How Data Centers Are Powered: Grid, Behind-the-Meter Gas, Solar + Storage, Nuclear and Hybrid Configurations

How does power get from the transmission line to the chip?

Every data center sits at the bottom of the same ladder of voltages. The point of interconnection is a transmission line at 69–230 kV in most regions; anything above about 20 MW needs a dedicated substation stepping down to at least 34.5 kV, a 100 MW site typically needs two independent feeds, and a new substation takes two to four years (amcompute, Mar 2026). ERCOT's transmission runs at 69, 138 and 345 kV (Compute Law Blog, Oct 2026). This is Phase 2 of How a Data Center Is Built, and the phase that sets the date.

From the substation, 13.8 or 34.5 kV feeders reach the building, where unit substations step down to 480/277 V or 208/120 V three-phase (TierPoint, Dec 2025). Behind them sit the generators and batteries: legacy designs carried diesel for 100% of load, while new builds carry 15–40% generator coverage with BESS "on virtually every project" (pv magazine, Jul 2026). Then a static or rotary UPS, busway to the row and, in a GB200 or GB300 NVL72 rack, a 54 VDC busbar (NVIDIA, May 2025). The long pole is the transformer: large power transformer lead times went from about 50 weeks in 2021 to about 120 weeks in 2024, and substation-class units now exceed 160 weeks (Data Center Knowledge, May 2026).

Read as text

• 345 kV to 54 V: the power chain inside the fence
• Typical voltages, lead times and losses at each stage, and the redundancy topologies that decide how much of each you buy.
• 01 · POI · Transmission · 69–345 kV · ERCOT runs 69 / 138 / 345 kV; first 765 kV line due Dec 2028. ≥20 MW needs its own substation; 100 MW typically two feeds. · Substation build 2–4 yrs
• 02 · Main transformer · → 34.5 / 13.8 kV · Large power transformer lead time ~50 wks (2021) → ~120 wks (2024) → >160 wks (2026); GSU units 80–210 wks. · >160 weeks lead time
• 03 · MV switchgear · 13.8 / 34.5 kV · Feeders to the building; 13.2, 13.8 and 27.6 kV service entrances are all common. Lineups sized N+1 or 2N. · Where redundancy is bought
• 04 · Generators + BESS · 15–40% coverage · Legacy: diesel 1:1 for 100% of load. New builds: partial generator coverage with 2–4 h batteries on virtually every project. · Build ≈ $13M/MW all-in
• 05 · UPS · Minutes or seconds · Static double-conversion rides on batteries for minutes; rotary / flywheel for seconds. BESS increasingly doubles as the backup. · UPS in 42% of power outages
• 06 · Unit substation + busway · 480/277 V · Or 208/120 V three-phase; 415/240 V busway to the row. A legacy chain makes 5–7 conversions at 2–4% loss each. · 12–15% lost fence to chip
• 07 · Rack · 54 VDC busbar · GB200 / GB300 NVL72 at ~120 kW. A 1 MW rack at 54 V needs up to 200 kg of copper, so 800 VDC arrives from 2027. · 800 VDC: +3–5 pts efficiency
• 8–12% · of facility power consumed by transformers, UPS and PDUs; hyperscale PUE 1.10–1.15 (mgrid 2025)
• 1.54 · industry-average PUE, flat for six years; 1.44 at sites of 20 MW or more (Uptime Institute 2025)
• 45% · of impactful outages are power: UPS 42%, transfer switch or utility 36%, generator 28% of those (Uptime 2025)
• REDUNDANCY TOPOLOGIES — THE OTHER MEANING OF "CONFIGURATION"
• N · → load · Exactly the capacity the load needs. No spare; any module failure drops the load. Uptime Tier I territory.
• N+1 · 41% of operators · → shared bus · One spare module on a common bus. Cheapest real redundancy; the bus itself is shared.
• Distributed redundant · catcher · → · STS · Modules run below capacity so any one catches the others' load. The static transfer switch is the single point of failure.
• 2N · 42% of operators · → A / B · Two complete, independent paths to dual-corded IT. Most expensive; concurrently maintainable (Tier III and up).
• 2N+1 · 3N/2 · Variations: a spare on each path, or three half-size paths serving two loads to cut the 2N cost premium.
• TIER I · BASIC CAPACITY · TIER II · REDUNDANT COMPONENTS · TIER III · CONCURRENTLY MAINTAINABLE · TIER IV · FAULT TOLERANT · Uptime dropped availability percentages in 2009 — "99.982%" is a vendor number.
• TIER I · BASIC CAPACITY
• TIER II · REDUNDANT COMPONENTS
• TIER III · CONCURRENTLY MAINTAINABLE
• TIER IV · FAULT TOLERANT
• Uptime dropped availability percentages in 2009 — "99.982%" is a vendor number.

