top of page

Data Centre Firm Power 2026–2030: Which Technologies Can Deliver Electricity Fastest?

46 minutes ago
47 min read

By Green Fuel Journal Research & Intelligence Team — see our Editorial Standards and AI Usage Disclosure for how this report was produced and verified. Research date: 20 November 2026.


Last reviewed: November 2026 · Next review trigger: GE Vernova’s fourth-quarter 2026 results (backlog expected to exceed $200 billion)


Scope and disclaimer

This report is research and analysis. It is not legal, financial, investment, engineering or safety-certification advice, and it does not endorse any company. Figures stated by companies are labelled as company claims. No reader should rely on this report as the sole basis for a decision.

See the full Disclaimer and Terms. Further Green Fuel Journal research is available in the knowledge base.


GreenFuel Journal graphic: Data Centre Firm Power 2026-2030, 90% of large gas turbines booked for 2028, blue server racks.

Executive Summary

Data center power is limited more by delivery time than by price: PJM’s 2028/29 capacity auction cleared at the $325/MW-day cap and still fell 6,831 MW short of the reliability requirement, with only about 525 MW of new generation and uprates. Jefferies estimates that grid power for large loads is now 5–7 years away, and Ember and an Interface report put connection waits at 7–10 years in the main European hubs and up to 13 years in the most congested. On reported ranges, reciprocating engines (15–24 months) and aeroderivative turbines (18–36 months) fit a 24-month target while large-frame turbines (about 36 to 80 months) do not, although the engine and turbine figures rest mainly on one source, SemiAnalysis.


The fastest dated firm supply in the sources reviewed comes from existing or restarting plants and modular onsite systems, not new large turbines. Constellation targets the Crane nuclear restart for the second half of 2027. Bloom Energy claims whole-site fuel-cell power in 90 days, a vendor figure with no independent confirmation. GE Vernova’s backlog plus slot reservations reached 122 GW in Q3 2026, with slots sold into 2029–2030, and sources put large-frame lead times between about three and six years.


Faster supply costs more and shifts risk onto the operator. Lazard puts new combined-cycle gas at $1,450–2,100/kW, SemiAnalysis puts reciprocating engines at $1,700–2,000/kW and fuel cells at $3,000–4,000/kW, and the IEA says reliable onsite gas can need 30–70% more capacity than the load. Regulators are passing the risk to operators: 77 large-load tariffs were pending or in place across 36 US states, and a 2026 review of large-load tariff terms found a median minimum bill of 80% of contract demand. No published source quantifies the stranded-asset exposure of bridge power, and that gap is the central open question in this report.


1. Executive Intelligence Synthesis

Direct answer: On reported ranges, modular onsite units (reciprocating engines, aeroderivative turbines and fuel cells) are the fastest route to firm power for a data centre in 2026–2028, ahead of new large-frame turbines and large-load grid connections. Most of that evidence rests on one newsletter and one vendor claim, and the sources measure time from different starting points. No single ranking holds across regions, project sizes or permitting conditions, so the report gives ranges, clocks and evidence labels.


1.1 Five Executive Signals


Signal 1: Grid access sets the 2026–2028 timetable

  • FINDING: Jefferies, reported by Utility Dive on 24 February 2026, estimates that grid power for large loads is now 5–7 years away (single source, geography not stated), and Ember and an Interface report put connection waits at 7–10 years in the main European hubs (one evidence family).

  • SO WHAT: A project that waits for the grid inherits that delay whatever its generation costs.

  • NOW WHAT: File the grid application immediately and treat any onsite supply as a bridge with a defined end date.


Signal 2: Modular onsite equipment is the fastest reported route, on thin evidence

  • FINDING: SemiAnalysis reports reciprocating engines at 15–24 months and aeroderivatives at 18–36 months, Bloom Energy claims 90 days for a whole site, and Enverus groups engines, small turbines and fuel cells at 18–24 months to commercial operation.

  • SO WHAT: The ranges rest on a single newsletter, a vendor claim and a grouped estimate, and they measure different clocks.

  • NOW WHAT: Require suppliers to put a dated delivery and commissioning schedule in the contract before relying on any of these figures.


Signal 3: Large-frame turbines are sold out and demand is shifting to smaller units

  • FINDING: GE Vernova’s backlog plus slot reservations reached 122 GW at Q3 2026 (64 GW backlog and 58 GW of non-binding reservations), with slots sold into 2029–2030.

  • SO WHAT: A large-frame order placed now delivers at the end of the 2026–2030 window or after it.

  • NOW WHAT: Compare smaller units, used or refurbished turbines and contracts on existing plants before committing to a large-frame slot.


Signal 4: Regulation is moving the cost of speed onto the operator

  • FINDING: 77 large-load tariffs were pending or in place across 36 US states by March 2026, and LBNL and Brattle report a median minimum bill of 80% of contract demand and a median minimum term of 12 years for tariffs since 2025.

  • SO WHAT: Reserved capacity is paid for whether or not it is used.

  • NOW WHAT: Size the load request to what will be built and model the exit terms before signing.


Signal 5: Request volumes overstate real load

  • FINDING: ERCOT’s large-load queue exceeded 233 GW at the end of 2025 against NERC’s 23 GW of ERCOT data centre load by 2035, NESO holds 72.8 GW of requests against 5.2 GW expected by 2030 (the horizons differ), and AEMO reports 36% of the previous year’s projects cancelled.

  • SO WHAT: Queue size is not demand, so supply plans built on requests carry large forecast error.

  • NOW WHAT: Test every supply plan against a low-demand case.


1.2 How to read the evidence

  • FINDING: Time-to-power figures in the sources reviewed use different starting points and end points, and many rest on one source.

  • SO WHAT: Comparing them without their definitions produces a false ranking.

  • NOW WHAT: Read every range in this report together with its clock and its evidence label.


The report uses four clocks.

First power is the first electricity delivered to the site. Full contracted capacity is the date the full contracted megawatts are available. 24/7 firm supply is continuous supply with the redundancy the operator requires. Long-term scalable supply is the capacity that can keep growing over the asset’s life. The sources mostly report equipment lead time, order-to-delivery time or time to commercial operation. None states that its clock starts at investment approval, so no figure in this report should be read that way.


Every central figure carries one of five labels. Primary means a regulator, operator or official document. Two-source means two sources of independent lineage. Single-source means one source only. Vendor claim means a company’s statement about its own product. No data means nothing authoritative was found.


Methodology note: The evidence labels follow the source type and the number of independent sources found for each figure. A label of Two-source is applied only when the sources do not share a common origin. Where sources are likely to share data, such as the Ember and Interface wait-time figures for Europe, the report treats them as one evidence family. The labels are an editorial judgement about source quality. They are not a statistical confidence measure.


Table 1: Six questions, six answers

Question

Answer on the evidence reviewed

Which technology delivers 100 MW fastest in 2026–2030?

Conditional. Reported ranges for reciprocating engines (15–24 months) and aeroderivatives (18–36 months) fit a 24-month target. Large-frame turbines (about 36 to 80 months) and Jefferies’ estimate of 5–7 years for large-load grid power do not. Most figures are single-source.

How long does it take to deploy gas turbines for a data centre?

GE Vernova says about three years for new heavy-duty orders, Wood Mackenzie says six years and Enverus says up to 80 months. The sources measure different clocks and are not reconciled.

Are reciprocating engines faster than large gas turbines?

On reported ranges, yes: 15–24 months against 36 months or more. The engine range is partly corroborated by Enverus, which groups engines with small turbines and fuel cells.

What does 24/7 firm power cost for a 100 MW data centre?

Capex ranges only, in $/kW. No source reviewed gives a delivered cost per MWh for onsite options. Section 7 shows the capex arithmetic.

Can solar, wind and batteries alone supply a hyperscale data centre around the clock?

No verified data. The sources reviewed hold capex for solar and storage but no storage-duration, overbuild or microgrid time-to-power evidence for data centres.

Is onsite gas faster and more reliable than waiting for a grid connection?

Faster than the reported grid waits (Jefferies’ 5–7 years; 7–10 years in European hubs), on reported ranges. No source compares reliability, and the IEA says reliable onsite gas can need 30–70% more capacity than the load.


2. Macro Context & Strategic Drivers

Direct answer: Global data centre electricity use was about 415 TWh in 2024 and the IEA expects about 945–950 TWh by 2030, while an S&P Global 451 Research low case puts 2030 at 1,587 TWh from a much higher 2025 base. The forecasts do not agree, so planning needs a range. Grid access is tightest in the US PJM and ERCOT regions, the main European hubs, Singapore and Australia’s New South Wales.


