The Energy Bottleneck: Why the Power Grid Is the Biggest Constraint on AI — and Which Data Centre Power Companies Stand to Win

By Balaji N, CFP · True Value Research · June 16, 2026

The Watt Problem

The most consequential constraint on artificial intelligence is not compute, software, or talent. It is electricity. Every large language model trained, every inference request answered, every GPU cluster spun up requires a continuous, reliable supply of power at a scale that is now straining the infrastructure of the world's most advanced economies.

A single NVIDIA H100 GPU draws roughly 700 watts under load. A standard hyperscale AI training cluster contains 10,000 to 100,000 GPUs. That is 7 to 70 megawatts for compute alone — before cooling systems, networking equipment, and power distribution overhead, which typically add another 40 to 50 percent. A single large AI data centre today consumes 200 to 500 megawatts. The data centres Microsoft, Google, Meta, and Amazon are planning for 2027 and beyond will require 1 gigawatt or more — the equivalent of a mid-sized nuclear power station, dedicated entirely to a single facility.

To understand the magnitude of this shift: the entire US data centre industry consumed approximately 200 terawatt-hours of electricity in 2023 — about 4 percent of total US consumption. The Department of Energy projects that figure will triple to 600 terawatt-hours by 2028, driven almost entirely by AI workloads. Some forecasts suggest it could reach 1,000 terawatt-hours — roughly 20 percent of US electricity consumption — before the decade is out.

The US power grid was not built for this. The grid's foundational architecture was designed in the 1950s and 1960s for a demand profile dominated by residential and industrial loads that grow slowly and predictably. AI data centres are creating demand spikes that are unprecedented in speed, scale, and concentration. New grid connections that once took 18 months to approve and build now face 5-to-7-year queues in the most constrained markets — Northern Virginia, the Texas Triangle, the Phoenix metropolitan area.

The result is a structural opportunity for companies across the full power-delivery stack. The constraint is real, the capital is committed, and the timeline is not decades but years. Below, we map the companies best positioned to benefit.

Why AI Power Demand Is Different

Traditional data centres run at relatively predictable utilisation rates, making their power draw manageable for grid operators. AI inference clusters are different: they must respond to unpredictable query volumes while maintaining near-zero latency, which requires maintaining large amounts of idle compute capacity on standby. AI training runs, meanwhile, draw full power continuously for weeks or months without interruption. Grid operators call this a "must-run" load profile — it cannot be curtailed during peak demand periods without destroying the value of the workload.

This creates three interlocking problems. First, hyperscalers need power that is available 24 hours a day, 365 days a year, regardless of weather or time-of-day pricing. Intermittent renewables — solar and wind — cannot meet this requirement without massive battery storage that does not yet exist at the required scale. Second, the power must be delivered at extremely high reliability: a 30-minute outage during an AI training run can corrupt weeks of work and cost millions of dollars. Third, the sheer concentration of demand — a single data centre campus drawing 500 megawatts in a county-sized area — requires grid upgrades that take years to permit and build.

These constraints are pushing hyperscalers toward unconventional solutions: direct power purchase agreements with nuclear plants, on-site gas turbines, fuel cells, and even interest in small modular reactors. Each solution creates a different set of investment opportunities.

The Nuclear Renaissance: Constellation Energy (CEG) and Vistra Corp (VST)

Nuclear power is the only large-scale source of carbon-free electricity that runs at full capacity regardless of weather conditions. For hyperscalers navigating both power reliability requirements and sustainability commitments, nuclear is the obvious answer — and the bidding war for nuclear capacity is already reshaping the energy sector.

Constellation Energy (CEG) operates the largest fleet of nuclear power plants in the United States, with 21 reactors generating approximately 10 percent of all US clean electricity. The company's pivotal moment came in September 2023 when it signed a 20-year power purchase agreement with Microsoft to restart the Three Mile Island Unit 1 reactor in Pennsylvania — dormant since 2019 — specifically to power Microsoft's AI data centres. The plant, rebranded Crane Clean Energy Center, is scheduled to restart in 2025. The deal demonstrated that hyperscalers would pay a substantial premium for long-duration, firm clean power, and it triggered a wave of similar announcements across the nuclear industry.