Each conversion costs something. Transformers, UPS and PDUs take 8–12% of facility power and hyperscale fleets run PUE of 1.10–1.15 (mgrid, Oct 2025), while the industry average has sat at 1.54 for six years and sites of 20 MW or more average 1.44 (Uptime Institute, 2025). A legacy chain of five to seven conversions loses 12–15% between fence and chip; 800 VDC claims back 3–5 points (datacentres.com, Aug 2026).

What do N+1, 2N and Tier III actually mean?

"Configuration" has a second meaning here: how many paths and spares the plant has. N is the capacity the load needs. N+1 adds a spare module on a shared bus. Distributed redundant (catcher) systems run several modules below capacity so any one can pick up the others' load, with a static transfer switch as the single point of failure. 2N is two independent paths; 2N+1 and 3N/2 are variations (DgtlInfra, Mar 2026). Operators split evenly: 42% run 2N, 41% N+1 (Uptime Institute, 2025).

Uptime's Tiers map onto these: Tier I is basic capacity with UPS and a generator, Tier II redundant components, Tier III a second delivery path so anything can be maintained live, Tier IV fault tolerant (Uptime Institute Journal, 2021). Uptime dropped availability percentages in 2009, so "99.982%" is a vendor number, not a Tier. Topology matters because power causes 45% of impactful outages, and UPS failures figure in 42% of those (Uptime Institute, 2025).

What are the seven ways to supply a data center with power?

More than 130 GW of co-located generation has been announced against a US data center pipeline above 160 GW; over 80% of it is gas, and S&P expects 22–32 GW on site by 2030 (Rabobank, May 2026).

Read as text

• Time-to-power vs cost: where each configuration lands
• Months to first power (log scale) against levelized cost, $/MWh. Bands are published ranges; dashed chips are estimates or capex-only.
• TIME TO FIRST POWER · COST · $/MWH
• TIME TO FIRST POWER
• COST · $/MWH
• Fuel cells (SOFC) · ~90 days · no published $/MWh · $3,000–4,000/kW capex
• Recips / aeroderivatives · 12–24 mo · $100–165/MWh · $1,700–2,000/kW recips
• Islanded gas microgrid · 12–30 mo · ≈ $86/MWh · modeled turbines (Heatmap)
• Solar / wind + storage · 18–36 mo · $93–109/MWh · off-grid · ~99% reliability
• ERCOT, transmission voltage · ≥18 mo · with headroom · ~$45–60/MWh · all-in (GR1D estimate)
• Existing-plant nuclear PPA · ~2 yr · online 2027 · ≈ $130/MWh · TMI / Crane
• Grid-only interconnect · 36–84 mo · ~$90–95/MWh · US benchmark (Rabobank)
• Heavy-duty frames / CC · 36–84 mo · $100–165/MWh · $2,300–2,500/kW CC
• Restarts · SMRs · new build · 3–5 yr · restarts · 2030s SMR · $141–220/MWh · (Lazard)
• BTM GAS BAND · $100–165/MWH
• Empty quadrant: no published option is both under 12 months and under ~$85/MWh. Speed is bought with cost or with grid headroom.
• Bridge-to-grid starts in the amber band and moves to the cyan band when the interconnect energizes: 12–24 months on gas, then grid supply with the gas kept as ~500+ h/yr peaking (Enchanted Rock model).
• ERCOT adders: 4CP ≈ $50,000/MW-yr ≈ $5.70/MWh at 100% load factor; 2025 real-time average $36.94/MWh (Potomac Economics).
• gas · grid · renewables + storage · nuclear · estimate / capex-only

