Plugin Capital Research : 8 November 2026 · Data cut-off 5 November 2026 ·
This is the third in our 2026 infrastructure sequence. From Innovation to Imperative argued that power scarcity, not capital, had become the binding constraint. The Global Neocloud Top 20 showed that the real dividing line between compute providers is the trio of power contracts, financing cost and customer credit. This article asks the next question: if the grid cannot deliver on schedule, should the data centre owner build the power plant — and what does that actually take?
Seven conclusions up front
- Demand is not the risk; the delivery date is. The IEA’s Base Case takes global data centre electricity from ~415 TWh in 2024 to ~945 TWh in 2030 — roughly 15% a year, more than four times the growth of total electricity demand. The updated 2026 vintage puts 2025 at ~485 TWh and 2030 at ~950 TWh. These are two different vintages and should not be blended.
- The binding constraint is time, not capital. The IEA counts more than 2,500 GW of renewables, storage and large-load projects stalled in connection queues worldwide, and estimates that around 20% of planned data centre projects could be delayed if grid constraints are not addressed.
- “Off-grid” is mostly a financing and permitting construct, not an electrical one. Of 38 cases we verified, only about three credibly qualify as islandable. Just 13 of 38 (34%) have any capacity energised or under construction; 25 (66%) are still paper.
- Equipment availability, not fuel preference, decides the technology. GE Vernova’s gas-equipment backlog grew from 62 to 83 GW during 2025; Siemens Energy and Mitsubishi are quoting 2029–2031 delivery. Reciprocating engines arrive in 12–24 months. That single fact explains why the fastest projects use mobile turbines, engines and fuel cells rather than heavy frames.
- Reliability is the hidden tax. Power remained the largest single cause of impactful data centre outages (45% in 2025, 54% in 2024), and 54% of significant outages cost more than $100,000. An islandable microgrid must reproduce black start, spinning reserve, N+1/2N redundancy and fuel firmness without a utility behind it.
- The world’s electrons are still mostly fossil, and energy is not capacity. Fossil fuels supplied 57.4% of global generation in 2025 even as renewables (33.8%) overtook coal (33.0%) for the first time in the modern era. With 2025 US fleet capacity factors of 24.4% for solar and 91.0% for nuclear, the cheapest energy is the hardest capacity.
- For APAC investors, the investable surface is the interface, not necessarily the turbine. Powered land, interconnection position, long-lead equipment slots, air permits and the financing layer are where durable value sits. Regulators from Dublin to Putrajaya are now turning “bring your own power” into a condition of connection.
I. The demand shock is now a delivery problem
The headline number is well known but still worth stating precisely. Data centres consumed about 415 TWh in 2024 — 1.5% of world electricity — having grown roughly 12% a year since 2017. The IEA’s Base Case projects ~945 TWh by 2030, just under 3% of global electricity and slightly more than Japan’s total consumption today; a 2035 extension reaches ~1,200 TWh. Accelerated servers (the AI workload) grow ~30% a year, against ~9% for conventional servers, and account for almost half the net increase. A typical AI-focused data centre draws as much power as 100,000 households; the largest under construction today will use twenty times that. (IEA, Energy and AI, Executive Summary; Energy demand from AI)
Two calibre warnings before the regional numbers. First, 415 TWh is consumption; the same IEA analysis models about 460 TWh of generation supplying data centres in 2024, a different measure that includes losses. Second, the IEA’s 2026 update puts 2025 at ~485 TWh and 2030 at ~950 TWh, with five named hyperscalers’ capex exceeding $400bn in 2025 and expected to rise ~75% to ~$715bn in 2026. (AEInsights, citing IEA Key Questions on Energy and AI; Energy Central, citing IEA Electricity 2026) We use the 2025 vintage for the regional table below because it is the only one with a published, region-disaggregated 2024/2030 series.
