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 costs, in what structure, and who bears the cost of the grid you are about to stress. The two questions are separable. 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 May 2026 model of a 1 GW campus, energy is the largest operating-cost category, yet it is only about $0.6B of roughly $8.5B in annualized total cost (about 7%). Servers cost about $5.0B a year and the facility about $1.4B: silicon dominates capex, and annualized capital dominates TCO. Once the chips are bought, though, power is a major cost the operator can still move, because siting and supply terms shape the delivered bill. 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 the steel already in the ground. Three variables decide whether a site is powerable on a reasonable schedule: transmission proximity, voltage class and substation headroom. Get them wrong and the cheapest energy market in the country is irrelevant, because the lines, rights-of-way, equipment and approvals can dominate the 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 reach the load (ERCOT's ~$33B statewide STEP plan, which pairs 2,468 miles of new 765 kV line with new and upgraded 345 kV and 138 kV work, plus PJM and SPP build-outs; see Chapter 3.2). A site beside an existing high-voltage line with spare thermal capacity is a different proposition from one that needs miles of new line. Every mile carries its own permitting, right-of-way acquisition and 100+-week lead times for conductor and structures. In practice the distance to a viable tap is a schedule variable, not a geography variable.

Voltage class sets both the equipment you buy and the losses you eat. A higher delivery voltage moves more power per ampere with lower I²R losses, but the step-down transformers cost more and take longer, and the customer substation is bigger. A 1 GW load taken at 345 kV is a different electrical interface, with a different substation cost, from 200 MW taken at 138 kV. The decision is effectively irreversible: voltage class is baked into the substation, the protection scheme and the interconnection agreement, and changing it means rebuilding the interface.

Substation headroom 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 already at its firm rating needs a new transformer bank (US substation-transformer lead times above 160 weeks, Wood Mackenzie, May 2026) 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 competing for the same headroom. Headroom is a depletable shared resource, and you are rarely the only one shopping for it.

Two rules of thumb let you run that screen from a desk. A substation's firm rating is what it can carry with its largest transformer out of service, so three 500 MVA banks give about 1,000 MVA firm, not 1,500. A line's loadability depends on its length as well as its voltage. On long lines, AEP's transmission reference puts a bundled 345 kV single circuit at about 400 MW, a 500 kV circuit at about 900 MW and a 765 kV line at 2,200–2,400 MW, each at 1.0 surge impedance loading over distances up to 300 miles. Short lines are limited by conductor and station-equipment thermal ratings instead, and can carry more. At 345 kV, 1 GW is about 1,670 A per phase at unity power factor, and keeping a gigawatt served with one circuit out takes at least two circuits that can each carry the whole load. A candidate whose substation or lines fail that arithmetic needs new steel, and its schedule then follows the long-lead equipment.

Operators routinely shop on one number, a delivered ¢/kWh, but that number averages components that behave very differently under an AI load shape. The delivered bill adds up the energy price you pay, any capacity and demand charges billed on their own, delivery charges, and whatever fixed charges and taxes apply. When you analyze exposure, split LMP into system energy and congestion, and add marginal losses in markets that price them. Count each embedded part once. Add a basis adjustment only if your contract bills it as its own line. Each part has its own driver, its own volatility and its own hedge:

• Energy (the commodity). In organized markets, the bus Locational Marginal Price (LMP) is the dispatch primitive. The price a load pays may instead be a load-zone settlement price, a zonal LMP, a utility tariff or a negotiated retail product, and it 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. Congestion is embedded in LMP, as are marginal losses where the market implements marginal-loss pricing; other loss treatments adjust the billed energy quantity instead. 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, and a node beside 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 obliges the system to procure. PJM's capacity auction clearing prices spiked roughly an order of magnitude in 2025–2026 as data-center load tightened supply and demand, turning capacity from a rounding error into a material line item. FERC's September 29, 2026 order accepted PJM's Reliability Backstop Procurement but suspended it for five months, to February 28, 2027, subject to refund, and PJM said on October 12 that it intends to refile. The procurement is not available as a remedy, and its cost allocation, the exit terms for transmission owners and the collateral asked of load-serving entities are all still open. Carry any backstop charges and security requirements as open exposures in the model, and do not count the procurement as capacity you have secured (see Chapter 3.2). Federal emergency orders are a second, non-market response. DOE's serial Federal Power Act §202(c) orders keep retiring PJM plants such as Eddystone available, and PJM charges their cost to load. Model them as a charge on your load, not as capacity you can count on (Chapter 3.11).
• Demand charges. A regulated retail tariff may bill energy consumed and billing demand as separate lines. The demand line is the rate ($/kW-month) × billing demand (kW), often with a ratchet that keeps it near the annual peak for months. For a high-load-factor data center this is usually favorable, because a flat 90%+ load factor spreads the demand charge over enormous consumption. 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% system energy and 10% congestion: merchant-exposed and heavy in basis risk. The other may be 50% capacity and demand charges: regulated and schedule-stable, with little hourly exposure but a high floor. Hedging starts from this decomposition, because 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.