How does a grid connection work in ERCOT, and what does it cost?

A load of 75 MW or more is a "large load" under Senate Bill 6: a $100,000 study fee, security of $50,000 per MW, studies, an interconnection agreement, quarterly stability assessments and, last, ERCOT's energization approval (Compute Law Blog, Oct 2026). The queue shows how that is going: 410 GW of requests in March 2026, 87% data centers (ERCOT, Apr 2026), about 474 GW by August (Utility Dive, Aug 2026) against a record peak of 91.1 GW set on July 22, 2026 (EIA, Aug 2026), and only about 2%, or 5,302 MW, energized as of November 2025. With headroom at an existing substation, the realistic floor is about 18 months (Compute Law Blog).

Texas energy is cheap; the connection is what costs time. ERCOT's 2025 real-time price averaged $36.94/MWh, or about $38/MWh all-in (Potomac Economics, May 2026), and industrial retail rates in the West South Central region averaged 6.48¢/kWh in 2025 against 8.62¢ nationally (EIA, Oct 2026). A transmission-voltage load also pays 4CP: the four highest 15-minute system peaks of June through September set next year's transmission charge, roughly $50,000 per MW-year or $5.70/MWh at 100% load factor (The Energy Mag, Jun 2026). Our planning estimate for such a site at high load factor is $45–60/MWh all-in, against the $90–95/MWh national benchmark (Rabobank).

Two SB6 provisions change the risk. Transmission-voltage loads interconnecting after December 31, 2025 must participate in firm load shed (Compute Law Blog), and if non-exporting backup can serve at least 50% of demand, ERCOT may direct the site to drop load or run that backup in an emergency (Baker Botts, Jul 2025). Federal regulators push the same way: after FERC rejected Talen's behind-the-meter expansion for Amazon in November 2024, the contract was re-cut in June 2025 as grid-delivered supply (POWER, Jun 2025).

Behind-the-meter gas: recips, aeroderivatives, frames or fuel cells?

Behind-the-meter generation serves the facility before any utility meter, and the four technologies differ more in lead time than in fuel. GE Vernova's turbine backlog is 116 GW, with new orders booking 2031 delivery and about 20% of the backlog from data center customers (Utility Dive, Jul 2026).

Rabobank's $100–165/MWh assumes Henry Hub at $4–4.50/MMBtu, 8,000–9,000 Btu/kWh heat rates, 40–65% overbuild and ten-year amortization (Rabobank); a gigawatt of load needs roughly 140 MMcf/d of firm gas (Enverus, May 2026). Permits are the exposure: Mississippi's March 10, 2026 permit for 41 turbines (1.2 GW) at xAI's Colossus 2 replaced 27 units that had run unpermitted (DCD, Mar 2026). The TCEQ ladder runs from permit-by-rule for engines up to 3.2 MW (about 876 operating hours, 45-day processing) through minor NSR for 44–300 MW (241 days on average) to PSD review above 415 MW (347 days), with a 25 tpy major-source threshold in the DFW and Houston nonattainment areas (TCEQ, Apr 2026; Encino Environmental, Sep 2026). All of it is paused for data centers as of September 21.

Can solar and storage power a data center?