| FIGURE 1 – Global data centre electricity consumption by region, 2024 vs 2030 (TWh/yr) Source: IEA (2025), Energy and AI, Base Case; regional table reproduced in UNDP TIDE Centre, The Energy Sprint of the AI Race (May 2026), Table 2. World totals 416 / 946 TWh are the source’s own. Regional rows sum to 409.6 / 934.2 TWh; the residual (6.4 / 11.8 TWh, ~1.2–1.5%) is not allocated by the source and is shown here as ‘Other + residual’. 2024 is an estimate, 2030 a projection; do not blend with the IEA’s updated 2026 vintage of ~485 TWh for 2025. |
Data behind Figure 1 (TWh/yr, IEA Base Case, via UNDP TIDE Centre Table 2)
| Region | 2020 | 2023 | 2024 | 2030 | 2024→2030 |
| World | 269 | 361 | 416 | 946 | +127% |
| North America | 112 | 158 | 187 | 434 | +132% |
| — of which United States | 108 | 154 | 183 | 426 | +133% |
| Asia Pacific | 93 | 128 | 150 | 378 | +152% |
| — of which China | 62 | 84 | 102 | 277 | +172% |
| Europe | 57 | 66 | 68 | 113 | +66% |
| Central & South America | 1.5 | 1.5 | 1.7 | 3.3 | +94% |
| Middle East | 1.1 | 1.3 | 1.5 | 3.0 | +100% |
| Africa | 1.1 | 1.3 | 1.4 | 2.9 | +107% |
Regional rows do not sum to the source’s world total (409.6 vs 416 TWh in 2024; 934.2 vs 946 TWh in 2030). The ~1.2–1.5% residual is not allocated by the source; we have not silently normalised it. (UNDP TIDE Centre, The Energy Sprint of the AI Race, May 2026, Table 2)
North America: concentration, not just growth
In 2024 the United States accounted for 45% of global data centre electricity, China 25%, Europe 15%. The US adds roughly 240 TWh by 2030 (+130%) and will, by the end of the decade, consume more electricity for data centres than for aluminium, steel, cement, chemicals and all other energy-intensive goods combined. Data centres represent nearly half of all US electricity-demand growth to 2030 — and nearly half of US data centre capacity sits in just five regional clusters. (IEA Executive Summary; Sustainability Online, summarising IEA)
The market signals are consistent with a physical shortage rather than a cyclical one. PJM’s capacity auction cleared at $333/MW-day in December 2025 against $29 two years earlier; PJM estimated it would have approached $530 without the regulatory cap. In Northern Virginia, CBRE reported only 10.8 MW of available capacity against a construction pipeline above 2,400 MW. (RaboResearch interview, C-Suite Newsletter; secondary market reporting compiled in New Market Pitch)
| Calibre warning. US data centre electricity figures circulating in 2026 come from different vintages and methods (LBNL bottom-up estimates, IEA modelled totals, vendor and bank forecasts ranging to 420–728 TWh in 2030). We use the IEA/UNDP series throughout this post and do not average across methods. |
EMEA: the constraint is the connection agreement
Europe is the slowest-growing of the three big regions in percentage terms (68 → 113 TWh, +66%) but the most instructive on policy. EU data centre capacity is expected to rise from ~12 GW in 2025 to ~28 GW in 2030, and AI-focused colocation signings reached 420 MW in H1 2026 against 89 MW in H1 2025. JLL puts FLAP-D live capacity at ~3.8 GW with 1.4 GW under construction and 2 GW planned, and powered land at €2.26m per MW in 2026, up 82% from €1.24m in 2021. Greenfield sites have risen from 8% to 39% of the 2026–2028 European pipeline, and the average hyperscale greenfield project is now ~175 km from a hub city, up from 46 km. (New Polis, citing CBRE Q2 2026 and JLL EMEA Mid-Year 2026; People’s Daily, citing European Commission and IEA)
The regulatory response is where EMEA becomes a template. Ireland’s data centres went from 5% of metered electricity in 2015 to ~23% in 2025; EirGrid effectively froze new Dublin connections from 2021/22 until 2028. In December 2025 the CRU reopened the door on conditions: a large energy user must bring generation or storage on-site or nearby matching its maximum demand, cover at least 80% of annual consumption with additional renewable projects built in Ireland, and accept that constrained parts of the network may simply be unavailable. Singapore’s 2019–2022 moratorium was lifted only for best-in-class efficiency, with a later call offering at least 200 MW to applicants below PUE 1.3 and 50% green power. The Netherlands bans most new hyperscale facilities above 70 MW; TenneT’s North Holland high-voltage grid has no room for new large users until about 2032. Germany requires PUE ≤1.2 for new data centres from July 2026 and replaced first-come-first-served queueing with a project-readiness test in April 2026. The UK’s NESO reported 72.8 GW of data centre connection requests to 2039. (SDEA; Johns Hopkins, A State Playbook for Managing Data Center-Driven Load Growth; Data Center Dynamics on Amsterdam; Australia Data Centre Index on European and NSW grid rules)
Read that list again: in each case the regulator is no longer asking for a connection fee. It is asking the data centre to bring generation, storage, flexibility and efficiency as the price of admission. Off-grid power is ceasing to be a workaround and becoming a licence condition.
APAC: the fastest growth, and the most immediate policy pivot
China adds ~175 TWh by 2030 (+170%); Japan ~15 TWh (+80%); South-East Asia’s data centre electricity demand is projected to double by 2030 on growth in Singapore and Malaysia. (Sustainability Online, summarising IEA)
Our own region supplies the clearest numbers. As of June 2026, Tenaga Nasional’s potential electric service agreements (ESAs) reached 13 GW — 8.35 GW of ESAs plus 4.6 GW approved by the Data Centre Task Force. Of the 8.4 GW secured pipeline (5.65 GW connected, 2.2 GW under construction, 0.5 GW newly signed), Johor accounts for 67%. Actual data centre load doubled in twelve months, from 603 MW in June 2025 to 1,256 MW in June 2026 (+108%), and data centres now take 9.28–9.3% of Malaysia’s electricity — with MBSB Research suggesting up to 35% of Peninsular demand by 2035 and the ministry projecting 21 GW of DC demand by 2040. (TNB newsclip, The Star, 1 October 2026; The Star, 15 September 2026)
Three APAC facts matter most for this thesis. One, grid bottlenecks in Johor are already prompting TNB to accelerate upgrades and steer developers to the Klang Valley and East Coast; Wood Mackenzie identifies 132 kV main intake substations as the specific pinch point, with DC maximum demand of ~3.8 GW against state demand of ~2.6 GW. Two, the Green Lane Pathway compresses connection timelines from 36–48 months to about 12 — proof that a large part of the queue is administrative. Three, and most consequentially, the Malaysian government is studying the potential use of on-site power plants to support DC developments, while the DCTF now requires proof of anchor tenants before an ESA is executed and Budget 2027 allocates RM15bn to grid strengthening. (TNB newsclip; DC Market Insights, citing TNB Q2 FY2026 briefing and Wood Mackenzie; Southeast Asia Green Economy Report 2026)