A steady load is a good power customer, and it can sometimes secure rates a flickering industrial load cannot. Load factor is average draw divided by actual peak draw, so a flat facility scores high. That flat draw spreads a utility's fixed cost of a substation and a line over the maximum possible consumption, driving the effective ¢/kWh of the delivery and demand components toward their floor. A high-load-factor data center can therefore clear a deal on a tariff that would crush a peaky factory. How much of your contracted capacity you use is a different ratio. Run flat at 80–95% of contract demand and your load factor is high, yet you still pay for reservation you leave unused whenever the minimum billing demand sits above your actual peak.

The corollary is a warning. Your effective per-MWh cost rises whenever delivered MWh fall while billing demand and the other fixed charges stay put. Curtailment does it, and so do demand-response events, idle gaps between training jobs and a half-full hall during ramp. 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 MWh is the flat draw times 8,760 hours. Each month's billing demand is the larger of actual peak and the contract minimum. Energy and delivery cash are MWh times their prices, demand cash is billing kW times the kW-month price times 12, and the facilities charge is added on top. Divide the total by MWh delivered, rounding only at the end.

Take the component contract at the base draw ($71.8/MWh against the $80.0 offer) and the variable offer in the downside case ($88.7 against $80.0). While the 9.00 MW demand minimum governs, the crossover is the fixed demand-and-facilities cash divided by the gap between the variable price and the component energy-plus-delivery price: $1.18M ÷ $25/MWh = 47.2 GWh a year, or 5.39 MW average. That sits well under the 9.00 MW minimum, so the formula's assumption holds. Below 5.39 MW, the variable offer is cheaper. Curtailing energy does not erase a continuing demand or facilities charge.

For a real site, get the operative tariff: AEP Ohio’s DCT explanation shows how a utility tests billing demand separately from energy consumed. 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 has no single price: dispatch produces an LMP at every node. A bus LMP is a dispatch primitive, while the price a load pays follows its market and contract. ERCOT, for example, 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, under whatever loss rules the market uses, reach the bill through the pricing and metering steps that apply.

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 sit at 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 2026 has moved past "firm grid or not" to how much curtailment you will accept in exchange for earlier energization and a lower energy charge. Conditional or non-firm service can get a site energized sooner, provided the studies, the network capacity and an executed offer all back it. FERC approved SPP's CHILLS on June 5, 2026. It offers non-firm, as-available transmission service for up to seven years, bridging the time it takes to secure the resources or upgrades that firm service needs. On June 18, 2026, FERC issued show-cause orders to all six FERC-jurisdictional RTOs and ISOs, which must now justify or reform their large-load rules, including how they serve flexible loads. The orders open proceedings; they create no executable offer. PJM's co-location transmission services are a separate pathway from the Interim Resource Adequacy Service (IRAS) PJM proposed for new large loads. On October 9, 2026, FERC (197 FERC ¶ 61,019, ER26-3515-000) accepted PJM's Large Load definition (50 MW and up) and registry, effective October 12, but rejected IRAS and the Bring Your Own New Capacity rules, so IRAS is not a service you can underwrite (see Chapter 3.2).

Texas works differently. Under SB6, the serving utility must provide firm-load-shed curtailment protocols for any non-exempt transmission-voltage customer that interconnects after December 31, 2025. The protocol is a condition of service. It buys no faster queue position and no discounted energy. ERCOT's Batch Zero notice of September 3, 2026 likewise handed out conditional classifications; it did not grant permission to energize (see Chapter 3.2).

Curtailment is a priced lever, not a binary. Duke’s February 2025 national model puts about 98 GW of US headroom within reach of new load that accepts 0.5% annual curtailment (Chapter 3.2), a national result, not capacity at your substation. 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, in return for a faster, cheaper interconnection. An inference service is a different case when its latency and availability commitments ride on the curtailed site staying up. It needs tested storage or generation, spare serving capacity at another site, or some other proven continuity plan to carry it through each event. If nothing covers the event, a curtailment can break the SLA and wipe out the energy saving. Underwrite the cost of the expected curtailment hours for 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 and 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. It was a footnote five years ago. It is now often the deciding variable on whether a project gets approved, because rising residential bills in data-center-heavy regions have made cost allocation a live political fight.

