Data Center Power Supply Strategy for the U.S. Electricity Squeeze

For most of the last decade, the power question at a data center kickoff meeting was about the electrical room: utility service size, switchgear ratings, generator placement. In 2026 the question moved two voltage levels up the stack. The bottleneck is no longer the design — it is whether the grid can produce, ship, and schedule the megawatts at all, on any timeline that matches a GPU procurement cycle.

The numbers behind that shift are unambiguous. PJM's capacity auction, which cleared at $28.92 per MW-day for the 2024/2025 delivery year, cleared at $329.17 two auctions later — a factor of eleven, with the next two auctions pinned at the FERC-approved price cap [3][4]. PJM's own market monitor attributes the increase primarily to data-center load growth [4]. The U.S. data center fleet consumed 176 TWh in 2023, 4.4 percent of national electricity; the Department of Energy's official scenario range for 2028 is 325 to 580 TWh — 6.7 to 12 percent of everything the country generates [1][2]. Meanwhile the machinery of supply is backlogged years deep: gas turbine order books stretch past 2031, and the median generation project waits about five years in interconnection queues where barely one in seven applicants ever reaches commercial operation [6][7].

This article is a strategy read, not a market commentary. It lays out what the constraint actually is, and then works through the five playbooks operators are using in response — siting for speed, selling flexibility, owning generation, signing long-dated firm power, and designing the load itself to be grid-friendly. Where a number matters, we compute it. (For how the electrical architecture inside the hall is changing in response, see our pieces on rethinking power for AI data centers and the 800 VDC distribution architecture.)

Fig. 1 — A modern high-density compute hall. The racks are the easy part; every megawatt they draw now has to be found in a market with no slack.

How tight is it, actually

Three independent measurements describe the same squeeze:

• Consumption. Data centers used 58 TWh in 2014, 76 TWh in 2018, 176 TWh in 2023 — roughly a tripling in nine years, driven almost entirely by AI accelerators [1]. The DOE range for 2028 (325–580 TWh) implies adding somewhere between 149 and 404 TWh of annual demand inside five years [1][2].
• Peak demand. Grid Strategies found that utilities' five-year nationwide load growth forecasts jumped from 23 GW to 128 GW in just two years of forecasts, with ERCOT alone adding 37 GW [2]. PJM expects its summer peak to grow from 156 GW in 2026 to 183 GW by 2030 — a 17 percent expansion of the world's largest wholesale market in four years, mostly data centers [4].
• Price. Capacity — the payment generators receive simply to exist and be available — is the cleanest scarcity signal. Its five-auction trajectory is shown below.

Fig. 2 — Left: PJM capacity price, up 11.4× in two auctions and capped ever since; an uncapped 2027/28 simulation reached roughly $530/MW-day [4]. Right: U.S. data-center electricity consumption, with the 2028 DOE scenario band [1].

Convert the capacity price into an operating cost and the squeeze becomes a line item. A 100 MW campus whose peak load contribution (PLC) tag is 100 MW paid about $1.06 million for capacity in the 2024/25 delivery year. At the 2026/27 clearing price the same tag costs $12.01 million a year — an extra $10.96 million for nothing the operator changed. Spread over the campus's delivered energy at 85 percent load factor (744,600 MWh), capacity alone moved from $1.42 to $16.14 per MWh.

The building was never the constraint

Here is the fact that reorders every strategy discussion: the fastest item in a new campus's critical path is the part you can see. A data center building takes 12 to 18 months to construct [5]. Everything that feeds it takes longer:

Fig. 3 — The four clocks. Median interconnection wait is now about 55 months, up from three years in 2015 and under two in 2008; new combined-cycle gas plants that took 3.5 years to build in 2023 now take about five, and the largest turbine frames up to seven [5][6][7].

The queue deserves its own sentence of arithmetic. Roughly 2,290 GW of proposed generation and storage sits in U.S. interconnection queues — nearly twice the entire installed fleet — and historically only about 13 percent of queued projects reach commercial operation [6]. Scroll the queue by technology and the problem sharpens: solar and storage dominate the line (956 GW and 890 GW), while natural gas — the resource everyone now wants — holds just 136 GW of queue positions [6]. On the manufacturing side, GE Vernova's gas turbine backlog reached 116 GW by mid-2026 with new orders booking for 2031 delivery; Siemens Energy holds 69 GW with three-plus-year lead times [7]. A 500 MVA power transformer — the kind a hyperscaler campus needs by the row — carries a two-to-three-year lead [7].