| Evidence note. The 13 GW “potential ESA” figure is a queue of applications and approvals, not metered load; actual connected load was 1,256 MW. Reserved capacity and actual draw differ by roughly 4.5×. Any valuation built on the 13 GW number without an utilisation ramp is valuing a queue position, not a power asset. |
II. Why the grid cannot answer on that timescale
The IEA’s Electricity 2026 devotes a chapter to grids as the emerging bottleneck. More than 2,500 GW of renewable, large-load and storage projects are stalled in queues worldwide. Annual grid investment, currently about $400bn, needs to rise roughly 50% by 2030; prices for key grid components have nearly doubled in five years. The timing mismatch is stark: planning, permitting and completing new grid infrastructure takes 5–15 years, against 1–5 years for solar and wind, 1–3 years for data centres and 1–2 years for EV charging. The IEA estimates that ~20% of planned data centre projects could face delays if grid constraints are not addressed. (IEA, Electricity 2026 – Grids; The Electricity Hub on the 20% estimate)
The counter-evidence matters too, and it is often omitted. The same IEA chapter estimates that 1,200–1,600 GW of advanced-stage projects could be unlocked without building new lines — 750–900 GW through conditional non-firm connection agreements, the rest through dynamic line rating, advanced power-flow control, topology optimisation, reconductoring and voltage uprating. FERC Order 2023 and similar reforms are being implemented, but LBNL says it is too early to measure their effect. The queue is partly administrative. A project that assumes the grid will never improve is as poorly underwritten as one that assumes it will. (IEA, Electricity 2026 – Grids)
The US evidence is the most granular. Berkeley Lab’s Queued Up: 2026 Edition (data through end-2025) counts ~8,200 active projects, 1,312 GW of generation and ~749 GW of storage = 2,061 GW, down 10% year on year — but the decline came from withdrawals (>750 GW) outrunning new requests (~600 GW), while only ~53 GW reached commercial operation. Only 13% of capacity requesting interconnection between 2000 and 2020 had reached COD by end-2025; 75% was withdrawn. The median IR-to-COD duration for projects completed in 2025 was over five years, up from ~22 months for 2000–2008 builds. Notably, gas capacity in the queue rose 86% to 253 GW in 2025 while solar, wind and storage fell — a leading indicator of the fuel shift discussed in Section V. (LBNL, Queued Up: 2026 Edition)
| FIGURE 2– Time-to-power ladder: build, equipment, interconnection and grid lead times (months) Sources: IEA, Electricity 2026 – Grids (5–15 years for new grid infrastructure; 1–3 years for data centres); Lawrence Berkeley National Laboratory, Queued Up: 2026 Edition (median IR-to-COD >5 years for projects reaching COD in 2025); RaboResearch, The Sprint (May 2026) for reciprocating engines, aeroderivatives and solar/wind; GE Vernova / Siemens Energy / Mitsubishi results via trade press for heavy-duty delivery slots; industry procurement estimates for large power transformers. These are different objects — procurement lead times, full project timelines and queue durations — plotted together only to show order of magnitude. Transformer figures are trade estimates, not official statistics. |
| Calibre warning on Figure 2. These bars mix procurement lead times, full project schedules and queue-wait statistics. Large power transformer lead times of 36–60 months come from industry procurement trackers and trade press, not from an official statistical series; treat them as indicative. SMR “early 2030s” reflects RaboResearch’s estimate that less than 2 GW of US SMR capacity reaches commercial operation before 2033. |
The transformer bottleneck deserves its own line because it is the item most often missed in project models. Trade sources report power-class (50–200 MVA, 138 kV) units at 36–48 months and high-voltage (>200 MVA, 230 kV+) units at 42–60 months, with Tier 1 suppliers effectively closed for 2028 delivery and order books into 2030. Workarounds now in active use include spec-building transformers before a site is committed, refurbished units from decommissioned plants, mobile transformer leases and modular substations. (GridReadiness lead-time tracker; Nexus Power, Transformer Lead Times in 2026) Each workaround is, in effect, an option on a future energisation date — which is the asset class this article is about.
III. What “off-grid” actually means
The phrase is used loosely enough to be useless in an investment committee. Four distinct things get conflated:
- Physical islanding. Is there a grid connection at all, and can the site operate with it open?
- Ownership of generation. Does the off-taker own the plant, or merely buy its output?
- Ownership of interconnection rights. Does it own a queue position, a substation, an ESA, or dual 345 kV feeds?
- Contractual matching. A PPA or renewable certificate is a claim on output, not evidence of physical hourly delivery.
AWS’s $650m purchase from Talen in March 2024 bought a data centre campus, not a power plant. TeraWulf’s Lake Mariner owns the site and a retired coal plant’s dual 345 kV interconnection, not generation. Bitfarms (now Keel Infrastructure) genuinely owns two merchant coal-refuse plants with ~165 MW of nameplate capacity; Soluna bought a 150 MW wind farm; Alphabet bought Intersect’s development platform, not an operating thermal fleet. (Case-study working paper, Sections 2 and 3; Plugin Capital, The Global Neocloud Top 20)
| FIGURE 6 – The five-step spectrum from grid supply to islandable microgrid Original Plugin Capital Research classification, derived from the 38-case working paper. Case names are illustrative. ‘Islandable’ means no physical grid connection or a designed ability to island; ‘BTM’ means co-located generation behind the meter with a grid connection retained. Amazon GW Ranch and Project Jupiter islandability rest on journalistic / company claims respectively, not independent confirmation. |
The spectrum runs from a plain utility tariff, through front-of-meter PPAs and co-located behind-the-meter (BTM) generation, to owned generation with a grid tie, and finally to a genuinely islandable microgrid. Moving right buys control over the energisation date; it also buys fuel risk, emissions permitting, maintenance complexity, and the whole reliability stack that a utility used to provide. Our Neocloud study made the same point from the other side: “contracted power” across providers variously means signed PPAs, approved interconnections or capacity covered by leases, and cannot be used for precise ranking.
