Power 2026

These are distilled notes from Neel Somani’s power2026.ai primer — go to the original for the full version. Somani traded power and gas as a quant at Citadel; the market mechanics below are his framework, restated tightly. Where a chapter ties that framework to a specific company, filing, or project DeadRisk already covers — Homer City, TeraWulf’s lease, Vistra’s, Situational Awareness LP’s 13F — that connection is this page’s own addition, not his.

The AI-capex loop gets argued in chips and dollars. Somani’s point is that the binding constraint sits one layer down, in a market almost nobody in tech reads: US data centers already draw roughly 5% of the country’s electricity, and the projection he leans on has that draw doubling roughly every two years. The Power Wall tracks the permitting and political side of that fact — the moratoriums, the grid queues, the protests. This primer is the other half: once you have permission, how the electricity actually gets priced.

Every power market, everywhere, obeys the same identity at every location, every instant: Supply = Demand + Net Exports + Change in Storage. Price is whatever makes that hold. There is no other physics underneath it.

In a competitive market every generator running at a given node gets paid the same clearing price there — set by the marginal unit, the most expensive one still needed to meet demand at that node. Efficiency is measured by heat rate, in MMBtu per MWh: combined-cycle gas plants run about 6, coal about 10, and most simple-cycle peakers sit close to coal — only the newest aeroderivative units get down near 8–9. A peaker beats coal on fuel price and the ability to start in minutes, not on raw efficiency. Constellation Energy’s restart of Three Mile Island is the same arithmetic from the buyer’s side: a nuclear plant that was economic to revive specifically because the alternative was paying gas’s marginal price for twenty years.

Somani lays out five steps: scope the plant, fight for the site (sage-grouse habitat, mining claims and oil-and-gas rights all show up on the same GIS layer), raise debt — nine figures for a typical project, ten for something Homer City’s size — against a long-term power purchase agreement, build it, then refinance once it’s operating and de-risked.

The step everyone assumes is hardest — financing — isn’t, anymore. The real constraint is turbines. Combined-cycle gas turbines have three manufacturers that matter for a US utility-scale project worldwide, and their order books run years out. GE Vernova is one of the three. That’s a textbook cornered resource — value accrues to whoever owns the choke point, not whoever has the most capital.

The clearest single illustration is Homer City, Pennsylvania: a dying 2 GW coal plant being rebuilt as a 4.4 GW gas plant to serve data centers, at roughly $10 billion, with 1,000 workers on site as of its most recent construction milestone. It needed new air-quality and waterway permits on top of the capital — this is what “AI needs power” costs as a construction project, not a headline.

Homer City works partly because the interconnection already exists — a former coal plant’s grid connection, reused rather than built from scratch. That’s the inverse of Bloom Energy’s bet: behind-the-meter generation exists precisely for sites where a new interconnection is not there and won’t be in time.

Grid operators — ISOs — don’t set prices. They solve an optimization problem: minimize total system cost, subject to supply equaling demand at every node, transmission-line ratings, generator limits, and reliability reserves. The price at any one location, the locational marginal price, falls straight out of that math. It’s literally a Lagrange multiplier on the balance constraint. A typical US node prices somewhere in the $20–60/MWh range most hours — the baseline everything later in this primer departs from.

Two markets clear the same power. Day-ahead clears 24 hours before delivery, on a forecast; real-time updates every five minutes, on what actually happened. Generators commit to the day-ahead number, and the gap between forecast and reality settles at the real-time price. Almost all of a power market’s volatility lives in that gap.

PJM is the most liquid US market and covers more territory than its name suggests; gas now sets the margin coal used to set. Vistra generates directly inside both PJM and ERCOT — one of the few merchant operators exposed to two of these five personalities in a single portfolio.

MISO is retiring coal into a wind-heavy resource mix. CAISO has so much solar that midday prices collapse and evening prices spike when the sun drops — the “duck curve,” the same shape battery buildouts everywhere are now built to flatten.

ERCOT runs energy-only, with no capacity payments, and prices go negative when wind oversupplies. Since its late-2025 real-time market redesign it splits the old single cap in two: roughly $5,000/MWh in the day-ahead market, roughly $2,000/MWh in real time. Alberta runs the same energy-only design, capped near C$1,000/MWh today (legislated to rise), and its own duck curve is now visibly flattening as batteries come online — a preview of where ERCOT is headed.

