Power Availability & Power-Cost Structure

Chapter 3.2 treated interconnection as a yes/no gate against a schedule: can you get power to this site, and when. Once the answer is yes, the economics turn on what that power actually costs, in what structure, and who bears the cost of the grid you are about to stress. The two questions are separable on purpose. A site can clear the speed-to-power gate and still be uneconomic, because the megawatts arrive at a price, a shape, and a set of curtailment and cost-allocation terms that the GPU fleet cannot earn against.

In Epoch AI’s 2026 1 GW model, energy is the largest operating-cost category; the share depends on what the model includes as opex, especially GPUs, facility cost and financing. Silicon dominates capex; electricity dominates opex. The power-cost structure is therefore the dominant TCO lever still available after the building exists — you cannot re-spec the chips, but with enough foresight at siting you can change the delivered cost of energy by a factor of two or three. What follows works through each fork in the power-cost stack and the dollars that ride on getting it wrong.

Before any tariff analysis, a site faces a physical screen that has nothing to do with price and everything to do with what steel is already in the ground. Three variables decide whether a site is powerable on a reasonable schedule at all: transmission proximity, voltage class, and substation headroom. Get these wrong and the cheapest energy market in the country is irrelevant, because the required lines, rights-of-way, equipment and approvals may dominate the project schedule and cost before a single electron arrives.

Transmission proximity is the first cut. A gigawatt-class campus needs a high-voltage transmission tap — typically 230 kV or 345 kV, and increasingly 500/765 kV as backbone projects (ERCOT's statewide ~2,500-mile, ~$33B 765 kV expansion, PJM and SPP build-outs) reach the load. A site adjacent to an existing high-voltage line with spare thermal capacity is a fundamentally different proposition than one that needs miles of new line, every mile of which carries its own permitting, right-of-way acquisition, and 100+-week conductor and structure lead times. The distance to a viable tap is, in practice, a schedule variable, not a geography variable.

Voltage class sets both the equipment you must buy and the losses you will eat. The higher the delivery voltage, the more power you can move per ampere and the lower the I²R losses, but the more expensive and longer-lead the step-down transformers and the larger the customer substation footprint. A 1 GW load taken at 345 kV is an entirely different electrical interface — and a different substation cost — than 200 MW taken at 138 kV. This is an irreversible decision: the voltage class is baked into the substation, the protection scheme, and the interconnection agreement, and you cannot re-pick it mid-life without rebuilding the interface.

Substation headroom is a variable that can change the energization path materially. A utility substation with spare transformer capacity and an open bay can absorb a new large load by adding a feeder; one that is already at its firm rating requires a new transformer bank (~128–208 week lead time; up to ~60 months in constrained markets) or an entirely new substation. The highest-value pre-diligence a developer can do is to map, for every candidate site, the nameplate and firm rating of the serving substation, its current loading, and the queue of other large loads ahead of you for the same headroom. Headroom is a depletable shared resource, and you are rarely the only one shopping for it.

Operators routinely shop on a single number — a delivered ¢/kWh — but that number averages together components that behave very differently under an AI load shape. A defensible power-cost model decomposes the headline rate into four parts, each with its own driver, its own volatility, and its own hedge:

• Energy (the commodity). In organized markets, bus Locational Marginal Price (LMP) is the dispatch primitive; the price payable by load may instead be a load-zone settlement price, zonal LMP, utility tariff, or negotiated retail product. The applicable energy price varies by location, interval, market, and contract. The same MWh can cost a few dollars at a wind-rich West Texas node and a hundred-plus dollars at a congested load center in the same interval.
• Congestion / basis. LMP decomposes into a system energy price plus a congestion component plus a small loss component. Congestion is the value of transmission scarcity between generation and your node, and it is the most under-modeled line in a merchant power plan. A node downstream of a transmission constraint can settle persistently above the hub; a node co-located with stranded generation can settle below it, occasionally negative.
• Capacity. In capacity markets (PJM, ISO-NE, MISO) you pay a separate charge for the resource adequacy your load obligates the system to procure. PJM's capacity auction clearing prices spiked roughly an order of magnitude in 2025–2026 as data-center load tightened the supply-demand balance, turning capacity from a rounding error into a material line item.
• Demand charges. Regulated retail tariffs bill the largest line not on energy consumed but on billing demand: demand-charge rate ($/kW-month) × billing demand (kW), often with a ratchet that holds the charge near the annual peak for months. For a high-load-factor data center this is usually favorable — a flat 90%+ load factor amortizes the demand charge over enormous consumption — but for a spiky or curtailed load it can dominate.