IV. The four obstacles — and one economic trap
4.1 The equipment queue decides the technology
The most under-appreciated fact in this market is that fuel choice is downstream of supply-chain choice. GE Vernova’s gas power equipment backlog and slot reservation agreements grew from 62 to 83 GW during 2025, with Q4 orders of $22.2bn (+65% organically) and a total backlog of $150bn. CEO Scott Strazik guided to ~100 GW by end-2026, with 2029 and 2030 capacity essentially sold out. Siemens Energy took 102 gas turbine orders in Q1 FY2026, about a quarter data-centre related, with delivery slots into 2029–2030 and “very limited” 2028 supply; Mitsubishi booked 31 large-frame contracts in the first three quarters and is quoting ~2031 delivery for new large-frame units. (GE Vernova FY2025 results; China Energy News / CCPIT summarising OEM results calls)
RaboResearch’s May 2026 study supplies the alternative. Gas-fired reciprocating engines (Caterpillar, Wärtsilä, INNIO Jenbacher) account for 18.3 GW of announced US capacity, in units of 2.5–35 MW, at $1,700–2,000/kW all-in and 12–24 month procurement. Aeroderivative turbines (25–50 MW) arrive in one to two years. Solar PV and onshore wind take 18–36 months at $1,150–1,600/kW and $1,900–2,300/kW respectively. Heavy-duty turbines and CCGT make up the largest announced category (27.2 GW + 14.2 GW) but are poorly suited to the current cycle. In total, the US BTM pipeline exceeds 130 GW, more than 80% natural gas, with >20 GW of initial-phase co-located capacity already in advanced development. (RaboResearch, The Sprint; Rabobank summary)
Author’s inference, not a source statement: for a project targeting 2027 energisation, only mobile/containerised turbines, aeroderivatives, reciprocating engines or fuel cells are physically available; a heavy-duty 7HA/9HA-class frame ordered in late 2026 is a 2030–31 asset. That, far more than any corporate decarbonisation preference, explains why xAI used trailer-mounted turbines, why VoltaGrid/Oracle specified Jenbacher engines, and why Oracle replaced gas turbines with Bloom fuel cells at Project Jupiter. The trade-off is operational: a 2 GW campus built from 2.5 MW units can require hundreds of engines, with maintenance and long-run efficiency becoming binding constraints.
4.2 Uptime: the reliability tax
Uptime Institute defines Tier I–IV by topology — basic capacity, redundant components, concurrently maintainable, fault tolerant — and does not publish availability percentages on its Tier Certification page. The widely quoted 99.982% (Tier III, ~1.6 h/yr) and 99.995% (Tier IV, ~26.3 min/yr) are conventional industry derivations from that topology, not Uptime SLAs, and should be cited as such. (Uptime Institute, Tier Certification)
The outage data is unambiguous. In the 2025 Global Data Center Survey, power remained the top cause of impactful outages at 45% (n=96), down from 54% (n=97) in 2024; within power-related IT service outages, UPS failures account for 42%, transfer-switch (utility) failures 36% and generator failures 28%. The 2025 Annual Outage Analysis reports 54% of respondents’ most recent significant outage cost more than $100,000, with one in five above $1m. Uptime also warns that “data relating to outages should be treated skeptically”. (Uptime Institute, 2025 Global Data Center Survey, Figure 12; Annual Outage Analysis 2025 executive summary; Data Center Knowledge summary)
| Contradiction we did not smooth over. Uptime’s Chinese newsroom reports 57% above $100,000 and “half” above $1m for the 2025 survey, versus 54% and one-in-five in the primary English report. Different survey waves, samples and question wording. We use the primary report figures and flag the discrepancy rather than blending them. |
| FIGURE 7 – Schematic: a data centre microgrid and the optional grid tie Original concept diagram by Plugin Capital Research. Component list is illustrative and derived from Uptime Institute tier topology, Uptime 2025 outage data and RaboResearch 2026. Not an engineering single-line diagram; no vendor-specific configuration is implied. |
Reasoned inference, not a Uptime claim: Tier III/IV availability is achieved with N+1 or 2N redundancy plus a firm grid. Remove the grid and the operator must internalise black start, spinning reserve, fuel firmness, maintenance redundancy and every forced outage. This is precisely why most projects marketed as off-grid quietly keep a grid tie — and why RaboResearch’s observation that engines “can revert to backup duty” once permanent grid service arrives is the single most elegant risk-mitigation structure in the market. The bridge asset does not become stranded; it becomes the diesel genset the site needed anyway.