2.1 Demand is a range, not a number

  • FINDING: The IEA puts data centre electricity at about 415 TWh in 2024 and about 945 TWh in 2030, while S&P Global’s 451 Research puts 2025 at 860 TWh, almost double the IEA’s 485 TWh for the same year.

  • SO WHAT: The gap is a difference in baseline and method, not a rounding error, so no central figure exists.

  • NOW WHAT: Plan against a range and ask any forecaster which baseline and definition it uses.


The IEA’s April 2025 Energy and AI analysis puts 2024 demand at about 415 TWh (about 1.5% of global electricity) and the 2030 base case at about 945 TWh (just under 3%). A later IEA publication, undated on the page reviewed, puts 2025 at 485 TWh and 2030 at about 950 TWh, with 17% growth in 2025. S&P Global Market Intelligence’s 451 Research, on 5 November 2025, put 2025 at 860 TWh and 2030 at 1,587 TWh in its low case. The report does not reconcile or average these figures. Goldman Sachs Research, on 10 November 2025, expected demand to rise 175% by 2030 against 2023 and gave no gigawatt or regional figures.


The IEA’s regional growth figures for 2024–2030 in its base case are about 240 TWh for the United States (+130%), about 175 TWh for China (+170%), more than 45 TWh for Europe (+70%) and about 15 TWh for Japan (+80%). No India figure appears on the IEA page reviewed. The S&P 451 low case gives North America above 386 TWh rising to about 755 TWh, Europe 145 TWh to 238 TWh and Asia-Pacific 267 TWh to 493 TWh between 2025 and 2030, with China and India not broken out.



Methodology note: Chart 1 plots the IEA and S&P Global 451 Research figures as published, each on its own baseline and definition. The series are not averaged or reconciled, and the S&P figure for 2030 is its low case.


2.2 Regional picture

  • FINDING: Every region with data reports a grid-access constraint, but the constraint takes different forms: capacity shortfall in PJM, connection queues in Europe, rationed allocations in Singapore and conditional connections in Ireland and Alberta.

  • SO WHAT: A single global timeline for firm power does not exist, and each site must be judged on its own jurisdiction.

  • NOW WHAT: Start with the jurisdiction’s instrument in Table 2 before choosing a technology.


North America.

The only official federal data-centre forecast reviewed is the DOE and LBNL report of 20 December 2024: 176 TWh in 2023 (about 4.4% of US electricity), rising to 325–580 TWh by 2028. No LBNL 2026 update was found. NERC’s Long-Term Reliability Assessment, released on 29 January 2026 and reported by POWER Magazine, projects 224 GW of ten-year summer peak growth and high reliability risk by 2030 in MISO, PJM, ERCOT, WECC-Basin and WECC-Northwest. It puts data centre load at 18 GW in MISO and 23 GW in ERCOT by 2035.


These NERC figures come from the POWER Magazine report, and the NERC document itself was not reviewed. PJM’s 2028/29 auction, announced on 14 July 2026, cleared 138,318 MW at the $325/MW-day cap and was 6,831 MW short of the reliability requirement. According to Utility Dive, only about 525 MW of new generation and uprates cleared. The cost was $16.4 billion against an estimated $29.7 billion without the price collar.


David Mills, President and CEO of PJM, said on 14 July 2026 that demand for electricity

“continues to grow faster than electricity supply.”

BloombergNEF expects US data centre demand to reach 106 GW by 2035, reported by Utility Dive on 3 December 2025. EIA’s September 2026 outlook expects total US consumption to grow about 2% in each of 2026 and 2027 and does not split out data centres. No Canada-wide federal, British Columbia or Quebec policy was found.


EU and UK.

The IEA expects European data centre load to grow by more than 45 TWh (+70%) between 2024 and 2030. Ember puts European demand at 96 TWh in 2024, 168 TWh in 2030 and 236 TWh in 2035 (published 19 June 2025), and the absolute 2030 figure rests on Ember alone. Average connection waits of 7–10 years in the main hubs, up to 13 years in the most congested, come from Ember and an Interface report relayed by Euronews on 5 May 2026.


The two likely share underlying data. NESO expects 5.2 GW of UK data centre demand by 2030 against about 72.8 GW of connection requests to 2039, and about 59 GW is awaiting a transmission connection. Ireland’s energy regulator, the CRU, issued its large energy users policy on 12 December 2025 (CRU2025236). According to a DLA Piper summary, sites of 10 MVA and above must bring dispatchable onsite or nearby generation or storage matching 100% of their de-rated connection capacity.


The CRU decision itself was not reviewed. At EU level, the Commission proposed the Cloud and AI Development Act (COM(2026) 502) on 3 June 2026, and a Linklaters summary reports that it adopted a data centre sustainability rating regulation in September 2026. A Grids Package (COM(2026) 600) is also reported, but sources conflict on its date, so none is given here. No Eurostat, JRC, ENTSO-E, ACER or DESNZ data centre series was found.


China.

Forecasts of 2030 data centre load range widely: Rystad Energy 289 TWh (2.3% of demand), Wood Mackenzie 479 TWh (July 2025), a State Council research body 400 TWh and Goldman Sachs close to 600 TWh. The IEA gives growth of about 175 TWh (+170%). No source agrees on a central figure. The National Development and Reform Commission and the National Energy Administration issued notice No. 650 (published 30 May 2025) and notice No. 688 (published 20 May 2026) on green power direct connection. Notice 688 names computing facilities as a priority and caps grid export at 20%, with minimum self-consumption of 60% of generation and 30% of electricity use. TechNode reported on 11 November 2026 that only 13.6% of tracked projects were operating and 71.6% were in planning, a single-source figure whose originating agency is not named. The official texts of East Data West Computing and the 15th Five-Year Plan energy plan were not reviewed.


India.

Section 3 covers India in full.


Australia.

AEMO’s 2026 Electricity Statement of Opportunities, released on 25 August 2026, forecasts data centre use of about 5 TWh in FY26, about 15 TWh in FY30 and about 34 TWh in FY36, equal to about 3%, 7.8% and 13% of forecast operational consumption in the National Electricity Market. Its forecasting overview reports a pipeline of 225 projects totalling 67 GW (the media release cites 40 GW), and AEMO forecasts no reliability gaps before 2030.


Singapore and Malaysia.

Singapore’s second data centre call allocated 200 MW on 21 August 2026, 50 MW each to Digital Realty, Equinix, Keppel DC and STT GDC, with each applicant committing to more than 50% of capacity from green-energy pathways. A law-firm guide, the only source, gives grid connection lead times of 26 months for a 66 kV connection to an existing substation and 28 months where a new substation is needed, and does not limit the figures to data centres. Malaysia reportedly raised its connected load charge from RM8.50 to RM89.64 per kW and has 117 approved data centres with about 23,500 MW of maximum demand. Both Malaysian items come from one trade-press report of 12 November 2026 that gives no effective date, and neither has been confirmed by the utility.


Brazil and the Gulf.

Brazil’s Redata tax regime (PL 278/2026), sanctioned on 15 September 2026, requires all contracted electricity to come from clean or renewable sources. No Brazilian load forecast or connection data was found. In the Gulf, the sources reviewed hold announcements only: HUMAIN in Saudi Arabia plans 250 MW by end-2026, 1.7 GW in 2027 and about 3 GW by 2030, with the power source not stated. No Gulf regulator demand forecast, connection data, tariff or regulatory instrument was found. Japan, South Korea, Indonesia, Africa and Latin America outside Brazil are not covered beyond the IEA’s Japan growth figure.