The best public model is the Stripe, Scale Microgrids and Paces off-grid study: 44% solar with gas backup near $93/MWh against $86 for off-grid gas turbines alone, 90% solar with batteries near $109/MWh, all at roughly 99% reliability rather than five nines and about 50 times the land, which West Texas has (Heatmap, Jan 2025). Projects take 18–36 months (Rabobank) at $1,150–1,600/kW for solar and $1,900–2,300/kW for wind (Lazard via pv magazine, Jun 2025). Google's $4.75B Intersect acquisition, announced December 22, 2025, puts its first co-located site in Haskell County, Texas (Alphabet, Dec 2025).

The ERCOT precedent is Goodnight Wind. On July 24, 2026 the PUCT approved Docket 59220, pairing the 265.5 MW wind farm with Crusoe and Ensign's 260 MW data center on terms that now define Texas co-location: full curtailment within 30 minutes on ERCOT instruction, no paid demand response or ancillary services, and a cumulative 525.5 MW obligation (White & Case, Aug 2026).

ERCOT's operational grid requirements supersede individual data center hardware concerns.Public Utility Commission of Texas, Docket 59220, via White & Case

When will nuclear actually power data centers?

Through grid-delivered contracts with existing plants, soon; through new reactors, not this decade. Crane (the former Three Mile Island Unit 1, 835 MW) restarts in 2027 under a 20-year Microsoft PPA signed in September 2024; Meta's Clinton contract (1,121 MW) starts June 2027; Duane Arnold restarts in Q1 2029 with a $1.9B DOE loan; Kairos's Hermes 2 (50 MW) for Google, the only new-build reactor under construction for a data center customer, targets 2030 (Nuclear News Network, Oct 2026). Rabobank's rule of thumb is two years for an existing-plant PPA, several for a restart, most of a decade for a first-of-a-kind design, and under 2 GW of US SMR capacity before 2033 (Rabobank). TMI power is estimated near $130/MWh (Heatmap) and Lazard puts new nuclear at $141–220/MWh (pv magazine, Jun 2025). The Talen ruling means nuclear reaches data centers through the grid, not behind the meter.

What are bridge power and islanded microgrids?

Bridge-to-grid is on-site generation that carries the full load until the interconnect arrives, then becomes utility-dispatched peaking capacity for 500-plus hours a year in Enchanted Rock's model (Latitude Media, May 2024). SB6 adds a gate: a new load of 75 MW or more pairing with a generator that existed on September 1, 2025 triggers a 120-day ERCOT study and a 60-day PUCT review, with 30-minute curtailment, no paid demand response and liability if shared transmission runs 25% or more below expectation (White & Case).

Islanded microgrids drop the utility altogether. Meta's Socrates runs two 200 MW blocks off-grid on 30 generating units each with about 173 MW of batteries from Q3 2026 (Cleanview, 2026), and Doña Ana (2.2 GW) and Vantage Shackelford (~2 GW) follow the same logic (Epoch AI, Apr 2026). Islanding adds BESS and synchronous condensers for inertia and fault current (Data Center Knowledge, May 2026), and the owner inherits the ripple problem described next.

What changes when the load is GPUs?

Training clusters alternate between compute and communication phases and the power follows: swings of "tens or hundreds of MW" with energy at 0.2–3 Hz, inside the grid's inter-area and plant oscillation bands. GB200 firmware can floor power at 90% of TDP for a 10.5% energy penalty; the fix is GPU-level smoothing plus rack-level storage (Microsoft, OpenAI and NVIDIA via arXiv, Aug 2025). The grid has felt it: 1,800 MW of data center load tripped on one normally-cleared fault in February 2025, NERC issued a Level 3 alert on May 4, 2026 with a ride-through standard due by year-end (Tech Times, Jul 2026), and ERCOT's NOGRR282 ride-through rule took effect August 1, 2026 (Compute Law Blog).