4.3 Fuel, permits and social licence
xAI’s Colossus is the most documented case of speed beating process — and of the bill arriving later. Colossus 1 started with an 8 MW grid connection; aerial imagery in April 2025 showed ~35 mobile turbines (~422 MW). Shelby County permitted 15 permanent Solar SMT-130 turbines (247 MW) on 2 July 2025, with SCR controls and backup-only status; the grid connection was upgraded to 150 MW from MLGW/TVA in November 2024/2025, with xAI paying for upgrades and agreeing to curtail during peak demand. At Southaven, Mississippi, a US Senate EPW Committee letter of 15 April 2026 states that 27 gas turbines capable of nearly 500 MW were operating without permits. MDEQ’s Permit Board approved 41 permanent turbines / 1.2 GW in March 2026; Gwynne Shotwell told the White House xAI would “commit to develop 1.2 gigawatts of power as our supercomputer’s primary source”. The temporary fleet grew to 69 units by July 2026; an MDEQ agreed order on 30 July 2026 required retirement to begin in August 2026 and finish by July 2027. NAACP, SELC and Earthjustice litigation, a DOJ intervention and EPA’s January 2026 rule on temporary combustion turbines followed. (US Senate EPW letter to EPA, 15 April 2026; E&E News, xAI gets air permit; case-study working paper, A1)
Two further cautionary examples. Oracle / Bloom “Project Jupiter” (Doña Ana County, NM) withdrew its original gas-turbine and diesel air-permit applications in April 2026 after community pushback on water and emissions, and redesigned the campus around up to 2.45 GW of solid-oxide fuel cells in “one single microgrid” — with a 2.8 GW master agreement and 1.2 GW initial tranche, three different scopes that must not be added. (Oracle press release, 27 April 2026) Fermi America’s Project Matador markets a 17 GW private “HyperGrid” near Amarillo, but its own homepage simultaneously says first power in 2026 and 2027, ~6 GW is permitted, ~2 GW of generation assets secured, construction paused in February 2026 and the CEO/CFO departed in April 2026. (Fermi America; Cleanview tracker) Company announcements at IPO stage are forward-looking statements, not installed capacity.
4.4 The regulatory pendulum
On 1 November 2024, FERC rejected PJM’s amended interconnection service agreement that would have increased co-located load at Talen’s Susquehanna nuclear plant from 300 MW to 480 MW (with a path to 960 MW), by a 2–1 vote. The order (189 FERC ¶ 61,078) held that the applicants had not met the high burden for non-conforming provisions; Chairman Phillips dissented, calling it “a step backward for both electric reliability and national security” and citing PJM’s analysis that no transmission upgrades were needed up to 480 MW. The core objection was cost allocation: large loads taking power behind the meter might avoid transmission charges borne by others. (FERC order, Docket ER24-2172; FERC docket page; Chairman Phillips’ dissent)
On 11 June 2025, Talen and AWS announced a 1,920 MW front-of-the-meter PPA through 2042, with deliveries ramping to 840–1,200 MW by 2029 and full volume no later than 2032. Talen’s own 8-K states plainly: “The existing Susquehanna co-located load arrangement … will transition to a ‘front-of-the-meter’ arrangement after the completion of transmission reconfigurations expected in the Spring of 2026”; Susquehanna will inject into PJM, Talen will act as retail supplier, and PPL will handle delivery. POWER Magazine puts the deal at 17 years / ~$18bn — a journalistic figure not present in the primary release. (Talen Energy 8-K, Exhibit 99.1, 11 June 2025; POWER Magazine)
This is the cleanest documented evidence in our dataset that “off-grid” is a regulatory outcome, not just an engineering choice. The same pendulum swung back partially: FERC approved a waiver on 1 June 2026 allowing Constellation to transfer 760 MW of capacity interconnection rights from Eddystone to the Crane restart, even as PJM warned that some transmission upgrades for Crane might not complete until 2031. (World Nuclear News; Nuclear News Network restart tracker, updated 31 October 2026)
4.5 The economic trap: a bridge with no destination
RaboResearch’s “sustained cost of delivered energy” (SCDE) is the most useful framework we found, because it prices the asset over the life it may actually lead rather than the three-to-seven-year bridge it was sold as. At year five, gas reciprocating engines cost $107/MWh, aeroderivatives $106/MWh and combined cycle $101/MWh, against a US industrial grid rate of about $86/MWh. But at high capacity factors, engines and aeroderivatives hit recurring overhaul and rebuild cycles around years 10, 20 and 30 — a sawtooth — while CCGT efficiency improves its position to roughly $55/MWh by year 15 and ~$50/MWh by year 30. The economic crossover with gas engines occurs around year 10. RaboResearch separately estimates BTM gas at $100–165/MWh versus $90–95/MWh for grid alternatives. (Gas Turbine World, summarising RaboResearch SCDE; RaboResearch summary)
| FIGURE 8 – Sustained cost of delivered energy: published anchor values ($/MWh) Source: RaboResearch, The Sprint: Data centers are building a parallel energy system in the US (May 2026), SCDE methodology, as reported in Gas Turbine World. Published anchors only: year 5 — gas reciprocating engines $107, aeroderivatives $106, combined cycle $101; year 15 — CCGT ~$55, heavy-duty simple cycle ~$58; year 30 — CCGT ~$50, simple cycle ~$55; US average industrial grid ~$86. A dash means the value was not published in the retrieved source; no intermediate values have been interpolated. RaboResearch reports a ‘sawtooth’ cost pattern for engines and aeroderivatives around years 10/20/30 caused by overhaul and rebuild cycles, and separately estimates BTM gas at $100–165/MWh versus $90–95/MWh for grid alternatives. |