Constellation followed with agreements covering capacity from its other plants and announced plans to invest $1 billion to extend the operating life of its existing reactor fleet by 20 years. The company generated $3.9 billion in adjusted operating income in 2025 and guided to accelerating earnings growth through 2030 as its fixed-price power agreements roll over at materially higher rates. At its 2026 investor day, management outlined a path to $10 per share in operating earnings by 2030 — more than double its 2024 baseline. The stock has compounded at over 60 percent annually since its spin-off from Exelon in 2022.

Vistra Corp (VST) is the largest competitive electricity generator in the United States by installed capacity, with a fleet spanning nuclear, natural gas, coal, solar, and battery storage. Vistra's nuclear portfolio includes the Comanche Peak plant in Texas and, following its 2024 acquisition of Energy Harbor, plants in Ohio, Pennsylvania, and New Jersey. The Energy Harbor deal added approximately 4 gigawatts of nuclear capacity, transforming Vistra into a genuine nuclear operator rather than primarily a gas fleet.

Texas is Vistra's home market, and Texas is experiencing among the most acute AI-driven power demand pressures in the nation. The ERCOT grid — Texas's isolated electricity market — has seen wholesale electricity prices spike repeatedly as data centre load growth outpaces new supply. Vistra's flexible generation fleet, capable of ramping from minimal output to full capacity in minutes, commands premium pricing in this environment. The company reported $5.4 billion in adjusted EBITDA in 2025 and has returned over $4 billion to shareholders through buybacks and dividends since 2020. The stock was one of the S&P 500's best performers in 2024, up over 260 percent.

Grid Equipment: GE Vernova (GEV) and Eaton Corp (ETN)

Power does not travel directly from a generator to a data centre. It passes through a complex infrastructure layer of transformers, switchgear, substations, transmission lines, and distribution equipment. This layer — the grid itself — is the second major bottleneck in the AI power supply chain, and the companies that make and install grid equipment are facing the strongest order books in their histories.

GE Vernova (GEV) was spun off from General Electric in April 2024 and immediately became one of the most strategically important companies in the AI infrastructure ecosystem. GE Vernova manufactures gas turbines, wind turbines, grid automation equipment, and high-voltage power transformers — essentially every major component required to generate and deliver electricity at scale. Its gas turbine business is experiencing a demand surge unlike anything in its history, driven by the need for fast-dispatchable power to firm up AI data centre loads.

The company reported a backlog of over $115 billion as of early 2026 — representing approximately four years of revenue — and has been raising prices on new orders by 15 to 20 percent annually as demand overwhelms supply. Lead times for large power transformers, which take 18 to 36 months to manufacture and are critical for every grid connection, have extended to record lengths. GE Vernova's gas power segment has been profitable for decades, and the wind segment, which struggled for years, is now approaching breakeven on improved pricing and project selectivity. Management guided to high-single-digit organic revenue growth through 2028 with expanding margins as the backlog converts to revenue.

Eaton Corporation (ETN) is the less glamorous but equally essential player in the data centre power chain. Eaton makes the electrical infrastructure that lives inside data centres: switchgear that routes power safely, circuit breakers that protect equipment, busway distribution systems, and uninterruptible power supplies that bridge the gap between grid power and backup generators. Every data centre, regardless of its primary power source, requires Eaton's products — and the company estimates it has approximately 30 to 35 percent market share in data centre electrical equipment globally.

Eaton's data centre business has grown to represent roughly 40 percent of total company revenue and has been compounding at over 20 percent annually since 2022. The company reported $25 billion in revenue in 2025 with data centre-related orders growing at twice that rate. Importantly, Eaton's installed base creates significant recurring revenue from maintenance, upgrade cycles, and software subscriptions for its intelligent power management platforms. The company has also expanded into EV charging infrastructure, which benefits from many of the same electrical expertise and utility relationships.