Rack power is climbing faster than the rules. The modal rack is about 9 kW and only 9% of operators run 50 kW or more (Uptime Institute, 2025); GB200 NVL72 draws about 120 kW (mgrid); Rubin Ultra Kyber is specified at 600 kW (The Register, Mar 2025); NVIDIA targets 1 MW racks from 2027. Above roughly 200 kW the 54 VDC busbar runs out of copper. 800 VDC carries 157% more power per conductor than 415 VAC in NVIDIA's October 2025 whitepaper (its May 2025 post said 85%), with 45% less copper, up to 5% better efficiency and about 30% lower TCO (NVIDIA, May 2025; NVIDIA whitepaper, Oct 2025), and first-half 2026 orders concentrated in Texas and Georgia (datacentres.com). The rack side is covered in 800 VDC Data Centers and Vera Rubin.

What does this mean for a 20–150 MW site in Texas?

Five things fall out of the numbers. First, 75 MW is the line: below it a site is outside Batch Zero and the energization pause. Second, substation headroom is worth more than land: about 18 months where capacity exists against 36–84 months for a new transmission-level connection. Third, bridge generation at this scale means reciprocating engines or aeroderivatives at 12–24 months, with 40–65% overbuild and, for 150 MW, roughly 21 MMcf/d of firm gas on the Enverus ratio. Fourth, design for the rules: NOGRR282 ride-through, SB6 firm load shed, 4CP management, and 15–40% generator coverage with 2–4 hours of batteries. Fifth, the TCEQ pause holds any engine above the 3.2 MW permit-by-rule threshold until issuance resumes. Grid power at an estimated $45–60/MWh against $100–165/MWh for on-site gas is why bridge power is a bridge, not a destination; the schedule scenarios are in Realistic Data Center Timelines in 2026.

Next read: 800 VDC Data Centers and Vera Rubin covers the rack side of this chain, from the 54 V busbar to the 1 MW rack.

Frequently asked questions

How is a data center connected to the grid?

Through a dedicated substation at 69–345 kV above about 20 MW, with two feeds typical at 100 MW; a new substation takes two to four years. In ERCOT, loads of 75 MW or more enter the SB6 process: a $100,000 study fee, $50,000 per MW of security, studies, then energization approval.

What is behind-the-meter power for a data center?

Generation on the data center's side of the utility meter: gas engines, turbines, fuel cells, or solar with batteries. More than 130 GW has been announced for US data centers, over 80% of it gas, at about $100–165/MWh against a $90–95/MWh grid benchmark (Rabobank).

How fast is on-site gas compared with a grid interconnect?

Reciprocating engines and aeroderivative turbines run within 12–24 months and fuel cells in about 90 days, against 36–84 months for a transmission-level interconnect or a heavy-duty turbine plant. A Texas minor NSR air permit averaged 241 days before the September 2026 pause.

Can data centers run on nuclear power today?

Only through grid-delivered contracts with existing plants: Clinton and Crane, the former Three Mile Island Unit 1, start supplying Meta and Microsoft in 2027, and Duane Arnold restarts in 2029. The first new-build reactor for a data center customer, Kairos Hermes 2 at 50 MW, targets 2030.

Why do Texas regulators curtail data centers?

Senate Bill 6 requires new transmission-voltage loads to join firm load shed and lets ERCOT call on-site backup covering at least 50% of demand. Co-located loads accept full curtailment within 30 minutes, as in the Goodnight Wind order, where the PUCT held that grid requirements supersede data center hardware concerns.