Table 2: Regulatory instruments by region

Region

Instrument

Date

What it does

Status

US federal

FERC show-cause proceedings EL26-67 to EL26-72

18 June 2026

Large-load interconnection tariffs preliminarily found unjust and unreasonable across six grid operators

Proceedings open; law-firm summary only

US federal

FERC PJM co-location order, EL25-49-002

18 June 2026

Accepted three co-location transmission products; PJM’s proposed 1 June 2029 start needs more support

Further compliance due; law-firm summary only

US, Texas

Senate Bill 6; PUCT proposed large-load rule

June 2025; March 2026

Sites above 75 MW face site control, screening fee and financial security; ERCOT may direct backup generation or curtailment

Law binding; PUCT rule proposed, final adoption pending

Canada, Alberta

Data Centre Regulation, AR 117/2026

9 June 2026

75 MW and above must bring own generation; bridging limited to three years; bridged load curtails first

In force; phase 2A provisional

Canada, Ontario

Data Centre Playbook, ERO 026-0853

13 August 2026

New data centres pay full energy cost including upgrades

Draft

Ireland

CRU large energy users connection policy, CRU2025236

12 December 2025

Dispatchable onsite or nearby generation or storage as a connection condition

Binding; secondary source

EU

Cloud and AI Development Act, COM(2026) 502

3 June 2026

Data centre acceleration zones and priority for strategic projects

Proposal

China

NDRC and NEA notice No. 688

14 May 2026 (published 20 May)

Computing facilities prioritised for multi-user green power direct connection; 20% grid-export cap

Official notice

India

NITI Aayog Data Centre Evaluation Framework

October 2026 (month from file path)

Power readiness conditions for states; demand response above 20 MW

Advisory

Australia

Expectations of data centres and AI infrastructure developers

About March 2026

Bring own clean energy or storage and pay full connection share

Non-binding; date from press

Singapore

Second Data Centre Call for Application

21 August 2026

200 MW allocated with green-energy conditions

Provisional

Brazil

Redata special tax regime, PL 278/2026

15 September 2026

Tax relief tied to 100% clean contracted electricity

Sanctioned

Gulf

None found

Not applicable

No data centre power instrument was found

No data

Regulatory approaches differ most on who supplies firm power. Ireland and Alberta make onsite or own generation a condition of connection. Australia and Singapore attach green-energy expectations. Texas lets the grid operator direct backup generation or curtailment. China and Brazil use access or tax relief as the incentive. Reservation discipline appears in several places: Malaysia’s charge, Singapore’s tranches, Texas security requirements and AEP Ohio’s minimum bills all aim at speculative demand. For a comparable view of capital allocation and risk in another power asset class, see Offshore Wind Investment Intelligence 2026–2030, Part I.


2.3 Requests against expected load

  • FINDING: The gap between connection requests and expected load is wide in every region that publishes both: 233 GW against 23 GW in ERCOT, 72.8 GW against 5.2 GW in the UK and 67 GW of AEMO pipeline projects with 36% of the previous year’s projects cancelled.

  • SO WHAT: Requests include speculative and duplicate demand, so grid queues and forecasts built on them overstate real load.

  • NOW WHAT: Treat any forecast built on queue data as an upper bound and check whether a tariff or fee has filtered the requests.


The ERCOT queue figure is request volume, not expected load, and NERC’s own ERCOT data centre figure is 23 GW by 2035. NESO does not expect all of its 72.8 GW of requests to connect. AEMO reports grid connection utilisation of about 27%. AEP Ohio’s large-load forecast fell from 30 GW to 13 GW after its tariff was approved, a change the Smart Electric Power Alliance reported through Utility Dive on 31 March 2026 but did not attribute to the tariff alone. In India the government’s estimate of AI data centre load for FY32 is reported to have risen from 13.56 GW in a March 2026 figure to 26.3 GW in a July 2026 figure. Section 6.3 sets these figures in chart form.


3. India-Specific Analysis

Direct answer: India’s data centre load is estimated at 13.56 GW by FY32 in a reported March 2026 government figure and 26.3 GW in a July 2026 figure, while installed capacity is reported at 950–1,700 MW depending on the source. Firm-power evidence is limited to a SECI round-the-clock award at INR 5.25–5.26/kWh, a battery storage transmission waiver and advisory guidance. No binding central rule specific to data centre power was found.


3.1 Demand estimates conflict

  • FINDING: CBRE, relayed by Business Today on 1 April 2026, puts installed capacity above 1,700 MW, while CBRE, relayed by ThePrint on 9 April 2026, puts operational capacity at about 950–1,500 MW.

  • SO WHAT: The two figures differ by more than 700 MW at the top end, and the definitions behind them are unverified.

  • NOW WHAT: Use a range for current capacity and ask any adviser which definition it applies.


Business Today reports 440 MW added in 2025 (+160%), about 500 MW expected in 2026 and Mumbai holding more than 50% of operational inventory. The CBRE original was not reviewed. The Straits Times, in a report carried by The Star on 8 May 2026, gives 1.5 GW now and 13.2 GW by 2032 “per MeitY estimates” and counts data centres by city: Mumbai 61, Hyderabad 33, Delhi-NCR 31, Bengaluru 31 and Chennai 30.


On future load, the Minister of State for Power, Shripad Naik, said in a written parliamentary reply reported on 29 July 2026 that AI data centres would add 26.3 GW of load by 2031–32. The press report describes that load as primarily served by renewable energy capacity. Inc42 and AngelOne relay the same reply, so the figure is single-source, and no Press Information Bureau text was found. The earlier March 2026 estimate is reported as 13.56 GW.


ThePrint, citing the Central Electricity Authority’s National Generation Adequacy Plan, reports data centre load of more than 13.6 GW by 2032 and 16.4 GW by 2040, with installed capacity of 520.5 GW on 31 January 2026 against a requirement of 1,121 GW by 2035–36. The CEA document was not reviewed. Andhra Pradesh plans 5 GW of data centre capacity by 2030. Grid readiness is a limit: only 42% of a targeted 15,253 circuit-km of transmission has been commissioned.


Green Fuel Journal’s earlier analysis of grid reliability under rapid renewable expansion is in India’s Renewable Energy Surge: Can the Power Grid Stay Reliable Amid Rapid Expansion?


3.2 Instruments

FINDING: India has advisory guidance and enabling measures but no binding central rule specific to data centre power.

SO WHAT: Operators carry connection and tariff risk through state and distribution-company decisions that vary by location.

NOW WHAT: Negotiate power terms with the state distribution company early and record the conditions in writing.


The SHANTI Act, explained by the Press Information Bureau on 19 December 2025, opens nuclear power to private participation, sets a goal of 100 GW by 2047 and targets five indigenous small modular reactors by 2033. Nuclear supply is therefore outside the 2026–2030 window. The Ministry of Power’s transmission-charge waiver gives battery storage at renewable substations a 100% waiver if commissioned by 30 June 2028 (Press Information Bureau reply of 17 March 2026).


NITI Aayog’s Data Centre Evaluation Framework, dated October 2026 by its file path, asks states for a power readiness and ramp-up plan, automated demand response above 20 MW, hourly matching and firm renewables with storage. It is advisory. Mercom reported on 15 October 2026 that NITI Aayog also asked for model green tariff and open access principles. The Union Budget of 1 February 2026 gave foreign cloud companies a tax holiday, to 2047 according to most outlets and to 2045 according to Entrepreneur India. The two dates conflict, and the Budget document was not reviewed.


3.3 Firm-power evidence and cases

  • FINDING: SECI awarded a 1 GW round-the-clock tender at INR 5.25/kWh to five bidders and INR 5.26/kWh to two, with at least 90% peak-hour delivery, as PV Tech reported on 10 August 2026.

  • SO WHAT: This is a price signal for firm renewable supply in India, but it is not data-centre-specific and names no hyperscaler buyer.

  • NOW WHAT: Use it as a procurement benchmark and check the commissioning terms in the award documents, because at least one award carries a 24-month commissioning period.


At least one award in the round, Juniper Green’s 230 MW, carries a power purchase agreement of about 25 years. No Indian battery storage, coal or gas tariff for data centres was found. For an Indian example of dispatchable renewable supply,


TCS announced a 1 GW HyperVault campus in Hyderabad in September 2026, and the news report names only “green energy” as the power source. CtrlS has a target of 4.4 GW by 2030 and nearly 100 MW of captive solar at Nagpur, according to Analytics India Magazine on 16 November 2026. Rahul Dhar, President of Global Data Centre Operations, Enterprise Colo Delivery, Programme Management and Energy R&D at CtrlS, said on 16 November 2026 that operators are “bulk consumers of power” and called for faster approvals and a shorter forecasting period for power requirements. Jitesh Patil, CMD of TSSPDCL, Telangana’s southern power distribution company, told ThePrint on 9 April 2026 that “data centres suck power suddenly” and warned that surges of up to half a gigawatt could trip the grid through voltage fluctuation. No verified power figures were found for Reliance, Yotta, NTT, Sify or AdaniConneX.


3.4 What the evidence cannot answer

  • FINDING: No state-by-state megawatt table, no tariffs by city and no connection times for Mumbai, Chennai, Hyderabad or Bengaluru are available in the sources reviewed.

  • SO WHAT: A 100 MW cost and speed comparison across the four cities cannot be built from verified data, and this report does not attempt one.

