The most consequential infrastructure story of this decade is not semiconductors. It is not data center real estate. It is power. Neel Somani, a former quantitative researcher at a major hedge fund who covered power and gas markets, makes this case with precision and authority in Power 2026: Electricity Pricing in the Age of AI — a primer that deserves serious attention from anyone with capital exposure to the AI build-out.
Somani's opening observation is unambiguous. "AI demand is exploding," he writes. "But while there's been plenty of discussion on the GPU and memory supply chains, the real constraint to expanding AI capacity is power." Data centers already account for roughly 5% of US power consumption, with demand doubling every two years. On an unconstrained trajectory, that demand would outpace total US power generation by the mid-2030s. The constraint, in other words, is not hypothetical.
How Power Is Actually Priced
To understand the opportunity, one must understand the mechanism. Somani constructs the argument from fundamentals. Power pricing is governed by marginal cost: the price at which every unit clears is determined by the most expensive generator required to meet demand at that moment. "In a competitive market," Somani explains, "the market price that everyone sells at is exactly equal to that marginal cost." This is not a quirk of regulation. It is a structural feature of how the grid works, enforced by the independent system operator (ISO), the central nonprofit authority that balances supply and demand across thousands of physical nodes in real time.
The implications are significant. Natural gas generators are frequently the marginal unit setting the clearing price. Their efficiency, measured by "heat rate," determines how much fuel is consumed per megawatt-hour produced. Combined cycle gas turbines operate at roughly 6 to 7 heat rate; simple cycle peakers are far less efficient and far more expensive. "In general, operators use their CCGT first if they can," Somani notes, "and only if the power price is high enough, then they'll turn on their peakers." Price spikes follow directly from this logic whenever demand surges.
Building the Infrastructure
The path from concept to operating power plant is where most investors underestimate the complexity. Somani walks through site selection, permitting, financing, construction, and operation in granular detail. The conclusion is sobering: "Site selection is so constraining that you don't always have tons of options. You decide on the type of plant you want to build, then pick the site, not the other way around."
Financing a major plant, typically $300 million or more, requires contracted cash flows. Power purchase agreements (PPAs), heat rate call options (HRCOs), and long-term supply agreements with anchor tenants are the instruments that make construction debt serviceable. Somani illustrates this with a real-world reference: Anthropic's $19 billion lease with TeraWulf, 400 MW over 20 years starting in late 2027, implies roughly $271/MWh with no GPUs included. SpaceX's deal with Reflection, structured with a 90-day out for either party, implies approximately $5,000/MWh for power that includes GPUs ready to operate.
The Homer City case study crystallizes the opportunity. A former 2 GW coal plant in Pennsylvania is being redeveloped into a 4.4 GW natural gas facility, at an estimated cost of $10 billion, intended to anchor a major data center campus. With an air quality permit approved by the Pennsylvania DEP in November 2025 and 1,000 workers actively on site, it represents the kind of large-scale, capital-intensive redevelopment that AI demand is now making economically viable.
The Procurement Problem
Construction itself faces a materials crunch. "Major turbine manufacturers like GE Vernova, Siemens Energy, and Mitsubishi Power are completely out of stock," Somani writes, "with the backlog months to years out." The response has been improvisation: jet turbines repurposed as gas turbines, parts imported from China. The constraint is real and near-term.
Market Structure Divergences
Somani's treatment of regional power markets is one of the primer's most instructive sections. PJM remains the most liquid market in the US. ERCOT in Texas is famous for negative power prices driven by excess wind generation and a price cap of $5,000/MWh during scarcity events. CAISO in California faces the "duck curve," where aggressive solar growth pushes daytime prices toward zero while evening demand requires expensive peaker units to come online. Alberta, profiled as a detailed case study, combines an energy-only market structure with rapid renewables growth, creating extreme price volatility between daytime and evening hours that batteries are only beginning to flatten.
These divergences are not cosmetic. They determine which regions can economically support new data center development, which fuel sources will be dispatched, and where transmission congestion will concentrate the price risk.
An Unsolicited Recommendation
Somani does not shy away from a pointed policy view. "My take is that we need to temporarily relax the resistance to inefficient, dirtier sources of fuel like coal or even diesel, as we build startup-speed development capabilities within the US." He adds that "delivery and transmission are the true constraints in many regions," citing the example of cheap upstate New York nuclear power that cannot reach Manhattan due to insufficient transmission capacity. Expedited permitting at the state and federal level, he argues, is not optional.
Five Key Takeaways for Advisors and Investors
- Power, not GPUs, is the binding constraint. Capital flowing into AI infrastructure needs to account for electricity procurement as a primary risk factor, not an afterthought.
- Long-term contracted cash flows are the currency of project finance. PPAs and similar structures are what unlock 9-figure construction debt. Anchor tenant quality determines borrowing costs.
- Regional market structure matters enormously. ERCOT, PJM, CAISO, and MISO behave differently. Transmission congestion, fuel mix, and ISO rules are not interchangeable.
- The procurement supply chain for turbines is broken. Lead times are measured in years. Investors in new generation assets must underwrite this constraint explicitly.
- Near-term demand is the real challenge. Major AI labs may be comfortable with their power position by 2028. The demand is now, and 6 to 12 months of contracted revenue is not sufficient to underwrite a new plant.
Footnote:
1 Somani, Neel. Power 2026: Electricity Pricing in the Age of AI. 2026, https://power2026.ai/.