Playbook one: site where the power already exists

The first strategy accepts the queue as a fact of physics and instead shops for capacity that has already cleared it. In practice that means four site archetypes:

Site archetype	What you are really buying	Example
Brownfield / retired generation sites	An existing high-capacity interconnection and transmission path. Reusing a retired plant's grid access skips the queue position entirely — the grid already studied and built for those megawatts.	Microsoft's Crane deal works precisely because TMI Unit 1's interconnection still exists [8]. DOE explicitly lists coal-site repurposing as a federal strategy for data-center power [1].
Utility-anchored campuses	A utility builds dedicated generation ahead of demand, amortized across a long take-or-pay contract.	Entergy is building 3 GW of new generation to serve Meta's Hyperion campus in Louisiana by 2028; Brookfield and NextEra's Paducah, Kentucky campus pairs with a dedicated 2 GW gas plant and up to 2.6 GW of storage [7].
Tethered / co-located at existing plants	Physical adjacency to an energized generator; power flows across a short private path rather than a congested corridor.	Amazon bought Talen's 960 MW Cumulus campus beside the Susquehanna nuclear plant for $650 million, then extended the arrangement into a 1,920 MW power agreement [9].
Off-grid private networks	No interconnection at all: private generation sized to the load, on private land.	Pacifico Energy's GW Ranch in Pecos County, Texas plans up to 5 GW of private gas, storage and solar serving data center loads with no grid connection [7].

Table 1 — Four ways to buy power that is already flowing or nearly so. The trade-off is consistent: speed and certainty in exchange for anchoring your campus to one generation asset and its fuel.

The discipline this playbook demands is temporal, not geographic. The megawatts you decline to sign for are the ones your competitor signs for; queues in ERCOT now hold roughly 445 GW of large-load requests, 77 percent of them data centers [11]. Siting due diligence has effectively become interconnection due diligence: the first question for any site is which of the four clocks (Fig. 3) has already been started, by whom, and whether that headway transfers with the land.

Playbook two: sell flexibility, not just buy energy

The second playbook inverts the operator's posture toward the grid. A flexible data center is not merely a customer; in PJM and ERCOT it can be a paid grid resource — and the payments are no longer rounding errors.

The mechanism in PJM is the capacity tag. Your PLC is set by your facility's draw during the single system peak hour each year, and it multiplies through the next year's capacity bill at roughly $329 per MW-day. The arithmetic of getting that one hour right:

• Shed 1 MW during the system peak hour → next year's capacity bill drops by about $120,000 [3].
• Shed 5 MW → about $601,000, from sixty minutes of load discipline.
• A 100 MW campus that can shed 20 percent on peak-hour notice converts a $12.0 million capacity liability into a manageable one — and can additionally enroll the sheddable portion in PJM demand response, where a curtailable megawatt earns on the order of $120,000 a year at current clearing prices [3].

ERCOT's stack is deeper because Texas prices scarcity more aggressively. Four programs matter, and they stack:

Program	What it pays for	Value, per MW per year
4CP avoidance	Curtailing during the four 15-minute summer system peaks that set next year's transmission charges	$50,000–$150,000 [12]
ECRS	Responding within 10 minutes to contingency events	$30,000–$80,000 [12]
ERS	Standing by for emergency curtailment (10- and 30-minute products)	$5,000–$20,000 [12]
Scarcity avoidance	Not consuming during real-time price spikes (up to the $5,000/MWh cap)	$5,000–$50,000 [12]

Table 2 — ERCOT flexibility revenue for a 100 MW data center, stacked 4CP + ECRS + ERS: roughly $8.5 to $25 million a year in good years [12]. At the current ERCOT transmission rate of $66.76/kW-yr, a 500 MW load that nails the four peaks avoids $33.4 million of annual transmission charges [11].

EPRI's grid modeling reaches the same conclusion from the system side: when data-center load can be shifted temporally, the average price of the power the center actually consumes falls [2]. Training clusters are the natural first candidates — a checkpointed training run tolerates a two-to-four-hour pause far better than a serving workload, and batch inference can move across hours without users noticing.

Playbook three: own the generator

The third playbook stops renting the grid's atoms. Behind-the-meter (BTM) generation — gas reciprocating engines, aeroderivative turbines, fuel cells, batteries — has moved from niche to mainstream as queue times stretched, and it now scales to gigawatts [7]. The economics, worked for a 100 MW simple-cycle installation:

Component	Basis	Result
Capital recovery	$1,100/kW installed × 100 MW, 10-year CRF at 8%	$16.4M/yr = $22.0/MWh
Fuel	9,200 Btu/kWh heat rate at $4.50/MMBtu	$41.4/MWh
O&M	Simple-cycle recip, dispatch duty	$7.0/MWh
BTM total	no capacity charge, no transmission adder	≈ $70/MWh
Grid all-in comparison	energy ~$50 + capacity $16.1 + T&D ~$20	≈ $86/MWh

Table 3 — BTM economics at 85 percent load factor. The margin is real but not enormous; the decisive advantage is the delivery date. A recip order is measured in months; the grid alternative is measured in half-decades [7]. Pairing 100 MW of gas with 400 MWh of batteries (~$120 million at $300/kWh installed) covers startup sequencing and peak shaving.