  • NOW WHAT: Obtain the state tariff orders and distribution-company connection data directly before choosing between cities.


That comparison belongs in a dedicated India report. No IEA, CEEW, CSEP or JLL figures for Indian data centres, and no text of state data centre policies, were found. US and European lead times must not be applied to India, because the grid, tariff and permitting conditions differ.


4. Operational / Technical Deep-Dive

Direct answer: For a 24-month target, the reported ranges favour reciprocating engines (15–24 months) and aeroderivative turbines (18–36 months) over large-frame turbines (about 36 to 80 months) and Jefferies’ 5–7-year estimate for large-load grid power. Bloom’s 90-day fuel-cell claim is unconfirmed. The ranking depends on region, project size, permitting and the starting point of the clock, so no single order holds. Supplier schedules should be tested against the clock each source uses.


4.1 Defining time-to-power

  • FINDING: The sources reviewed measure equipment lead time, order-to-delivery time, time to commercial operation or time to first power, and none starts from investment approval.

  • SO WHAT: A turbine lead time says nothing about permitting, gas supply, interconnection or commissioning, which all add time.

  • NOW WHAT: Ask for the full schedule from investment approval to 24/7 firm supply and compare options on that basis.

SemiAnalysis reports that air permitting can take a year or more even in Texas, a single-source figure. No source reviewed covers gas supply infrastructure, which the project must still secure.


4.2 Time-to-power matrix

  • FINDING: On reported ranges, modular onsite units and existing or restarting nuclear plants have the shortest times, while large-frame gas and large-load grid connections have the longest.

  • SO WHAT: The shortest times rest on single sources and vendor statements, so the order is provisional.

  • NOW WHAT: Use Table 3 to screen options against a target month and then test the shortlist with supplier schedules.


Table 3: Reported time to power by technology

Technology

Reported range

Clock stated by source

Evidence label

Notes

Fuel cells (Bloom)

Whole-site power within 90 days (Bloom); installation and commissioning in weeks (SemiAnalysis); 18–24 months (Enverus)

Bloom: site power; SemiAnalysis: installation and commissioning; Enverus: commercial operation

Vendor claim; single-source; grouped estimate

Sources conflict; Enverus groups fuel cells with engines and small turbines

Reciprocating engines

15–24 months (SemiAnalysis); 18–24 months (Enverus)

Lead time; commercial operation

Two-source, partial

Enverus grouping is not engine-specific

Aeroderivative turbines

18–36 months; small units about 12 months; installation 2–4 weeks after delivery

Lead time

Single-source

SemiAnalysis only

Industrial gas turbines

12–36 months; used or refurbished under 12 months

Lead time

Single-source

SemiAnalysis only

Large-frame gas (combined cycle)

About 3 years (GE Vernova, new orders); 6 years (Wood Mackenzie); up to 80 months (Enverus)

New-order slot; order to delivery; order to commercial operation

Conflicting

Not reconciled; SemiAnalysis adds about 2 years to build after turbine delivery

Nuclear restart

Crane: second half of 2027; Palisades: power purchase agreements require service by March 2027; Duane Arnold: early 2029

Restart date

Two-source for Crane

Palisades fuel loading paused (Boston Globe, 16 October 2026)

New nuclear and SMRs

No verified US or European time before 2030; India targets five indigenous SMRs by 2033

Target date

No data (Europe, US)

India: official target, not a delivery record

Geothermal

Fervo and Google 396 MW online 2028; Cape Station phase 1 (100 MW) slated for 2026

Target online date

Company-announced

Phase 1 delivery not verified

Solar plus storage microgrids; hydro

No verified data

Not applicable

No data

Capex for solar and storage is in Table 4

Large-load grid connection

Jefferies: grid power for large loads 5–7 years away; European hubs 7–10 years (up to 13); Singapore 26 months (66 kV, existing substation) to 28 months (new substation)

Connection wait

Jefferies: single-source; Europe: one evidence family; Singapore: single-source

Jefferies does not state a geography; the Singapore figures are not data-centre-specific; no verified Australian figure

Methodology note: Table 3 reports each range as the named source states it, with the clock that source used. It does not average, rank or convert ranges. Comparing a range to a 12, 24, 36 or 60-month horizon is arithmetic on the reported range and is not a project forecast. No source reviewed contains project-level data for a 100 MW delivery, so any statement about 100 MW in this report is illustrative.


Methodology note: Chart 2 plots each source’s range as a separate bar, coloured by evidence label, so conflicts remain visible. Dated milestones for Crane and Fervo are excluded because they are dates, not durations.


4.3 Technology by technology

  • FINDING: Large-frame turbines are the bottleneck, and smaller equipment is absorbing the demand that large turbines cannot meet before 2029.

  • SO WHAT: The choice between technologies is also a choice between supply chains with different order books.

  • NOW WHAT: Check the supplier’s booked capacity for the year you need power, not only its published lead time.


Large-frame gas. GE Vernova’s gas backlog was 44 GW at Q1 2026 excluding reservations. Its backlog plus slot reservations was 116 GW at Q2 and 122 GW at Q3, made up of a 64 GW backlog and 58 GW of non-binding reservations, and at least 130 GW was expected by year-end. Wood Mackenzie counts 110 GW of global orders at the end of 2025 against 60–70 GW of manufacturing capacity and expects turbine prices of $600/kW by the end of 2027, up 195% since 2019.


Aurora Tenorio, Senior Analyst for Supply Chain at Wood Mackenzie, wrote on 1 April 2026 of

“six-year lead times and order books sold through 2027.”

Scott Strazik, CEO of GE Vernova, said on 23 April 2026 that gas turbines are

“really not the gating item”

in a three-year project cycle. The two views are not reconciled. Jefferies, reported by Utility Dive on 24 February 2026, puts large-frame manufacturers over 90% booked for 2028 and over 70% for 2029, and expects at least 19 GW of equipment capacity available by 2028, 49 GW by 2029 and 76 GW by 2030.

Smaller gas units. Enverus, reported by POWER Engineering on 5 October 2026, finds that reciprocating engines, small and medium turbines and fuel cells make up 61% of behind-the-meter gas capacity. It expects 25.5 GW of new US industrial demand to be served behind the meter through 2030, of which 22.5 GW (88%) is data centres.


The IEA expects about 15–27 GW of onsite gas for data centres by 2030, mostly in the US, and notes that about one-fifth of tracked US projects had begun land clearing or construction. Jefferies puts behind-the-meter equipment makers about 65% booked for 2028 and about 33% for 2029.


Fuel cells. Bloom Energy’s fuel cells carry the highest reported capex among the gas-based options at $3,000–4,000/kW (SemiAnalysis, single-source), and the 90-day speed is Bloom’s own claim. The company reports 60% lower-heating-value efficiency per DOE. Section 5.4 covers the Brookfield and AEP agreements.


Nuclear. Constellation’s Crane restart is expected in the second half of 2027 (Section 5.2). Lazard puts new nuclear at $12,300–17,570/kW, while EIA’s 2023-dollar figures are $7,861/kW for two large reactors and $8,936/kW for a six-module small reactor plant. The two sources differ by a factor of 1.5 to 2 for large nuclear because the vintage and basis differ. No source reviewed gives a US or European small-reactor time-to-power before 2030.


Geothermal. Lazard puts geothermal capex at $5,135–6,635/kW and LCOE at $67–111/MWh, figures Lazard labels illustrative, against EIA’s 2023-dollar binary plant figure of $3,963/kW. Fervo and Google signed a 396 MW agreement on 2 September 2026 with an option for about 600 MW more, and the deal is scheduled to come online in 2028. Enhanced geothermal cost was not found. Green Fuel Journal’s analysis of the technology is in Next-Generation Geothermal Investment 2026–2030, Part I and Part II.


Solar plus storage. Lazard puts utility solar capex at $1,250–1,850/kW and four-hour storage at $270–365/kWh, with solar-plus-storage LCOE of $61–156/MWh. EIA’s 2023-dollar battery figure is $1,744/kW ($436/kWh) and predates the 2025–26 price changes. The IEA expects 20–25 GW of data centre batteries by 2030. No microgrid time or cost evidence for data centres was found. See Clean Power in 2026: Why Solar + Storage Are Now Beating Coal and Gas and Global Solar Market Slowdown 2026–2027 for supply-side context.


Grid connection. Table 3 gives the reported waits. Pamela MacDougall, head of energy markets and regulation at AWS EMEA, said in an interview reported on 3 February 2026 that “certainty of the delivery date has continued to be delayed” across Europe. The same report puts a data centre build at about two years and grid power at up to seven.


