Neel Somani, a former quantitative researcher who covered power and gas at a major hedge fund, has published Power 2026, a book length primer arguing that electricity, not chips, is now the constraint on AI capacity. Somani states that data centers already draw roughly 5 percent of US power consumption, and that demand from data centers is on a path to double roughly every two years, a pace that would in theory outrun total US generation by the mid 2030s. The book is not news and not a study. It is a teaching text aimed at founders, investors, and traders, and its value to an AI industry reader is that almost nobody covering AI compute deals can actually read a power market.
That gap shows up constantly in how the industry gets covered. Nine-figure cloud commitments, multi-gigawatt data center campuses, and a $9.1 billion agreement between a hyperscaler and a bitcoin miner for compute capacity tend to get reported as chip stories or real estate stories. Somani’s framing treats them as power procurement stories first. In his account, the price of electricity alone can decide whether a proposed campus ever gets built.
The concept worth learning is marginal cost pricing. In a competitive commodity market, every unit sold clears at the price it costs to produce the single most expensive unit still needed to satisfy demand. Somani calls that the marginal unit. If a grid needs 100 units of power and the cheapest 99 come from low-cost generators while the last unit requires a pricier plant, every generator on the grid gets paid that higher clearing price, not just the expensive one. On most US grids the marginal unit is a natural gas plant, so its fuel cost and its heat rate, the amount of fuel it burns per megawatt hour, tend to set the price for an entire region during a given hour.
Somani’s second concept is the independent system operator, or ISO, the nonprofit authority that runs this pricing exercise continuously. An ISO collects bids from utilities and offers from power plant owners, computes the clearing price at each location on the grid, and dispatches generators to hold the system in balance. Skip that coordination and mismatched supply and demand pushes grid frequency away from 60 hertz, which risks damaging generating equipment that costs upward of $100 million per unit. PJM, the mid-Atlantic market that covers the Virginia data center corridor, along with MISO, CAISO, and the Texas grid operator ERCOT, each run their own version of this optimization with different rules for reliability and capacity.
Where markets diverge is how they pay generators to simply exist. Most US markets run a capacity market, a standing payment that keeps backup plants available even when they rarely run. Texas and Alberta reject that model for what Somani calls an energy-only market: generators earn nothing beyond what they clear for power actually produced, with prices allowed to spike toward a cap, $5,000 per megawatt hour in Texas, during scarcity to cover fixed costs that marginal-cost pricing alone would not.
For a plant owner or a data center operator hedging exposure, the standard tool is the spread trade. A spark spread pairs a long position on power with a short position on natural gas, scaled by a plant’s heat rate, so the trade tracks a plant’s actual margin rather than either commodity alone. Somani cites Anthropic’s 400-megawatt, 20-year lease with TeraWulf, the bitcoin miner turned data center host, implying a price near $271 per megawatt hour, as the scale of commitment developers need to underwrite nine-figure construction loans.
Anyone negotiating the next power purchase agreement for a data center campus should be pricing it the way a spread trader would: against the marginal generator likely to set the local clearing price, not against the headline number in the press release.
This analysis draws on “Power 2026: Electricity Pricing in the Age of AI,” a primer by Neel Somani published at power2026.ai.