Inside the Data Centre: Vertiv Holdings (VRT)

Vertiv Holdings (VRT) occupies the most direct position in the data centre power value chain: it supplies the power and thermal management infrastructure that operates inside the four walls of the facility itself. Every data centre requires power conditioning equipment to convert grid power to the precise voltages and frequencies that servers require, uninterruptible power supplies, battery systems, and — increasingly — liquid cooling systems to manage the extraordinary heat output of modern AI chips.

This last capability has become Vertiv's most important growth driver. Air cooling, the standard approach for conventional server racks, cannot handle the heat density of AI GPU clusters. A rack of H100s generates 10 to 40 kilowatts of heat — three to ten times what air cooling can effectively manage. Liquid cooling, which circulates coolant directly to chip surfaces, is becoming a requirement for every AI data centre, and Vertiv is one of only two global companies with the engineering capability and manufacturing scale to supply these systems.

Vertiv reported $8.3 billion in revenue in 2025, up 36 percent year-over-year, with operating margins expanding from 8 percent to 13 percent as scale benefits and pricing power flowed through. The company entered 2026 with a $7.5 billion backlog and raised guidance multiple times through the year. Management has outlined a path to $15 billion in revenue by 2028 and 20 percent operating margins, driven almost entirely by liquid cooling adoption in AI facilities. At current demand trajectories, they may be conservative.

On-Site Power: Bloom Energy (BE) and the Gas Turbine Builders

Given the grid queue backlogs in key data centre markets, some hyperscalers are bypassing the grid entirely for portions of their power load, generating electricity on-site instead. This has created an unexpected opportunity for companies that supply distributed power generation equipment.

Bloom Energy (BE) manufactures solid oxide fuel cells that convert natural gas or hydrogen into electricity with 65 percent efficiency — roughly double the efficiency of a conventional gas turbine — while producing far lower emissions and no combustion byproducts. Fuel cells can be sited at a data centre campus, generate power without connecting to the transmission grid, and scale incrementally as load grows. For a hyperscaler facing a five-year grid interconnection queue, on-site fuel cells provide immediate power availability at a price that, while higher than grid electricity in normal markets, is justifiable given the alternative cost of delayed operations.

Bloom signed a landmark agreement with American Electric Power in 2025 to deploy 1 gigawatt of fuel cells for data centre customers — the largest fuel cell order in history. The company has also announced agreements with Korea's SK Group to supply fuel cells for AI campuses across Asia. Revenue has grown from $972 million in 2023 to an estimated $2.1 billion in 2026, and the company achieved its first full-year profitability in 2025. The longer-term thesis rests on hydrogen: Bloom's fuel cells are designed to run on hydrogen as well as natural gas, positioning the company to benefit as green hydrogen production scales.

On the gas turbine side, beyond GE Vernova, Siemens Energy and Baker Hughes are benefiting from the same demand surge. Baker Hughes, primarily known as an oilfield services company, has a significant gas technology segment that manufactures turbines specifically designed for distributed power generation adjacent to industrial facilities. Its order book hit a 15-year high in late 2025.

Grid Construction: Quanta Services (PWR)

Quanta Services (PWR) is the largest electrical grid construction company in North America — the engineering, procurement, and construction firm that physically builds the transmission lines, substations, and distribution infrastructure required to connect new power sources and demand centres to the grid. In a normal capital cycle, Quanta benefits from gradual utility investment in grid maintenance and expansion. In the current cycle, it is experiencing something categorically different.

The combination of AI data centre load growth, renewable energy interconnections, and the electrification of transportation and industry is creating the largest grid construction wave since the post-World War II rural electrification programmes. Quanta reported $22 billion in revenue in 2025 and entered 2026 with a backlog exceeding $30 billion — the largest in its history. The company is constrained primarily by the availability of skilled electrical workers, not by demand. It has responded by acquiring smaller specialty contractors and investing in training programmes.