Flexible load. The Duke Nicholas Institute, reported by Utility Dive on 11 February 2025, estimated 76 GW of new flexible load at 0.25% curtailment and 126 GW at 1% across 22 large US balancing authorities. A 256-GPU cluster in Phoenix cut power 25% for three hours in a preprint of 1 July 2025 whose peer-review status is not stated. EPRI’s UK test reached up to 40% reduction on a cluster out of production, and Google reports 1 GW of demand response in utility contracts (POWER Magazine, 19 March 2026). None shows firm flexibility at campus scale.


4.4 Cost benchmarks and conflicts

  • FINDING: Reported capex for onsite options runs from $1,100/kW for gas peaking to $17,570/kW for new nuclear, and the fastest modular options sit at the upper end of the gas range.

  • SO WHAT: The fastest options are not the cheapest per kilowatt, and the capex figures are US-based or of unstated geography.

  • NOW WHAT: Convert capex to cost per firm kilowatt after overbuild before comparing options.


Table 4: Reported capex and LCOE by technology

Approach

Capex ($/kW)

LCOE ($/MWh)

Source status

Combined-cycle gas

Lazard 2026: 1,450–2,100 (2,400–2,600 for early-stage projects with post-2028 operation); BNEF via TechCrunch: 2,157 (2025); EIA 2023 dollars: 868–921

51–129 (Lazard)

Capex: three sources that conflict; LCOE: single-source

Gas peaking and aeroderivative

Lazard peaking: 1,100–1,650; SemiAnalysis aeroderivative: 1,700–2,000; EIA 2023 dollars: 1,606

144–276 (Lazard peaking)

Two sources overlap only at the edges

Industrial gas turbines

1,500–1,800 (SemiAnalysis)

Not found

Single-source

Reciprocating engines

1,700–2,000 (SemiAnalysis)

Not found

Single-source

Fuel cells (Bloom)

3,000–4,000 (SemiAnalysis)

Not found; Lazard does not cover fuel cells

Single-source

Geothermal

Lazard: 5,135–6,635 (labelled illustrative by Lazard); EIA 2023 dollars: 3,963

67–111 (Lazard)

Capex: two sources that differ by vintage; LCOE: single-source

New nuclear

Lazard: 12,300–17,570; EIA 2023 dollars: 7,861 (large) and 8,936 (small modular)

175–255 (Lazard)

Sources conflict by vintage and basis

Solar plus storage

Lazard solar: 1,250–1,850; storage: 270–365/kWh (four-hour); EIA 2023 dollars battery: 1,744/kW

61–156 (Lazard solar plus storage)

Capex: two sources, different vintage

Methodology note: Table 4 reproduces each source’s figures with its basis and vintage and does not adjust for inflation, scope or location. EIA’s figures are in 2023 dollars and were published in January 2024, before the 2025–26 equipment price increases, so they understate current procurement cost. The capex figures are US-based or of unstated geography, and no European, Chinese or Indian benchmark for onsite or grid-supplied firm power was obtained, apart from India’s auction tariffs. Lazard’s own 2026 figures show gas LCOE at a 15-year high.



Several conflicts remain open. Turbine lead times range from about three years (GE Vernova) to six years (Wood Mackenzie) and up to 80 months (Enverus), probably because they measure order-to-delivery, order-to-operation and new-order slot differently. Combined-cycle capex runs from Lazard’s $1,450–2,100/kW to BNEF’s $2,157/kW, which sits just above Lazard’s range and below its $2,400–2,600 early-stage figure, while EIA’s older figure is far lower. SemiAnalysis puts aeroderivative cost above Lazard’s gas-peaking range, and the scopes may differ. The overbuild factor is 30–70% in the IEA’s view and 10–20% for modular fleets and up to 50% for large-frame units in Enverus’s view. Wood Mackenzie and Jefferies disagree on whether equipment availability improves before 2030, and Section 8.2 sets out both views.


5. Named Company Case Studies

Direct answer: Operators have secured firm power through four routes: dedicated gas plants (Meta in Louisiana), nuclear restarts and contracts (Constellation’s Crane), geothermal power purchase agreements (Fervo with Google) and onsite fuel cells (Bloom). Outside the US, China Datang’s Ningxia direct supply and AirTrunk’s Hunter project show the renewable route, and Microsoft’s exit from Loughton shows the grid risk. Most figures are company claims or trade-press reports.


5.1 Meta Hyperion, Louisiana

  • FINDING: Meta is paying for 10 gas plants for its Hyperion campus (7 new and 3 approved in 2025), totalling 7.5 GW of gas plus up to 2.5 GW of renewables and storage, at about $11 billion (Fortune, 27 March 2026).

  • SO WHAT: The scale shows what a developer builds when the grid cannot serve a campus, and the $11 billion figure is a single-source estimate.

  • NOW WHAT: Read the contract term, because the plants sit on a 15-year Entergy contract and Louisiana regulators still had to approve the new plants.


A joint venture with Blue Owl carries up to $27 billion of total development cost. Fortune raised the question of ratepayer risk after the contract term if load falls.


5.2 Constellation, Crane restart

  • FINDING: Constellation expects the Crane restart (formerly Three Mile Island 1) in the second half of 2027, with the Nuclear Regulatory Commission’s final environmental assessment issued in September 2026 (ANS, 12 November 2026).

  • SO WHAT: Existing nuclear plants offer the only dated firm supply in the sources reviewed that does not depend on turbine slots.

  • NOW WHAT: Treat restart dates as targets and track regulatory milestones, because Palisades’ fuel loading was paused.


The Boston Globe reported on 16 October 2026 a restart cost of $1.6 billion with a $1 billion DOE loan. Palisades’ power purchase agreements require service by March 2027, and Duane Arnold is targeted for early 2029.


5.3 Fervo Energy and Google, Cape Station

  • FINDING: Fervo and Google signed a 396 MW geothermal agreement on 2 September 2026, with an option for about 600 MW more and online dates from 2028.

  • SO WHAT: Geothermal moves from announcement to contract but not to delivery before 2028.

  • NOW WHAT: Count geothermal toward 2028 and later supply, not the 2026–2027 gap.


Tim Latimer, CEO and co-founder of Fervo, said on 2 September 2026 that the agreement shows enhanced geothermal

“is ready to power the next generation of computing infrastructure.”

This is a company statement. Cape Station phase 1 (100 MW) is slated to deliver in 2026, and delivery has not been verified.


5.4 Bloom Energy, Brookfield and AEP

  • FINDING: Brookfield raised its framework with Bloom Energy from $5 billion to $25 billion on 9 July 2026, and AEP agreed in November 2024 to up to 1 GW of fuel cells.

  • SO WHAT: Capital is committing to fuel cells even though the 90-day delivery claim is Bloom’s own and the cost is the highest among gas-based options.

  • NOW WHAT: Test the 90-day claim through reference sites and contract terms before building a schedule on it.


5.5 China Datang, Ningxia

  • FINDING: China Datang’s Zhongwei project in Ningxia began direct supply to a data centre cluster in February 2026 and entered formal operation in early May 2026 (The Next Web, 25 June 2026).

  • SO WHAT: It shows a dedicated renewable supply running ahead of the grid, within the framework of notice 688.

  • NOW WHAT: Note that direct connection caps grid export at 20% and does not itself supply firm capacity.


Phase one is 2 GW (500 MW of solar, 1.5 GW of wind and storage) at about 8.7 billion yuan, with expected output of 4.3 TWh a year against forecast consumption of 2.29 TWh and four dedicated 110 kV lines. Phase two would reach 4.6 GW. Al Jazeera described a 500 MW wind and solar project on 28 May 2026, a different scale that this report does not reconcile. Anders Hove, senior research fellow at the Oxford Institute for Energy Studies, said on 28 May 2026 that China’s power system is dispatched “mainly at the provincial level.”


5.6 AirTrunk, Kurri Kurri, Australia

  • FINDING: AirTrunk secured exclusivity on more than 1 GW of Hunter-region projects on 12 November 2026, including 260 MW of solar and more than 800 MWh of battery storage (operational 2030) and up to 350 MW of wind (2032).

  • SO WHAT: Its own renewables arrive after 2030, so firm power for 2026–2030 must come from the grid.

  • NOW WHAT: Plan grid supply for the first phase and treat the renewables as later firming.