The premium is usually rational: power is only 10–15% of a data centre’s total cost of ownership, but a delayed facility leaves $3–5bn of annual GPU depreciation and lost revenue idle. The premium stops being rational if the grid never arrives and the “temporary” asset becomes a decade-long primary power source with a cost curve that worsens with age. RaboResearch names three unresolved uncertainties: equipment longevity, the regulatory viability of sustained BTM operation, and the bankability of generation assets that outgrow their original purpose. Its strategist is blunter about who should play: unless you are at ≥500 MW–1 GW, can commit to a 20–25 year contract and have a balance sheet that can absorb a stranded asset, self-supply is dangerous — the 1997–2002 merchant gas turbine cycle ended in a wave of write-downs. (C-Suite Newsletter interview with RaboResearch)
V. What powers the world — and what actually powers data centres
| FIGURE 3 – Global electricity generation mix, 2025, and change versus 2024 (TWh) Source: Ember, Global Electricity Review 2026 (data year 2025; 31,779 TWh total; 91 countries representing 93% of global demand plus 215-country dataset). Wind (2,701 TWh), other fossil (846 TWh) and other low-carbon (811 TWh) are derived from the published shares (8.5% / 2.7% / 2.6%) and reconciled to the published total; small rounding differences may remain. Change data: solar +636, wind +205, gas +36, nuclear +35, other low-carbon +9, hydro +3, other fossil −12, coal −63 TWh. |
Ember’s Global Electricity Review 2026 (data year 2025) reports 31,779 TWh of global generation. Renewables reached 33.8% (10,730 TWh), overtaking coal at 33.0% (10,476 TWh) for the first time in the modern era; gas was 21.8% (6,919 TWh), nuclear 8.9% (2,812 TWh), hydro ~14.0% (4,436 TWh), wind 8.5% and solar 8.7% (2,778 TWh). Low-carbon sources together reached 42.6%; fossil still 57.4%. The 2025 change data is the real story: solar +636 TWh (18× the increase in gas), wind +205, nuclear +35, gas +36, coal −63. Solar met 75% of all demand growth. Batteries fell ~45% in cost and deployment rose 46% to ~250 GWh — enough to shift only 14% of new solar generation from midday to other hours. (Ember, Global Electricity Review 2026; full PDF; Argus Media summary with absolute TWh)
| FIGURE 5- World grid mix (2025) versus modelled physical supply to data centres (2024) Sources: Ember, Global Electricity Review 2026 for the world grid (fossil 57.4%, nuclear 8.9%, renewables 33.8%); IEA, Energy and AI (2025) for data centres (coal ~30%, gas ~26%, nuclear ~15%, renewables ~27%; ~2% other/rounding). These are NOT like-for-like: different years, different sources, global generation versus modelled physical supply to data centres including on-site generation. Corporate PPAs and renewable certificates are excluded from the physical measure. |
Data centres do not have a single fuel mix, and corporate procurement claims are not physical delivery. The IEA’s modelled physical supply to data centres in 2024 was roughly coal 30%, renewables 27%, gas 26%, nuclear 15% — with the US more gas-heavy (gas >40%, renewables 24%, nuclear ~20%, coal ~15%). By 2035, the IEA still expects fossil fuels to supply around 40% of data centre electricity even as renewables and nuclear expand. (ORF Global Quarterly, Table 2, citing IEA Energy and AI; CloudsPress, explaining the IEA physical-supply measure)
| FIGURE 4 – US fleet capacity factors, 2025, and nameplate MW needed for 100 MW continuous Sources: Left — US Energy Information Administration, Electric Power Monthly, Tables 6.07.A and 6.07.B, calendar year 2025 (nuclear 91.0%, gas combined cycle 58.4%, coal 48.7%, wind 34.2%, utility-scale solar PV 24.4%). Right: author’s illustrative calculation, nameplate = 100 MW ÷ capacity factor. This is annual-energy parity ONLY and deliberately understates the requirement: it ignores hourly profile, a 14-hour night (~1,400 MWh usable storage for 100 MW), N+1/2N redundancy, seasonal derating, curtailment and land. US fleet figures, not global averages. |
Here is the crux. A data centre does not buy annual terawatt-hours; it buys every hour of firm capacity. Using EIA 2025 US fleet capacity factors, the nameplate capacity required to match the annual energy of a 100 MW continuous load is ~110 MW of nuclear, ~171 MW of gas CCGT, ~292 MW of wind or ~410 MW of solar PV — and that is before hourly firming. A 14-hour night alone needs ~1,400 MWh of usable storage at 100 MW, before N+1 redundancy, seasonal derating and land. (EIA Electric Power Monthly, Tables 6.07.A / 6.07.B)
The resulting ranking for a 2027 energisation date is therefore not a ranking of cleanliness or even of LCOE:
| Option | Firm 24/7? | Realistic 2026–29 availability | Main constraint |
| Gas reciprocating engines / aeroderivatives | Yes, with N+1 | 12–24 months | Fuel supply, air permits, maintenance at scale, carbon liability |
| Heavy-duty CCGT | Yes | 2029–2031 delivery slots | OEM backlog; ~$2.3–2.5bn/GW |
| Utility-scale solar + wind + BESS | Only with firming | 18–36 months | 24.4% / 34.2% capacity factors; multi-day storage; land |
| Existing nuclear (PPA / restart / uprate) | Yes | 2027–2032 | Scarce assets, regulatory approval, cost allocation |
| New nuclear / SMR | Yes | Early 2030s; <2 GW US before 2033 | First-of-a-kind, licensing, no near-term capacity |
| Geothermal / fuel cells | Yes / claimed | 2027–2030 | Site-specific resource; scale; cost |
Sources: RaboResearch, The Sprint, EIA, GE Vernova, Ember.