The mechanism that sets the answer is the large-load tariff, written to push the cost of the buildout onto the loads that cause it. Each utility writes its own, with its own service terms. As of September 2026, EEI counted approved tariffs in 25 states and pending ones in seven more. The Oregon POWER Act template is worth knowing because variants of it are spreading. The Oregon PUC approved PGE’s Schedule 96 under it in May 2026 (Order 26-154, UM 2377), and service began July 8, 2026. A load of 20 MW or more, counting contracted demand and aggregated campus tenants, falls in the class. The load pays 100% of the distribution facilities used exclusively by the customer, with proportional allocation for shared facilities (PGE Rule I). Minimum generation and transmission demand charges at 90% of contracted system capacity set a minimum billing-demand floor for those components: you pay for 90% of contracted capacity whether you draw it or not, which protects the utility against stranded investment if you under-build or leave. It does not oblige you to buy 90% of your contracted annual kWh. 10–30-year contract terms scale with load size, reaching 30 years at 220 MW and above, and a 1¢/kWh surcharge applies at 100 MW or more of allocated system capacity. These are the PUC figures (2025 statute; 2026 order). Check the operative Schedule 96 and UM 2377 sheets and the project service agreement before you put any of them in a bill model, and note that Pacific Power’s data-center tariff (UE 463) is a separate proceeding.

California wrote the same cost-causation principle into law on September 21, 2026. SB 886 directs the CPUC to set data-center tariffs for interconnection, transmission, distribution and generation by January 1, 2028, with an early-termination fee for a customer that leaves within 10 years of interconnection. AB 2383 and SB 1168 add tariff and rate-structure duties. Until the CPUC proceedings set the terms, a California site has a direction to underwrite (the load pays for its own grid costs) but no rate.

Texas separates SB6 emergency-curtailment protocols from its 75 MW-and-up large-load interconnection standards (Project 58481). PUCT §25.194, in force from October 8, 2026, sets a $100,000 study fee and lets ERCOT protocols add a flat fee of their own. Before ERCOT studies the load, the customer must also post security of $50,000 per MW of requested peak demand, about $50M for a 1 GW request. Once the customer signs the standard interconnection agreement, the security becomes the greater of $50,000 per MW of contracted peak or the upgrade costs the study allocates. The customer also pays its direct interconnection costs in cash, invoiced before the utility incurs them, and the payment is not refundable. ERCOT still recovers ongoing transmission costs through 4CP allocation, and that charge sits apart from the direct payments and the security posted for system upgrades. A PUCT proposal of July 24, 2026 in Project 58000 would add a minimum billing demand, set from contracted peak, the highest annual non-coincident peak or 12CP demand. Payments would run for at least 20 years from the date service capacity becomes available. As of October 12, 2026 it is still a proposal. SB6 gives PUCT until December 31, 2026 to revise how wholesale transmission costs are allocated. PUCT is evaluating transmission cost recovery under SB6 §6 in Project 58484 and writing the rule in Project 58000.

Virginia, the largest US market, added a flat $0.011/kWh data-center consumption tax in July 2026, and it reaches self-supplied electricity too. HB30, which became Chapter 1, Item 3-5.24 when enacted on June 29, 2026, caps the tax's deposits to the general fund at $600M a year. Anything collected above that comes back as pro-rata refunds, net of administration costs, starting in FY2028. This tax is not dedicated network-upgrade recovery.

The consequence for the model is direct and large. Together these mechanisms can add a fraction of a cent to several cents per kWh of effective cost. Minimum billing demand also fixes part of the bill as a demand charge, owed however far your energy use falls. That makes the curtailment savings discussed above partly illusory, because 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 is in Chapter 16.1.

Once the chips are bought, power is one of the big operating costs you can still control. You buy the GPUs at market price and write them down on a contested depreciation schedule (→ Chapter 1.8), and utilization and refresh decisions stay consequential. Siting is how you move the power cost: the node, the voltage class, the tariff, the curtailment terms, and grid versus behind-the-meter supply. Where you build and the supply terms you sign change the delivered bill a lot. By region, metro power typically runs roughly 1.5–2x the cheapest industrial rates in the US (EIA, July 2026). EIA’s industrial averages for that month put Virginia at 10.07¢/kWh, about 1.4x Texas (7.07¢) and 1.7x Louisiana (6.08¢), and the gaps are wider for California and New Jersey. The 8.47¢/kWh Texas tile above is EIA’s commercial-sector average for the same month, a different customer class. Treat all of these statewide averages as benchmarks. A site’s real savings come from pricing its own voltage class, tariff, supply contract and curtailment terms.

The leverage is clearest against the revenue the asset earns. Against a topline of roughly $12–13B a year per GW of contracted critical IT (SemiAnalysis, June 2026; Chapter 1.8), a mis-structured energy bill can run to many hundreds of millions a year. A 2¢/kWh difference in delivered power cost, 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. That saving is material, and any procurement saving should be compared with efficiency and utilization gains on the same facility-meter and useful-work boundary. This is why the screen runs availability first, structure second, and everything else after.

Separate commitments from price exposure, the way Chapter 1.1 sorted the scoping decisions by reversibility. Some commitments are expensive to change, or change only under rights the contract spells out: voltage class, the installed interface, contract tenor, reserved capacity, exit fees and minimum billing demand. Prices behave differently. A fixed-price PPA holds its strike for the contracted term, but capacity prices, basis spreads, indexed energy and regulated rates can all move even when the service commitment underneath them cannot. Curtailment is an operating exposure: cover it or price it. Record termination, reassignment and tariff-election rights next to each price assumption. 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.