The campus consumes 550.5 MW per state planning records and connects at 132 kV through Ausgrid. Robin Khuda, Founder and CEO of AirTrunk, said on 12 November 2026 that

“digital growth must bring new energy with it.”

This is a company statement reported by PV Tech.


5.7 Microsoft and Nscale, Loughton, UK

  • FINDING: Microsoft reportedly withdrew from Nscale’s 50 MW Loughton site, scalable to 90 MW, over power supply in October 2026, and UK Power Networks reportedly cannot supply the site until the early-to-mid 2030s.

  • SO WHAT: It is the clearest case in the sources reviewed of a project stalling on grid delivery, though the claim rests on press reports that UK Power Networks has not confirmed.

  • NOW WHAT: Confirm the utility’s dated connection commitment before committing capital to a site.


ITPro, citing the Telegraph, reported the exit on 19 October 2026, and Euronews, citing The Morning Intelligence newsletter, reported the supply timing on 29 September 2026. Microsoft said the decision related to a single site, and Nscale said the grid issue did not affect its other contracts. The original completion date was early 2027.


Table 5: Case comparison

Company

Project

Firm-power route

Key figure

Timeline

Label

Meta

Hyperion, Louisiana

Dedicated gas plants

7.5 GW gas; about $11 billion (estimate)

Announced 27 March 2026

Reported; cost single-source

Constellation

Crane restart

Nuclear restart

$1.6 billion; $1 billion DOE loan

Second half of 2027

Reported; timing two-source

Fervo and Google

Cape Station, Utah

Geothermal agreement

396 MW

2028

Company claim

Bloom Energy

Brookfield and AEP deals

Onsite fuel cells

$25 billion framework; up to 1 GW (AEP)

2024–2026

Company claim

China Datang

Zhongwei, Ningxia

Wind, solar and storage direct supply

2 GW phase one; 8.7 billion yuan

Operating from May 2026

Reported

AirTrunk

Kurri Kurri, NSW

Solar, storage and wind after 2030

More than 1 GW exclusivity; 550.5 MW campus

2030 and 2032

Reported

Microsoft and Nscale

Loughton, Essex

Grid supply

50 MW scalable to 90 MW

Early 2027 originally; early-to-mid 2030s per report

Reported; utility has not confirmed


6. Friction, Risk & Systemic Bottlenecks

Direct answer: Five problems are most likely to delay data centre power before 2030: sold-out large turbines, reported grid waits of 5–10 years in the regions with data, speculative demand that distorts forecasts, grid stability rules for sudden load swings, and unquantified stranded-asset risk from bridge power. No published source resolves the last, and it bears most on operators who choose bridge power. Sections 6.1 to 6.5 take each problem in turn.


6.1 Equipment

  • FINDING: The IEA reports tight gas turbine and transformer supply chains, and GE Vernova’s slots are sold into 2029–2030.

  • SO WHAT: Equipment delays affect every route that needs generators, switchgear or transformers.

  • NOW WHAT: Secure transformers and switchgear slots at the same time as generation.


Wood Mackenzie reported in July 2025 that China’s domestic transformer lead time is 48 weeks against 143 weeks in the US, a single-source figure. No verified Australian equipment lead-time figure was found.


6.2 Grid access

  • FINDING: Connection waits reach 7–10 years in European hubs and 26–28 months in Singapore, Jefferies puts grid power for large loads 5–7 years away, and only 42% of India’s targeted transmission has been built.

  • SO WHAT: Where the wait exceeds the project timeline, onsite power is the only route to opening on schedule.

  • NOW WHAT: Ask the utility for a dated connection commitment and treat undated offers as the long case.


Kristi Hobbs, vice president of system planning and weatherization at ERCOT, said at a board meeting on 9 December 2025 that

“we have outgrown the process that was established for reviewing these large loads.”

Section 5.7 describes the UK Power Networks case.


6.3 Phantom demand and forecast quality

  • FINDING: Forecast ranges are wide within each source family: China’s 2030 figures run from 289 TWh to nearly 600 TWh, and the IEA and S&P 451 differ by almost a factor of two for 2025.SO WHAT: A supply plan sized to one forecast can be badly wrong.

  • NOW WHAT: Set the supply plan against both the low and high forecasts and stage commitments.


Section 2.3 gave the request-to-load gaps. China adds another signal: only 13.6% of tracked projects are operating. Chart 4 compares ERCOT and NESO requests with expected load.



Methodology note: Chart 4 compares connection requests with expected load as each source reports them. The ERCOT and NESO figures cover different horizons (2035 and 2030) and different load definitions, so the chart shows the size of each gap and does not compare the two regions with each other.


6.4 Grid stability

  • FINDING: ERCOT’s ride-through rules for large loads (NOGRR 282 and NPRR 1308) took effect on 1 August 2026, and the IEA says server load can swing by more than 50% of rated capacity within a second.

  • SO WHAT: Sudden load changes can trip protection and affect other customers, so grid operators now set technical conditions.

  • NOW WHAT: Build ride-through capability and ramp limits into the electrical design from the start.


Patil’s warning in Section 3.3 shows the same concern in India.


6.5 Critical unresolved issue: bridge power and who bears stranded-asset risk

  • FINDING: Onsite gas can arrive faster than a grid connection, but no published source quantifies who pays if the grid arrives early or the load never does.

  • SO WHAT: The speed advantage of bridge power has an unpriced liability attached to it, and the liability falls on operators, ratepayers or both depending on the contract.

  • NOW WHAT: Model the exit cost and the useful bridge window before committing to any onsite supply.


On speed, large-frame equipment is the bottleneck, as Section 4.3 showed, and the market is shifting to smaller units. Behind-the-meter gas also needs extra capacity to be reliable (Section 4.4), air permitting can take a year or more, and no source covers gas supply infrastructure. Jefferies calls behind-the-meter gas a bridge solution that might last up to ten years. If grid power for large loads is 5–7 years away, as Jefferies also estimates, the useful bridge window may be only a few years. That conclusion is this report’s inference from the two figures, not a sourced one.


On risk allocation, regulators are pushing cost onto the large load through the tariff terms in Table 6, including collateral of $1.5 million per MW at Dominion and exit fees of 36 months of minimum charges at AEP Ohio. Ratepayers keep some exposure. Meta’s Hyperion plants sit on a 15-year Entergy contract, and Fortune raised ratepayer risk after the term if load falls. In PJM’s 2028/29 auction the price collar held the cost to $16.4 billion against an estimated $29.7 billion uncapped, a difference of $13.3 billion (arithmetic on the two reported figures). Outside the US, regulators use reservation charges (Malaysia), capacity tranches (Singapore), full connection-cost allocation (Australia, Ontario and Alberta) and bridging caps (Alberta’s three-year limit and proposed 1,600 MW system cap) instead of quantified estimates.


Green Fuel Journal’s analysis of a related stranded-asset question in shipping is in LNG as a Transition Fuel 2026–2030: Will LNG Ships Become Stranded Assets?


Several questions remain open. No source quantifies stranded-asset exposure for behind-the-meter gas or grid upgrades, and no peer-reviewed study was found. The disagreement between Jefferies and Wood Mackenzie on whether turbine lead times ease from 2027 is not settled. Whether GE Vernova’s slot reservations convert to orders is unknown. Whether queue requests reflect real projects is contested. The rules that would clarify risk may arrive late: PJM’s co-location products have a proposed start of 1 June 2029, after the 2026–2028 window in which most of the gap opens.


Julia Hoos, Head of USA East at Aurora Energy Research, said on 15 July 2026 that

“the current system doesn’t work to bring online new capacity or stimulate demand response.”

The Jefferies note, the Wood Mackenzie report and an E3 rate study are cited through secondary reports and the originals were not reviewed, and the outcome of PJM’s backstop auction is unverified.


7. Capital & Investment Implications

Direct answer: Capex for fast modular options runs from $1,700 to $4,000/kW on SemiAnalysis’s single-source estimates, and after overbuild each firm kilowatt costs more than the headline figure. Operators carry minimum-bill, term, collateral and exit costs under large-load tariffs that median 80% of contract demand and 12 years. Capital is flowing into gas, nuclear, geothermal and fuel cells, but most announced sums are company claims or single-source estimates. Section 7.1 shows the capex arithmetic for 100 MW.


7.1 What the capex ranges imply

  • FINDING: After the overbuild factor, 100 MW of firm onsite power from modular gas costs about $187–240 million in plant capex on the reported ranges, and fuel cells cost about $330–480 million.