Quanta's competitive moat is difficult to replicate: it has decades of relationships with every major US utility, a fleet of specialised heavy equipment, and a workforce with the certifications required to work on high-voltage transmission infrastructure. New entrants face barriers that take years to overcome. The company has also been expanding into renewable energy construction and EV charging infrastructure, diversifying its exposure while remaining anchored to the core grid business.

The Small Modular Reactor Wildcard

The technology with the highest potential impact — and highest uncertainty — is the small modular reactor, or SMR. Conventional nuclear plants cost $10 to $15 billion and take 10 to 15 years to build. SMRs, which generate 50 to 300 megawatts in factory-built, modular units, promise to compress costs to $1 to $3 billion per unit and timelines to 5 to 7 years. If the technology proves out at commercial scale, SMRs could become the preferred power source for co-located hyperscale campuses.

Several hyperscalers are already placing bets. Amazon invested in X-energy for SMR deployment. Google announced agreements with Kairos Power. Microsoft has made its intentions clear with the Three Mile Island revival and has publicly stated interest in SMR procurement as the technology matures.

Oklo (OKLO), backed by OpenAI CEO Sam Altman, is pursuing a next-generation fission design using a fast reactor architecture that can run on recycled nuclear fuel. The company went public via SPAC in 2024 and carries all the characteristics of an early-stage speculative bet: no commercial reactor yet deployed, but management talent, a differentiated technology approach, and an extraordinary network of potential customers. It is not a core holding for risk-conscious investors, but it represents the type of optionality the AI power transition creates.

Mapping the Investment Landscape

The AI power bottleneck creates a tiered investment opportunity across three risk categories:

Near-term, high-visibility cash flows are concentrated in companies with large backlogs and existing infrastructure. Constellation Energy and Vistra benefit from locking in nuclear power at premium prices through long-term contracts with investment-grade counterparties. Eaton and GE Vernova are shipping equipment today against backlogs that extend into 2028. Quanta Services is building transmission infrastructure with customers who have already committed capital. These companies do not require the AI thesis to play out — the orders are already on the books.

Mid-term scaling plays include Vertiv, which must continue executing on its liquid cooling ramp, and Bloom Energy, which is transitioning from early commercial to volume deployment. Both carry more execution risk than the first group but also more upside if their revenue forecasts prove achievable.

Long-duration, speculative optionality sits with the SMR developers. These companies could become central to the AI power supply chain by 2030 if they achieve commercial deployment, but the regulatory and engineering uncertainties are substantial.

Risks Worth Tracking

The bear case for this theme rests on two scenarios. First, a pause or retrenchment in AI capital expenditure: if hyperscalers reduce their infrastructure spending — whether because AI revenue growth disappoints, because energy costs rise faster than expected, or because a competing technology reduces compute requirements — the demand underpinning this analysis weakens materially. Second, grid and permitting reform that dramatically accelerates interconnection timelines could reduce the premium that hyperscalers currently pay for firm, on-site, or nuclear power.

Neither scenario appears likely in the near term. AI infrastructure spending commitments made in 2025 and 2026 are multi-year in nature and largely irreversible — you cannot un-order a transformer or un-sign a nuclear power purchase agreement. But investors should monitor hyperscaler capital expenditure guidance closely for any signals of retrenchment.

Conclusion: Power Is the Moat

The energy bottleneck is not a temporary friction that the AI industry will engineer around. It is a structural constraint rooted in physics, geography, and decades of underinvestment in the US power grid. Resolving it will require hundreds of billions of dollars of capital investment over the next decade — capital that is already flowing to the companies profiled here.

The most durable investment opportunity belongs to the companies that cannot be easily replicated: nuclear operators with licensed, operating plants (Constellation, Vistra); grid equipment manufacturers with multi-year backlogs and skilled workforces (GE Vernova, Eaton); and the sole large-scale builder of US transmission infrastructure (Quanta Services). These companies have physical assets and regulated relationships that no amount of software or capital can shortcut. In a world where the defining constraint on AI is electricity, that is a moat worth owning.

This article is for informational purposes only and does not constitute investment advice. Always conduct your own due diligence before making investment decisions.