The regulatory fault line here is co-location. In November 2024, FERC rejected (2–1) an amended interconnection agreement that would have expanded behind-the-meter sales from the Susquehanna nuclear plant to Amazon's adjacent campus from 300 to 480 MW, on cost-shift and reliability grounds [10]. The market's answer was to restructure in front of the meter instead: June 2025 brought a 17-year, up-to-1,920 MW front-of-meter PPA between Talen and AWS [9][10]. The deal structure is now a template — keep the megawatts grid-accounted, contract them long, and let the adjacency deliver the speed.

Politics has followed. In March 2026, the major hyperscalers signed a White House "Ratepayer Protection Pledge" committing to "build, bring, or buy" the energy for their data centers and to pay the full cost of associated grid infrastructure — a formalization of the principle that the AI buildout self-supplies rather than free-rides on capacity paid for by everyone else [9]. Expect regulators to keep pulling in the same direction: the era of the passive, always-full-load grid customer that pays yesterday's rates is closing.

Playbook four: sign the long-dated firm electron

The fourth playbook is the hyperscaler signature move: 15-to-25-year power purchase agreements that finance firm, carbon-free generation directly. Nuclear dominates the list because it is the only established technology that delivers gigawatt-scale, 24/7, carbon-free output — the exact profile an AI training campus needs:

Buyer	Plant / technology	Capacity	Status and first power
Microsoft	Crane Clean Energy Center (TMI Unit 1 restart), 20-year PPA; $1B DOE loan closed Nov 2025	835 MW	Restart targeted 2027 [8]
Amazon	Susquehanna (existing PWR), 17-year PPA	1,920 MW	Active — power flowing today [9]
Amazon	Calvert Cliffs, 20-year PPA incl. 190 MW uprate	690 MW	Signed Sep 2026; uprate 2030–32 [9]
Meta	Clinton (existing PWR), 20-year PPA; plus TerraPower, Oklo agreements	1,100 MW; up to 6.6 GW total	2027 start; SMR units 2032–35 [9]
Google	Kairos Power KP-FHR SMRs (world's first corporate multi-SMR agreement)	500 MW	First reactor ~2030, full 500 MW by 2035 [9]

Table 4 — Selected nuclear PPAs backing U.S. data centers [8][9]. Across all tracked deals, the four hyperscalers have committed on the order of 10 GW.

Read the table with a scheduler's eye, though. The only megawatts that arrive before 2030 come from restarts and uprates of plants that already exist; every SMR row is a bet on technology that, commercially, barely exists — as of early 2026, precisely four SMR units operated anywhere in the world, all in Russia and China [10]. A campus pro-forma that pencils only if an SMR arrives on schedule is buying timeline optionality, not power. The sound version of this playbook treats nuclear PPAs as the 2030s layer and finances the 2027–2029 gap with playbooks one through three.

Playbook five: build the load the grid wants

The fifth playbook changes the product rather than the procurement. NERC's large-load task force identified the core planning problem: data centers are historically "firm" loads — they request full capacity, always, and grid planners must hold generation for them at one-in-ten reliability standards [5]. A load that can genuinely modulate changes that calculus, and every market is now writing rules to reward it and to charge the load that will not:

• Flexible interconnection. Accept a curtailment envelope (for example, 100 MW firm plus a curtailable band) and the interconnection studies get shorter and cheaper; the grid builds against a softer peak.
• Frequency and contingency response from existing hardware. The UPS fleet and battery plants already installed for ride-through are, in aggregate, a large fast-response resource; markets like ERCOT's ECRS pay for exactly that response profile if the controls allow participation [12].
• Workload-aware curtailment. Checkpointed training, deferrable batch inference, and temporal shifting of non-latency-critical work — EPRI's modeling shows temporal load shifting directly lowers the average price of power a data center consumes [2].
• Honest demand forecasts. Utilities' five-year forecasts jumped fivefold in two years [2]; speculative duplicate requests are part of why. Sites that bring verified, staged load schedules move through interconnection studies faster than sites that ask for everything at once.