The decomposition matters because two sites with the same delivered ¢/kWh can have completely different risk profiles. One may be 90% energy and 10% congestion (a merchant-exposed, basis-risk-heavy site); the other may be 50% capacity and demand charges (a regulated, schedule-stable site with little hourly exposure but a high floor). Hedging starts from this decomposition — you cannot buy an instrument against a component you have not priced. → PPA structures, basis risk, and the hedging toolkit (CRRs/FTRs, heat-rate hedges) live in Chapter 3.4.

The main reason an AI data center is a good power customer — and the reason it can sometimes secure rates a flickering industrial load cannot — is its load factor. A facility running training or steady inference at a flat 80–95% of contract demand around the clock is the dream customer for a utility recovering the fixed cost of a substation and a line: it amortizes those fixed charges over the maximum possible consumption, driving the effective ¢/kWh of the delivery and demand components down toward their floor. That amortization lets a high-load-factor data center clear an economic deal on a tariff that would crush a peaky factory. The corollary is a warning: anything that lowers your load factor — curtailment, demand-response events, training-job idle gaps, a half-full hall during ramp — raises your effective per-MWh cost on the fixed components, because the same demand and delivery charges now spread over fewer megawatt-hours. Curtailment costs you the energy you forgo and the fixed charges you keep paying on power you are not drawing.

Build the invoice before comparing its rate. Annual energy in MWh = flat MW × non-leap-year hours. Monthly billing demand = max(actual peak, contractual minimum). Energy and delivery cash each equal MWh × their respective prices; demand cash = billing MW × kW per MW × the kW-month price × billing months. Add fixed facilities cash, then divide the total by energy delivered. Use the unrounded energy and charges throughout; the displayed GWh and money are rounded only at the end.

Select the component contract at the base draw and the variable contract in the lower-draw case. While the demand minimum governs, the crossover is fixed annual demand-and-facilities cash divided by the difference between the variable offer and component energy-plus-delivery price; divide that MWh quantity by year hours to obtain average MW. Confirm the crossover lies below the minimum before using that branch. Curtailing energy does not erase a continuing demand or facilities charge.

AEP Ohio’s DCT explanation supports the distinction between billing-demand tests and energy consumed; these prices and floors are independently assumed. Obtain the operative tariff for a real site. Carry the ramp downside to 2.5 and price PPA exposures in 3.4.

In an organized market — ERCOT, PJM, MISO, SPP, CAISO, ISO-NE — energy does not have one price; dispatch has an LMP at every node. A bus LMP is a dispatch primitive, while a load's payable price follows its market and contract — for example, ERCOT typically settles wholesale load at a Load Zone Settlement Point Price, and a retail data center may instead face an REP or utility tariff. Site screening therefore compares the price at the load's actual settlement point or tariff, not an assumed bus LMP; congestion and losses still flow through the applicable pricing chain.

West Texas illustrates the exposure. ERCOT publishes prices at Resource Nodes, Hubs, and Load Zones; a VPPA settled at a resource node and a data center priced at a Load Zone SPP or retail tariff occupy different settlement points, so congestion changes the spread between them. This is locational basis risk: a hedge struck at a hub or resource node does not automatically protect load priced at a zone or tariff, so site screening and hedge design must use the project's actual settlement chain. → the full basis-risk and VPPA-settlement treatment, including negative-settlement stress tests, is in Chapter 3.4; the lender's view of that exposure in Chapter 2.5.

The power decision of the 2026 era has moved past "firm grid or not" to how much curtailment you will accept in exchange for queue speed and a lower energy charge. Every major RTO is building a faster interconnection lane for loads that agree to be curtailed: ERCOT's mandatory curtailment / 'kill switch' for large loads interconnecting after the end of 2025, SPP's price-responsive curtailment and non-firm tracks, PJM's non-capacity-backed and interim arrangements, and the FERC flexible-load study concept. Duke’s national estimate is conditional: Duke’s February 11, 2025 study modeled 98 GW of US grid headroom integratable at just 0.5% annual curtailment — about 44 full-load-equivalent hours a year. That is a national modeling result, not an inventory of deliverable capacity at your substation: a site-specific power-flow and stability assessment and an executable service offer are what turn it into megawatts you can energize.

