how depreciation plays out. For investors, tracking what companies say about capex guidance and data center commissioning progress on earnings calls is more meaningful than chasing short-term headlines.
The advice for newcomers boils down to two principles: first, look at “contract visibility”—companies with long-term power purchase agreements and anchored large-customer orders weather cycles better than purely conceptual partnerships; second, look at cash flow and earnings conversion, and beware of companies that only tell stories while profits never materialize.
5. Conclusion: Kilowatt-Hours Are the Foundation Logic of the Compute Era
From GPUs to HBM to today's electricity, every wave of the AI compute rally has essentially been about finding the scarcest link and repricing it. As chips gradually cease to be the only bottleneck, electricity—the hardest constraint of all—is passing the baton to the power equipment, generation asset, and grid supply chains. For investors looking to position in US stocks along the AI theme, understanding the data center “power anxiety” matters more than memorizing a few more chip models—after all, even the most powerful chip needs power first before it can spin up.
