Power 2026

⚠️ Correction to my own first assessment

On a first pass (intro page only) this was graded "mostly qualitative, a primer, thin on figures." That was wrong and the fuller read reverses it. The intro undersells the book: it carries real market mechanics, named deals with numbers, the pricing formulas, and 41 footnotes. Recorded because the mistake is instructive — grading a multi-chapter source from its landing page is how a good source gets rejected.

Why this sits on datacenters

This rung had zero papers. The rung's question is "where it runs, and what stops it running (power included, in full)" — and this is a practitioner treatment of exactly the stopping constraint, written by someone whose seat was trading power and gas.

Structure

Part 1 — All About Power Plants: Preface/Motivation · Fundamentals of Commodities Pricing · The Power Plant · To Build A Power Plant · Case Study: Homer City · Meeting the Growing Demand

Part 2 — How To Trade Power: Power Markets in the United States · Case Study: Alberta · The Production Cost Model · Practical Approximations for Power Pricing · Types of Power Trades

The two anchor numbers

The most useful thing here for MenFem is the spread between power-only and power-with-GPUs:

Deal	Structure	Implied
Anthropic ← TeraWulf	$19B lease, 400 MW over 20 years, starting H2 2027	~$271/MWh, no GPUs — but includes building and cooling
SpaceX → Reflection	90-day out for either party	~$5000/MWh, power that comes with GPUs ready to go

Both are the author's own derivations — the SpaceX figure is explicitly labelled "napkin math," the Anthropic one "implied revenue." Neither is a market print and neither should be cited as one. What they are good for is the order-of-magnitude gap between selling electricity and selling compute, which is the datacenters → inference-economics bridge.

Market mechanics worth having

• DA vs RT — "the 'day-ahead' (DA) market which is computed the day before power is delivered, and a 'real-time' (RT) market which is recomputed every 5 minutes on the day of."
• LMP, precisely — "The Lagrange multiplier of each location's balance constraint is exactly the marginal cost for providing an additional MWh of power at that location. It comes for free when you solve the optimization problem using Gurobi or CPLEX." This is the definition most AI-and-power coverage never gets right.
• Capacity — California uses resource-adequacy requirements (utilities like PG&E negotiate with generators to keep them alive); PJM runs an ISO-operated capacity market.
• Interconnection queue — "the line of people waiting to construct their generators at various locations. The ISO doesn't let them all construct at once."
• Behind-the-meter — consuming on-site generation can "simplify or avoid the generator interconnection process, but you need the space to put all of those GPU racks, cooling infrastructure, and batteries to balance your generation with your consumption." The BTM trade-off is queue-avoidance bought with siting and balancing cost.
• Datacenters as ancillary services — flagged as an opportunity, not an established practice.

Instruments and how they price

• PPA — an anchor tenant (hyperscaler or AI lab) commits to $/MWh for a term; "that future stream of cash flows is easier to underwrite against."
• Heat Rate Call Option (HRCO) — "designed to mimic the economics of your power plant. If your power plant makes $50K one day, then in theory, exercising the HRCO should pay $50K."
• Forward — financially settled against the eventual spot DA price.
• Spark spread — Power Price − (Heat Rate × Natural Gas Price). Dark spread is the same against coal.
• Financing — for an established developer with a prime hyperscaler tenant, "a reasonable ballpark is ~2.25% above SOFR."

Homer City case study

2 GW of coal (three units) being replaced by a 4.4 GW natural gas plant at ~$10B, with 1000 people actively working on construction and a heat rate of "around 6 … considered very good." PA DEP air-quality permit approved November 2025. In PJM.

Demand

• "Data centers already account for ~5% of US power consumption" — footnote to a US DOE report
• "data center power demand doubling every two years" — footnote to Goldman Sachs
• If sustained, "demand would in theory outpace total US power generation by the mid-2030s" — the author frames this as a theoretical extrapolation, not a forecast, and it should be repeated that way

Sourcing pattern

41 numbered footnotes. The pattern is consistent and worth knowing before citing:

• Physical plant specs and market structure → cite authoritative sources (EIA, ISO technical manuals, FERC). Solid.
• Headline demand figures → footnoted (DOE, Goldman). Solid.
• Deal economics → author synthesis and derivation, labelled as such. Use for shape, not as a print.

Evidence grade

Good for mechanism, careful-but-derived for deal pricing, properly sourced for demand. Not a paper and not a dataset — a practitioner primer. Its value to this rung is that it makes the power constraint legible well enough to reason about, and it supplies two implied-price anchors nothing else in the KB has. The datacenters rung still has zero papers after this; a primer is a foundation, not a substitute.