Curtailment is therefore a priced lever, not a binary. The question is whether your workload's goodput economics survive the expected curtailment-hours. A batch-inference or checkpoint-tolerant training fleet can absorb a few hundred curtailment-hours a year by rescheduling work or riding through on storage, paying for the privilege of a faster, cheaper interconnection. An always-on inference business serving an SLA cannot — every curtailed hour is breached revenue, and the cheap non-firm energy is a false economy. Underwrite the cost of the expected curtailment-hours given your workload, netted against the queue-time and energy savings you are buying — the curtailment cap alone tells you neither. → the load-flexibility and grid-services revenue framing lives in Chapter 15.8; curtailable/non-firm interconnection mechanics in Chapter 3.2.

A gigawatt of new load does not arrive on the existing grid for free — it triggers transmission and distribution upgrades, sometimes hundreds of millions of dollars of them. The defining regulatory and political question of 2025–2026 is who pays: the data center that caused the upgrade, or the ratepayer base that did not. This was a footnote five years ago. It is now frequently the deciding variable on whether a project gets approved at all, because rising residential bills in data-center-heavy regions have made cost allocation a live political fight.

The mechanism through which the answer is set is the large-load tariff — a new customer class, now approved in 23+ states, written specifically to push the cost of the buildout onto the loads that cause it. The Oregon POWER Act template is representative and worth memorizing because variants of it are spreading — the Oregon PUC approved PGE’s Schedule 96 under it in May 2026 (Order 26-154, UM 2377): a 20 MW+ threshold defines the class; the large load pays 100% of the distribution-upgrade cost it triggers; minimum generation and transmission demand charges at 90% of contracted system capacity (a take-or-pay floor — you pay for 90% of what you contracted whether you draw it or not) protect the utility against stranded investment if you under-build or leave; 10–30-year contract terms scale with load size, reaching 30 years at 220 MW and above; and a 1¢/kWh surcharge applies above 100 MW of allocated system capacity. These are the PUC template figures (2025 statute; 2026 order) — verify the operative Schedule 96 / UM 2377 sheets and the project service agreement before entering any of them as a bill operand, and note that Pacific Power’s data-center tariff (UE 463) is a separate proceeding. ERCOT's SB6 regime is the Texas analogue, with a 75 MW large-load definition, mandatory curtailment, and a ≥$100,000 screening study fee. Virginia, the largest US market, layered on a flat $0.011/kWh data-center consumption tax effective July 2026.

The consequence for the model is direct and large: these allocation mechanisms can add anywhere from a fraction of a cent to several cents per kWh of effective cost, and the minimum-take floor converts a variable energy bill into a fixed obligation that behaves like debt. A take-or-pay floor means the curtailment savings discussed above are partially illusory — you keep paying the demand floor on power you are not drawing. The 'who pays' regime in your jurisdiction is a first-order input to the power-cost stack and, increasingly, to whether the project clears its social-license gate at all. → the regulatory mechanics and the federal-vs-state collision over large-load interconnection are owned by Chapter 3.2; the macro load-growth and cost-shift narrative in Chapter 16.1.

After the chips are bought, power is the largest controllable variable in the lifetime cost of the asset. Silicon dominates capex and is essentially fixed at procurement — you buy the GPUs at market and depreciate them on a contested schedule (→ Chapter 1.8). The building shell is a small and slow-moving line. But energy — the largest operating category in the cited Epoch model, with its stated denominator — is the cost you can still move by a factor of two or three through decisions made at siting: which node, which voltage class, which tariff, how much curtailment, behind-the-meter or grid.

The leverage is starkest netted against the revenue the asset earns. Against that topline (Chapter 1.8's revenue-per-GW figure), a mis-structured energy bill can run to many hundreds of millions a year. A 2¢/kWh difference in delivered power cost — entirely achievable between a well-sited, well-hedged location and a poorly-structured one — is on the order of $175M/yr on a 1 GW facility at 100% load factor. No engineering efficiency in the rest of this guide moves that much money, which is why the screen runs availability first, structure second, and only then everything else.

Sort the power-cost components by reversibility, the same way Chapter 1.1 sorted the scoping decisions. The irreversible components are fixed at contract: the voltage class and substation interface, the large-load tariff election and its take-or-pay floor, the contract tenor, and a fixed-price PPA's strike. The reversible components float and must be hedged or actively managed: the merchant energy and congestion exposure, the capacity-auction outcome, and the curtailment-hours you actually incur. A 15-year asset financed against a merchant energy curve is a far riskier instrument than one behind a matched-tenor fixed-price supply; which of the two you are is decided here, and the lenders in Chapter 2.5 will price the difference ruthlessly.