Every instrument in this market is a bet on one of the constraints above. Forwards lock a future price. Spark spreads — buy power, sell gas — replicate owning a gas plant’s economics without owning the plant; a 7-heat-rate spark spread hedges a 7-heat-rate CCGT exactly.

FTRs and basis trades bet on congestion between grid nodes when a transmission line hits its rated limit — a purely financial claim on the same physical bottleneck The Power Wall documents from the permitting side.

The move that makes data-center debt investable: lock a hyperscaler or AI lab into a decades-long power purchase agreement as an anchor tenant, then borrow against that contract instead of against the merchant power market. TeraWulf’s $19 billion, 20-year lease with Anthropic — 401 MW in TeraWulf’s own disclosure — prices out (401 MW × 8,760 hours × 20 years ≈ 70.3 million MWh) to an implied $271 per contracted megawatt-hour. That’s several times a typical wholesale LMP from Chapter 4, because it’s paying for guaranteed data-center capacity — shell, cooling, delivery — not raw electricity; the premium is what a large, well-capitalized tenant’s contracted credit buys over a merchant power contract.

The other end of the duration curve is SpaceX’s compute deal with Reflection AI: up to $6.3 billion for Nvidia GB300s at Colossus 2, $150 million a month through 2029 — with a 90-day out for either party. Somani’s napkin math prices it around $5,000/MWh. That’s power with GPUs ready to go, and the 90-day out is exactly what makes it impossible to underwrite a plant against. Roughly eighteen times TeraWulf’s $271 is not a price discrepancy; it is the market quoting duration and bundled compute as separate products. Note what the buyer did next: Reflection signed a $1 billion-plus follow-on with Nebius three weeks later — a tenant this short-dated shops, which is the point of paying for the out. The seller’s side of that trade sits inside SpaceX, whose xAI turbine fleet is the generation being resold.

Vistra ran a version of the same trade from the seller’s side: a 20-year agreement selling 2,609 MW of nuclear output directly to Meta — the anchor-tenant logic working through an existing generator instead of a developer’s new lease. Applied Digital, IREN, CoreWeave and Nebius run variations of the same structure from the miner-conversion and neocloud side.

Behind-the-meter compute skips the anchor-tenant trade entirely: generation built beside the GPUs, no interconnection queue to wait in, at the cost of owning land, cooling and battery risk a grid connection would otherwise have shared. That’s exactly Bloom Energy’s business.

One fund’s public filings show both sides of this thesis priced into real capital. Situational Awareness LP — named for the essay Leopold Aschenbrenner wrote before the fund existed — disclosed Bloom Energy as its single largest long position and SanDisk as its second-largest in the 13F for the quarter ended 2026-03-31: the long side of exactly the power-and-memory chokepoints this primer describes. The same filing carried put options — a bearish or hedging position, not a long bet — on Intel, AMD and Oracle. A 13F shows a snapshot, not a strategy, but the snapshot is public: long the physical chokepoints, short or hedged on the compute layer sitting above them.

Turbines and transmission are both binding constraints, at different stages. Turbines gate how fast new supply can be built — a multi-year order book, no matter how much capital shows up. Transmission gates where any supply, new or already built, can actually clear: Upstate New York has cheap nuclear power right now, and Manhattan still can’t get much of it, because the wires aren’t big enough. That’s a different bottleneck, downstream of the first one, not a competing answer to the same question.

Underneath both constraints sits a mismatch of clocks. Somani reports the major labs are comfortable with their power position a couple of years out, around 2028; the desperate demand is now — and six to twelve months of demand cannot underwrite a nine-figure construction loan that needs years of contracted cash flow. That mismatch, more than any single bottleneck, is why capital keeps routing to what already exists: Constellation’s relicensed reactor, TeraWulf’s reused coal interconnection, 90-day-out compute resales. New supply answers the 2028 question. Only existing supply answers the 2026 one.

Put the whole thing together and the same megawatt-hour can be worth less than zero and $5,000 within the same day — physical constraints (transmission limits, generator startup costs, renewable intermittency) colliding with market rules (marginal-cost pricing, energy-only design versus capacity markets). Which constraint binds, where, and when is most of what there is to understand about pricing the AI buildout.