Sources

1. How power reaches an AI data center — amcompute, 2026-03-30
2. ERCOT large load interconnection in Texas — Compute Law Blog, 2026-10-04
3. Data center power distribution — TierPoint, 2025-12-11
4. Data center power: a comprehensive overview — DgtlInfra, 2026-03-28
5. Why AI data center projects face years of delays after approval — Data Center Knowledge, 2026-05-12
6. Data center power shifts favor BESS, solid-state transformers — pv magazine USA, 2026-07-15
7. Battery storage gains ground as data centers seek diesel alternatives — Data Center Knowledge, 2026-05-15
8. NVIDIA 800 V HVDC architecture will power the next generation of AI factories — NVIDIA, 2025-05-20
9. 800V DC power architecture: how NVIDIA's Kyber standard is rewiring the data center — datacentres.com, 2026-08-26
10. Uptime Institute data center PUE stagnation 2025 and liquid cooling — mgrid, 2025-10-01
11. Uptime Institute Global Data Center Survey 2025 — Uptime Institute, 2025
12. Explaining Uptime Institute's Tier classification system — Uptime Institute Journal, 2021
13. The sprint: data centers are building a parallel energy system in the US — Rabobank, 2026-05-06
14. 2025 State of the Market Report for the ERCOT Electricity Markets — Potomac Economics, 2026-05
15. Interconnection and Grid Analysis Update — ERCOT Board of Directors, 2026-04-20
16. Texas PUC approves data center audit — Utility Dive, 2026-08-21
17. Texas audit could delay 49.8 GW of data center load, cost projects up to $15 billion, BNEF warns — POWER, 2026-08-06
18. Texas pauses all environmental permitting for data centers — Beveridge & Diamond, 2026-09-29
19. Texas Senate Bill 6: understanding the impacts to large loads and co-located generation — Baker Botts, 2025-07-18
20. Texas 4CP: AI and bitcoin loads — The Energy Mag, 2026-06-11
21. Short-Term Energy Outlook, Table 7c: US regional electricity prices — EIA, 2026-10
22. Talen, Amazon launch $18B nuclear PPA: a grid-connected IPP model for the data center era — POWER, 2025-06-12
23. GE Vernova gas turbine backlog climbs to 116 GW — Utility Dive, 2026-07-23
24. Oracle taps VoltaGrid for 2.3 GW modular gas fleet to power AI data centers across Texas — POWER, 2025-10-15
25. Parker Hannifin to supply more than 1GW of natural gas turbines to Stargate's Abilene campus — DCD, 2025-10-28
26. OpenAI Stargate: where the US sites stand — Epoch AI, 2026-04-17
27. Musk's xAI gets go-ahead for 41 natural gas turbines in Mississippi to power Colossus data centers — DCD, 2026-03-11
28. World's largest private power grid: Fermi America's Matador campus — Interesting Engineering, 2026-02-26
29. Brookfield, Bloom Energy expand partnership to $25 billion framework for onsite data center power — Power Engineering, 2026-07-09
30. Why data centers are looking to natural gas for behind-the-meter power — Enverus, 2026-05-28
31. Air permitting for data centers (AAPCA presentation) — TCEQ Air Permits Division, 2026-04-07
32. Air permitting for data centers in Texas — Encino Environmental, 2026-09-03
33. Can solar microgrids power data centers? — Heatmap, 2025-01-07
34. Despite low gas prices, solar and wind remain cheapest sources of power in US (Lazard LCOE+ 2025) — pv magazine USA, 2025-06-17
35. Alphabet announces agreement to acquire Intersect to advance US energy innovation — Alphabet, 2025-12-22
36. PUCT affirms curtailment authority over co-located data centers in first net-metering approval — White & Case, 2026-08-05
37. Nuclear power for data centers, explained (project tracker) — Nuclear News Network, 2026-10-01
38. Enchanted Rock is selling utilities on a flexible data center connection — Latitude Media, 2024-05-24
39. Behind-the-meter data centers: Meta New Albany (Socrates) — Cleanview, 2026
40. Power stabilization for AI training datacenters (arXiv 2508.14318) — Microsoft, OpenAI and NVIDIA, 2025-08-20
41. AI data centers triggered 1,800 MW grid drop; NERC issues highest alert — Tech Times, 2026-07-04
42. NVIDIA charts course for 600 kW racks — The Register, 2025-03-19
43. Hourly peak load in ERCOT set a new record, exceeding 91 GW on July 22 — U.S. Energy Information Administration, 2026-08-03
44. 800 VDC Architecture for Next-Generation AI Infrastructure (whitepaper) — NVIDIA, 2025-10