What this means for everyone else

The five playbooks above were written by balance sheets north of $10 billion. But the same scarcity prices flow down to every colocation buyer, GPU lessor, and mid-size operator — and the decision framework compresses cleanly:

Question	If yes	If no
Is your workload checkpointable or batch-shiftable?	Enroll in flexibility programs first (Table 2) — it is revenue with no construction.	Budget the capacity-tag inflation as a permanent line item ($16+/MWh and rising in PJM).
Signing a colocation or GPU-rental contract in 2026–27?	Ask which of the four clocks is already running and demand an energization date with liquidated damages.	Assume any "powered shell" quote is only as firm as its substation application [5][7].
Building in ERCOT?	4CP discipline is worth six figures per MW; automate it, and watch the SB6 rulemaking [11][12].	Building in PJM: PLC-tag management is the single highest-ROI hour of the year [3].
Need power before 2028?	Playbooks one and three are the only ones that deliver on that horizon.	Nuclear PPAs and SMRs are a 2030s layer; treat them as such [8][10].

Table 5 — The compressed decision grid. For the equipment layer underneath whichever strategy you choose — switchgear, busway, CDUs, and the racks in Fig. 1 — send us your project scope for a quote.

One closing observation. Every previous constraint in this industry — floor space, then cooling, then chips — was eventually answered by engineering and capital, and each answer redistributed market share toward whoever saw it first. Power is a harder constraint because it is mediated by queues, regulators, and turbine factories, but it is the same test. The operators who treat electricity as a design input from day one — who start the clocks, sell the flexibility, and build the load the grid wants — are the ones who get to energize on schedule while everyone else waits in line.

References

1. U.S. Department of Energy. "DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers" (LBNL 2024 U.S. Data Center Energy Usage Report: 176 TWh in 2023, 325–580 TWh in 2028, 4.4% → 6.7–12%). energy.gov
2. Lawrence Berkeley National Laboratory. Large Load Literature Review, August 2025 update (Grid Strategies: five-year forecast 23 → 128 GW, ERCOT +37 GW; EPRI: data-center load shifting lowers coincident power price; DOE flexibility and coal-repurposing strategies). eta-publications.lbl.gov
3. PJM Inside Lines. "PJM Auction Procures 134,311 MW of Generation Resources" (2026/27 BRA results: $329.17/MW-day at the FERC cap, +5,400 MW forecast load). insidelines.pjm.com
4. New Jersey Board of Public Utilities. An Investigation of PJM's Capacity Market, August 2026 (price ladder $28.92 → $269.92 → $329.17 → $333.44 → $325.00; peak 156 GW 2026 → 183 GW 2030 → 222 GW 2036; 5.6% short of IRM; market monitor attribution). nj.gov
5. NERC. Characteristics and Risks of Emerging Large Loads (data center buildings 12–18 months; interconnection request-to-COD median ~5 years, p25 40 / p75 70 months; firm-load planning risk). nerc.com
6. Pilot Energy, "The interconnection queue" (LBNL Queued Up 2025 data: ~2,290 GW active, median 55 months, 13% completion, gas 136 GW of queue vs 956 GW solar). pilotenergy.com
7. PE Presswire, "Power, Not Capital, Is Now the Bottleneck on AI Data Center Deals" (GE Vernova backlog 116 GW mid-2026, orders for 2031; Siemens 69 GW; CCGT build 3.5 → ~5 years, frames to 7; Entergy 3 GW for Hyperion; Paducah 2 GW gas + 2.6 GW storage; Pacifico GW Ranch 5 GW off-grid; transformer 2–3 year leads). pepresswire.com
8. U.S. Congress, House Committee on Science, Space, and Technology. Powering Demand: Nuclear Solutions for AI Infrastructure, hearing charter, June 2025 (Microsoft–Constellation 20-year PPA for 835 MW TMI Unit 1 restart; DOE $1B loan; LBNL figures). govinfo.gov
9. CM Energy Insight, "When Big Tech Buys the Power Station" (Talen–AWS 17-year 1,920 MW PPA; Cumulus campus $650M; Calvert Cliffs 690 MW 20-year PPA; Meta Clinton 1.1 GW and up to 6.6 GW; Google–Kairos 500 MW by 2035; White House Ratepayer Protection Pledge, March 2026). cmenergyinsight.com
10. World's Top Data Centers, "Why Data Center Operators Are Looking at Nuclear" (FERC's 2–1 rejection of the Talen ISA amendment, Nov 2024; only four commercial SMR units worldwide as of early 2026; restarts vs SMR risk framing). worldstopdatacenters.com
11. Capture Cascade, "PUCT Project 58484" (Texas SB6; 4CP reform draft: more coincident peaks, minimum demand charges 10–15 years; $66.76/kW-yr transmission rate; $33.4M/yr avoidable per 500 MW; ERCOT large-load queue ~445 GW, 77% data centers). capturecascade.org
12. LōD, "ERCOT Demand Response Programs: Operator Guide" (4CP $50k–$150k/MW-yr; ECRS $30k–$80k; ERS $5k–$20k; scarcity avoidance $5k–$50k; $5,000/MWh cap; automation requirement). lod.io