The honest conclusion is uncomfortable for a green-tech real assets manager: gas wins the 2026–2029 window on speed, nuclear wins the 2030s on firmness, and renewables win on energy cost but require firming that the current battery build-out cannot yet supply at hyperscale. That is a sequencing problem, not a technology verdict — and it is exactly why a credible conversion path (fuel cells, CCS, geothermal, 24/7 carbon-free matching, or a grid tie with a defined retirement date for bridge assets) belongs in the underwriting, not in the marketing appendix.
VI. The evidence: 38 cases, sorted by what they actually prove
| FIGURE 9 – Highest verified status of 38 off-grid / power-asset cases Source: Author’s exclusive reclassification of the 38 cases in the accompanying working paper (Plugin Capital Research, 6 November 2026): each case assigned a single highest verified status. Status tags in the underlying research overlap and must not be summed to 38. ‘Paper status’ is not the same as ‘no physical progress’: RaboResearch counts >20 GW of initial-phase co-located US capacity in advanced development with equipment procured, fuel secured and permits approved. |
We verified 38 distinct cases of data centres building, buying or contracting dedicated power, then assigned each a single highest-verified status. The result is sobering: 7 operational, 6 under construction, 19 contracted/announced, 3 permitted, 2 proposed, 1 blocked. One third of the sample has steel in the ground; two thirds is a press release, a permit or a term sheet. This does not mean no physical progress — RaboResearch counts >20 GW of initial-phase co-located capacity in advanced development with equipment procured and fuel secured — but it does mean that announced GW and energised GW are different assets.
Table 1 · Behind-the-meter and off-grid generation (selected)
| Case | Date | Status | Scale | Technology | Truly islandable? |
| xAI Colossus 1 / 2 / Southaven | Jun 2024 – Aug 2026 | Operational + permitted | 150 MW grid; 41 permanent turbines / 1.2 GW | Mobile + permanent gas turbines, Megapacks | No — BTM + grid backup |
| Stargate Abilene, TX (Crusoe / OpenAI / Oracle) | Sep 2025 – 2026 | Operational (phased) | 360.5 MW permitted onsite; ~1.2 GW campus | GE LM2500XPRESS, Solar Titan 350, BESS | No — ERCOT-connected |
| Crusoe / Tallgrass → Google “Project Tembo”, WY | Jul 2025 / Nov 2026 | Permitted, superseded | 1.8 GW → 2.7 GW resource mix | Gas + 702 MW fuel cells with CCS | Original design yes; Google version hybrid (590 MW grid service) |
| Williams “Socrates”, OH | Jun 2025 – 2026 | Under construction | 400 MW (2 × 200 MW) | Gas BTM | Contested — Williams says “behind-the-meter”, not off-grid |
| VoltaGrid / Oracle, TX | Oct 2025 | Contracted | 2,300 MW | INNIO Jenbacher engines, ABB, Energy Transfer fuel | Unverified |
| Bloom / Oracle “Project Jupiter”, NM | Apr 2026 | Contracted; prior gas design cancelled | 2.45 GW project / 2.8 GW master | Solid-oxide fuel cells, one microgrid | Claimed, not independently confirmed |
| Amazon “GW Ranch”, Pecos County TX | Aug 2026 | Permitted / proposed | 7.65 GW, 35 turbines | Dedicated gas BTM | Reported islandable; no Amazon primary found |
| TeraWulf “Nautilus”, PA | Feb 2023 | Operational | 50 → 100 MW | Direct BTM from 2.5 GW Susquehanna nuclear at ~2¢/kWh | Effectively islanded BTM — mining load, not AI |
| Bitfarms / Keel, PA | Mar 2025 | Operational plants | 165 MW nameplate owned | Two coal-refuse merchant plants, PJM | Partial self-supply; PJM-connected |
| Soluna, TX | Apr 2026 | Operational | 150 MW Briscoe wind farm owned | Wind + co-located modular DC | No — ERCOT Batch Zero |
| Fermi “Project Matador”, Amarillo TX | 2025–2026 | Proposed / partial permit; stalled | 17 GW marketed; ~6 GW permitted | Gas, 4 × AP1000, solar/BESS | Hybrid — 200 MW grid link planned |
| Homer City Energy Campus, PA | Apr 2025 | Under construction | up to 4.5 GW | 7 × GE Vernova 7HA.02 CCGT | No — PJM and NYISO connected |
| Meta Hyperion / Entergy, LA | Aug 2025 | Under construction | 2,262 MW (3 × 754 MW) | Utility-owned CCGT + 500 kV | No — contrast case: utility owns the plants |
Sources: Senate EPW letter on xAI, Cleanview Stargate Abilene, Crusoe/Tallgrass release, Williams, VoltaGrid, Oracle/Bloom, TeraWulf Nautilus, Bitfarms/Stronghold, Soluna, Fermi, Homer City, S&P Global on Meta/Entergy. Full 38-case detail with URLs and caveats is in the accompanying working paper.