  • SO WHAT: The overbuild factor adds 10–20% to modular plant capex before any other cost, and fuel cells cost about 1.8 to 2 times as much as engines per kilowatt on these ranges.

  • NOW WHAT: Apply the calculation below to your own load, overbuild factor and capex quotes.


WORKED EXAMPLE: sizing and pricing 100 MW of firm onsite power


Illustrative assumptions (assumed for this example, not sourced): firm load of 100 MW (100,000 kW); target for first power of 24 months from order.

Sourced inputs: overbuild factor of 1.1–1.2 for modular units and 1.5 for large frame (Enverus); capex ranges and time ranges as in Tables 3 and 4 (SemiAnalysis, Lazard, Enverus).

Option

Installed capacity (MW)

Capex ($/kW)

Plant capex ($ million)

Capex per firm kW ($)

Reported time against 24-month target

Reciprocating engines

110–120

1,700–2,000

187.0–240.0

1,870–2,400

15–24 months: fits

Aeroderivative turbines

110–120

1,700–2,000

187.0–240.0

1,870–2,400

18–36 months: fits only at the short end

Fuel cells

110–120

3,000–4,000

330.0–480.0

3,300–4,800

90 days (vendor claim) or 18–24 months (Enverus): fits

Large-frame combined cycle (Lazard capex)

150

1,450–2,100

217.5–315.0

2,175–3,150

About 36 to 80 months: does not fit

Methodology note: The 100 MW load and the 24-month target are illustrative inputs chosen for this example. The overbuild factors, capex ranges and time ranges are the sourced figures, each with the label shown in Tables 3 and 4. All results are arithmetic on those inputs. They are not quotes from a supplier, not forecasts, and not a delivered cost per MWh, which no source reviewed contains.


Calculation: installed capacity equals 100 MW multiplied by the overbuild factor. Plant capex equals installed kilowatts multiplied by the capex per kilowatt, and capex per firm kilowatt equals plant capex divided by 100,000 kW. For reciprocating engines at the low end, 110,000 kW × $1,700 = $187.0 million, and at the high end 120,000 kW × $2,000 = $240.0 million. Using the IEA’s 30–70% overbuild range instead, installed capacity becomes 130–170 MW and plant capex becomes $221.0–340.0 million. The table excludes fuel, operating cost, financing, permitting, gas supply infrastructure, transformers and any redundancy beyond the overbuild factor.


7.2 Who carries the risk

  • FINDING: Large-load tariffs and fees shift stranded-asset risk to the operator through minimum bills, long terms, collateral and exit charges.

  • SO WHAT: A 12-year median term and an 80% median minimum bill make early exit or lower-than-planned load expensive.

  • NOW WHAT: Size contract demand to the load you will actually draw and model the exit cost.


Table 6: Reported large-load cost-allocation terms

Term

Reported value

Source

Minimum billing demand

Median 80% of contract demand (examples range from 60% to 100%)

LBNL and Brattle, August 2026

Minimum contract term

Median 12 years since 2025 (5 years before)

LBNL and Brattle, August 2026

Collateral and exit fees

$1.5 million per MW (Dominion); 36 months of minimum charges (AEP Ohio)

LBNL and Brattle, August 2026

AEP Ohio tariff

85% minimum demand, up to 12 years including a four-year ramp, projects above 25 MW

PUCO order, 9 July 2025, via POWER Magazine

Alberta

C$15,000 per MW commitment fee

Bennett Jones, 9 July 2026

Texas

Screening fee of at least $100,000; draft rule proposes $50,000 per MW security and a non-refundable $50,000 per MW interconnection fee

Foley, 31 March 2026 (draft rule)

Malaysia

Connected load charge RM8.50 to RM89.64 per kW (reported)

w.media, 12 November 2026 (unverified)

7.3 Capital signals

  • FINDING: Announced capital is large but mostly company-stated: Brookfield’s $25 billion Bloom framework, Meta’s estimated $11 billion for Hyperion gas plants and GE Vernova’s $197 billion company-wide backlog.

  • SO WHAT: The sums show direction, not delivered capacity.

  • NOW WHAT: Separate operating, contracted and announced capacity before using any of these figures in a model.


BNEF, reported by TechCrunch on 27 April 2026, put combined-cycle plant cost at $2,157/kW in 2025, up from under $1,500/kW in 2023, with build time 23% longer, a single-source figure. Wood Mackenzie reports that turbines make up 20–30% of combined-cycle project cost.

Strazik said on 29 October 2026 that GE Vernova expects its backlog “to exceed $200 billion” in the fourth quarter, and H1 2026 gas orders were priced 10–20 points above Q4 2025. Fuel cost, financing cost and operating cost are not covered by the sources reviewed.


8. Future Scenarios & Forecast, 2026–2035

Direct answer: Three scenarios frame the 2026–2030 outcome: a bridge-power era that lasts to the end of the decade, equipment relief by 2028–2030, and maturing flexibility and co-location rules. Each is defined by observable signposts, and none carries a probability. Evidence beyond 2030 is thin, so the report uses only the longer-horizon figures in the sources reviewed, and readers should treat each scenario as a prompt for planning, not a prediction.


8.1 Scenario design

  • FINDING: No source reviewed assigns probabilities to supply, equipment or rule outcomes.

  • SO WHAT: Any probability weight would be Green Fuel Journal’s judgement presented as data.

  • NOW WHAT: Use the signposts in Table 7 to see which scenario is unfolding and adjust commitments as evidence arrives.


The longer-horizon figures available are Ember’s 236 TWh for Europe in 2035, BloombergNEF’s 106 GW for the US in 2035, AEMO’s 34 TWh for FY36, NERC’s ten-year view and India’s nuclear goals to 2047.


8.2 Three scenarios

  • FINDING: Scenario A holds that equipment and connection constraints persist to the end of the decade; Scenario B holds that equipment supply eases from 2028; Scenario C holds that flexibility and co-location rules mature.

  • SO WHAT: Each implies a different value for bridge power and a different exit date.

  • NOW WHAT: Choose commitments that still work in the scenario you consider least likely.


Scenario A,

a bridge-power era to 2030. Wood Mackenzie expects turbine lead times to stay long near term and fall from 2027 with little effect on additions until 2030 or later. Jefferies’ 5–7-year estimate for large-load grid power and PJM’s proposed 1 June 2029 co-location start fit this path. Onsite supply carries a longer useful life and a greater exposure to fuel and permitting.


Scenario B,

equipment relief by 2028–2030. Jefferies expects at least 19 GW of available gas equipment capacity by 2028, 49 GW by 2029 and 76 GW by 2030, and GE Vernova’s CEO describes a three-year cycle. Under this path the premium paid for fast modular units falls and large-frame units become available again.


Scenario C,

maturing flexibility and co-location rules. Duke’s modelling of 76 GW of flexible load and Google’s 1 GW of demand response in contracts (Section 4.3) are the evidence for this path. FERC’s proceedings, Texas SB 6 curtailment powers and ERCOT’s ride-through rules would define the terms. Demonstrations so far are small, so this path depends on evidence that is not yet available.


Table 7: Scenario signposts

Signpost

Latest reading

Points toward

GE Vernova backlog and reservations

122 GW at Q3 2026, of which 58 GW non-binding

A if reservations convert and lead times stay long; B if they lapse

PJM capacity auction shortfall

6,831 MW short for 2028/29

A if shortfalls persist

FERC co-location compliance

Further compliance due after the 18 June 2026 order; proposed start 1 June 2029

C if filings accepted early

Crane restart

Second half of 2027

Supports firm supply from existing plants in all scenarios

AEMO pipeline conversion

225 projects, 67 GW; 36% of the prior year cancelled

B if cancellations continue

Texas large-load rule

Draft proposed March 2026; final adoption pending

C if curtailment terms finalised

Methodology note: The scenarios are structured analyst judgement. They carry no probability weights because the sources reviewed provide no basis for any. Each scenario was defined by one named source’s expectation or by a dated regulatory milestone, and the signposts are the latest readings available. The scenarios are not forecasts, and they are not mutually exclusive.


9. Strategic Recommendations

Direct answer: Operators who need firm power before 2029 should file the grid application now, use modular onsite supply as a dated bridge, and size contracts to expected load. Investors should separate operating, contracted and announced capacity and discount single-source and vendor figures. Regulators should publish queue statistics and align co-location rules before 2029. None of this is advice for a specific project. Section 9.4 gives the first contract step for an executive team.


9.1 Industry

  • FINDING: Jefferies’ 5–7-year estimate for large-load grid power and European hub waits of 7–10 years mean most developers cannot rely on the grid alone before 2029.

  • SO WHAT: The choice is how long and how large a bridge to build.

  • NOW WHAT: Take five steps.


  • First, file the grid application immediately and ask for a dated commitment.

  • Second, request supplier schedules that state the clock and include permitting and commissioning.

  • Third, size onsite capacity with the overbuild factor, as in Section 7.1.

  • Fourth, negotiate minimum-bill and exit terms against a low-demand case.

  • Fifth, define the date at which onsite supply will be retired or repurposed.


9.2 Investors

  • FINDING: Most announced sums and speed claims in the sources reviewed are company-stated or single-source.

  • SO WHAT: Capital allocated on announced figures carries delivery risk.

  • NOW WHAT: Apply three tests.


  • First, separate operating, contracted and announced capacity for every project.

  • Second, discount vendor claims, such as Bloom’s 90-day figure, until reference sites confirm them.

  • Third, check whether a contract carries minimum-bill, collateral and exit terms, and watch whether GE Vernova’s slot reservations convert to orders.


9.3 Policymakers

  • FINDING: Regulators in Malaysia, Singapore, Texas and Ohio are already filtering speculative demand, and PJM’s co-location products have a proposed start of 1 June 2029.

  • SO WHAT: Rules that arrive after 2028 miss the period when most of the supply gap opens.

  • NOW WHAT: Prioritise four actions.


  • First, publish queue statistics that separate requests from expected load.

  • Second, bring co-location and large-load rules forward to before 2029.

  • Third, adopt reservation discipline, such as security requirements or fees.

  • Fourth, commission an estimate of stranded-asset exposure from behind-the-meter gas, because no published source provides one.


Table 8: Decision matrix for firm power options

Option

Time to first power (reported)

Cost evidence

Equipment availability

Regulatory exposure

Stranded-asset exposure

Reciprocating engines

15–24 months; two sources, partial

$1,700–2,000/kW; single-source

Behind-the-meter makers about 65% booked for 2028

Air permit up to a year or more; single-source

Bridge window unquantified

Aeroderivative turbines

18–36 months; single-source

$1,700–2,000/kW; single-source

Same booking figure as above

Same permit note

Unquantified

Fuel cells

90 days (vendor claim) to 18–24 months

$3,000–4,000/kW; single-source

No supplier capacity data

No data

Unquantified

Large-frame gas

About 36 to 80 months; conflicting

$1,450–2,157/kW across sources

Slots sold into 2029–2030

Air permit as above

Longer asset life; unquantified

Existing or restarting nuclear

Crane second half of 2027

$1.6 billion restart cost; no $/kW

Not applicable

Regulatory milestones pending

Long-term contract risk

Geothermal

Online 2028 for the Fervo and Google deal

$5,135–6,635/kW (Lazard, illustrative)

No data

No data

Not assessed

Large-load grid connection

Jefferies: 5–7 years (geography not stated); European hubs 7–10 years

Tariff terms in Table 6

Not applicable

Large-load tariffs and fees

Minimum-bill exposure

Methodology note: Table 8 is structured analyst judgement. Each cell restates the reported evidence for that criterion, with its label, and does not carry a score or star rating. The criteria are time to first power as reported, cost evidence, equipment availability, regulatory exposure and stranded-asset exposure. A cell reads “no data” or “unquantified” where the sources reviewed have nothing to report. The matrix does not rank the options and does not weight the criteria.


9.4 The first decision for the executive team

  • FINDING: The sources measure delivery on different clocks, and none quantifies the exit cost of bridge power.

  • SO WHAT: A contract signed without a stated clock and an exit cost leaves the two most uncertain variables with the operator.

  • NOW WHAT: Before signing any onsite power contract this quarter, require the supplier to state in writing the delivery date, the clock it measures from and the exit cost if grid power arrives earlier.


10. Executive FAQ


Which technology can deliver 100 MW of firm power to a data center fastest between 2026 and 2030?

Reciprocating engines and aeroderivative turbines fit a 24-month target on reported ranges of 15–24 and 18–36 months. Large-frame turbines (about 36 to 80 months) and Jefferies’ estimate of 5–7 years for large-load grid power do not. Most of these figures are single-source, and Bloom’s 90-day fuel-cell claim is unconfirmed, so the answer depends on region, size and permitting.


How long does it take to deploy gas turbines for data center power?

About three to six years for new large-frame turbines on reported figures. GE Vernova says about three years for new heavy-duty orders, Wood Mackenzie says six years and Enverus says up to 80 months, and the three measure different clocks. GE Vernova’s slots are sold into 2029–2030.


Are reciprocating engines faster to deploy than large gas turbines?

Yes, on reported ranges. SemiAnalysis reports 15–24 months for engines against about 36 months or more for large frame, and Enverus corroborates 18–24 months for engines grouped with small turbines and fuel cells. The engine figures are partly single-source.


Can onsite power really be faster than waiting for a grid connection?

Yes, on reported ranges, because Jefferies puts grid power for large loads 5–7 years away and European hub waits are reported at 7–10 years. Onsite gas needs 10–70% more capacity than the load according to Enverus and the IEA, and air permits can take a year or more. No source compares the reliability of the two routes.


What does 24/7 firm power for a data center cost through 2030?

Published sources give capex only: about $1,700–2,000/kW for engines and aeroderivatives and $3,000–4,000/kW for fuel cells (SemiAnalysis, single-source), and $1,450–2,100/kW for large-frame combined cycle (Lazard). After the overbuild factor, 100 MW of modular gas costs about $187–240 million in plant capex. No delivered cost per MWh for onsite options was found.


Can solar, wind and batteries alone keep a hyperscale data center running around the clock?

No verified evidence answers this yet. The sources reviewed hold capex for solar and four-hour storage but no data centre microgrid time, storage-duration or overbuild evidence. The IEA expects 20–25 GW of data centre batteries by 2030.


11. Research Limitations, Scope and Legal Disclaimer

Limitations of the evidence

Several central figures in this report rest on a single source, and several primary documents were reviewed only through secondary summaries. Those documents are the NERC Long-Term Reliability Assessment, the FERC orders in dockets EL26-67 to EL26-72 and EL25-49-002, the Irish CRU decision, the EU Cloud and AI Development Act and Grids Package texts, the UK AI Growth Zones paper, India’s Press Information Bureau texts for the 26.3 GW and 13.56 GW replies, China’s East Data West Computing and 15th Five-Year Plan energy plan texts, and the Australian Government’s expectations page. Not every source page was re-checked against its live version before publication.


SemiAnalysis is the only source for onsite equipment capex and lead times. Lazard’s LCOE figures, BNEF’s combined-cycle cost and Jefferies’ equipment bookings and grid-power estimate each reach this report through one source. India’s 26.3 GW load estimate rests on one parliamentary reply relayed by the press. Rystad is the only source for China’s capacity figures, and Ember is the only source for Europe’s absolute 2030 demand. Malaysia’s connected load charge comes from a single trade-press report and needs verification. Publication dates for the IEA’s later energy and AI publications, the Rystad China page, the NESO evidence and the NITI Aayog framework were not stated on the documents.


Several data points were searched for and not found. These include IEA absolute regional totals and any IEA India figure, global 2030 figures from BNEF and McKinsey, interconnection wait statistics by US grid operator, Gate 2 outcomes for UK data centres, data centre electricity series from Eurostat, the Joint Research Centre, ENTSO-E and ACER, and consumption statistics from China’s NEA, NDRC and CAICT. No non-US cost benchmarks were found except India’s auction tariffs, and no source gives a delivered cost per MWh for onsite options, a solar-plus-storage microgrid cost or time for data centres, a small-reactor time-to-power before 2030, a quantified stranded-asset estimate, a Gulf regulatory instrument or Brazilian load data. Company cases outside those in Section 5 were not covered, including xAI, Stargate at Abilene, Talen, Baidu, ByteDance, State Grid and CGN. No in-window quotes from named European Commission, CRU, EirGrid, NESO or Ofgem officials, hyperscaler chief executives, DOE officials or named Chinese officials were found. Japan, South Korea, Indonesia and Africa were not covered. Readers should confirm any load-bearing figure on the original source before relying on it.


12. References & Strategic Sources


This report is backed by authoritative research, institutional analysis, industry intelligence, and strategic data sources.


© 2026 Green Fuel Journal, a publication of Sekason Research Limited. All rights reserved. Do not reproduce without written permission. Short excerpts with a link to GreenFuelJournal.com are permitted.

bottom of page