Table 2 · Acquisitions, restarts and long-term firm-power contracts (selected)
| Case | Date | Status | Scale | What was actually acquired |
| Microsoft / Constellation, Crane (TMI-1) | Sep 2024 | Restart under construction | 835 MW, 20-yr PPA, $1.6bn | Output of a restarted reactor — not a plant, not off-grid |
| AWS buys Talen’s Cumulus campus | Mar 2024 | Acquisition closed | $650m; up to 960 MW planned | A data centre campus, not a power plant |
| FERC ER24-2172 rejection | Nov 2024 | Blocked | 300 → 480 MW co-located load | Regulatory determination on BTM cost allocation |
| Talen / AWS restructured PPA | Jun 2025 | Contracted | 1,920 MW to 2042 | Front-of-the-meter retail PPA |
| Alphabet acquires Intersect | Dec 2025 (closed Mar 2026) | Closed | $4.75bn cash + debt; multiple GW | A development platform + co-located Haskell County, TX site |
| Vistra / Meta nuclear PPAs | Jan 2026 | Contracted | >2,600 MW incl. uprates | Output of existing PJM nuclear plants |
| Constellation / Meta, Clinton IL | Jun 2025 | Contracted from 2027 | 1,121 MW + 30 MW uprates | 20-yr output contract |
| Constellation / Amazon, Calvert Cliffs | Sep 2026 | Contracted | 690 MW incl. 190 MW uprate | Output + uprate |
| Google / Kairos Power SMR | Oct 2024 | Contracted | 500 MW by 2035 | Future SMR fleet PPA |
| Amazon / X-energy | Oct 2024 | Contracted | ~$500m Series C-1; Cascade 320→960 MW | Equity + future SMR capacity |
| Oklo / Switch | Dec 2024 | Non-binding MPA | 12 GW through 2044 | Non-binding master power agreement |
| Google / Fervo “Corsac” geothermal | May 2025 | Permitted | 115 MW | Enhanced geothermal + NV Energy Clean Transition Tariff |
| MARA / Long Ridge Energy & Power | Apr 2026 | Contracted | >1 GW potential | Ownership of an existing power plant site |
| SoftBank / SB Energy, Piketon OH | Mar 2026 | Proposed | 10 GW generation, 10 GW DC | Grid-connected, DOE-announced |
| Brookfield / NextEra, Paducah KY | Jul 2026 | Proposed | 4.6 GW dedicated generation | $100bn “bring your own power” campus, no anchor tenant |
Sources: Constellation 8-K, FERC order, Talen 8-K, Alphabet/Intersect release, RaboResearch (Intersect closing), DOE Portsmouth fact sheet, Brookfield Paducah PDF. Additional case URLs in the working paper.
Three cases that carry the argument
Microsoft / Constellation, Crane Clean Energy Center. A 20-year PPA for the full 835 MW output of the restarted Three Mile Island Unit 1, backed by ~$1.6bn of investment and a $1bn DOE loan closed in November 2025. NRC issued its final environmental assessment and FONSI on 18 September 2026; FERC approved the Eddystone→Crane capacity-rights transfer on 1 June 2026; Constellation targets 2027, though PJM has warned that some transmission upgrades could run to 2031. Crucially, Microsoft’s data centres will not be directly connected to Crane — the plant injects into PJM. This is firm carbon-free power, not off-grid power. (SEC 8-K; World Nuclear News; CRES Forum site visit note; ABC/AP on the DOE loan)
Talen / AWS. The complete arc — a $650m campus purchase (Mar 2024), a rejected BTM ISA (Nov 2024), and a 1,920 MW FTM retail PPA (Jun 2025) — is the best evidence that regulatory design, not engineering, determines whether “off-grid” survives. AWS now pays all applicable transmission and distribution charges, which was the central concern in FERC’s rejection.
Alphabet / Intersect. A $4.75bn cash-plus-debt acquisition (22 Dec 2025, closed Mar 2026) of a platform with multiple gigawatts of energy and data centre projects in development or under construction, including the first co-located data centre and power site in Haskell County, Texas. Alphabet explicitly frames it as building “new power generation in lockstep with new data center load”. This is the clearest example of a hyperscaler moving from buying electrons to owning the energisation pipeline. (Alphabet press release; RaboResearch)
One widely repeated claim that needs correcting
Headlines in 2026 routinely say hyperscalers are “buying power plants”. We could not verify a single case of a hyperscaler acquiring a large operating thermal plant outright. What we did verify are: long-term PPAs and output contracts (Microsoft/Constellation, Talen/AWS, Meta/Vistra, Meta/Constellation, Amazon/Constellation), restarts and uprates, purchases of data centre campuses co-located with generation, acquisitions of development platforms (Intersect), of merchant coal-refuse plants by a bitcoin miner (Bitfarms/Keel), of a wind farm (Soluna), and of power-plant sites (MARA/Long Ridge). Ownership of generation, ownership of interconnection rights and ownership of output are three different assets with three different risk profiles. Conflating them leads to systematically overvalued “power bankability” claims.
VII. Implications for investors — the APAC lens
1. Underwrite time-to-power, not $/MWh. A $10–70/MWh BTM premium is small against $3–5bn of annual GPU depreciation on an idle cluster. The underwriting question is not “what does power cost” but “what is the probability-weighted energisation date, and what is the asset worth if it slips four quarters”. In our own earlier work we concluded that the core value of modular construction is compressing time-to-energisation; the same logic applies one layer down the stack.
2. Own the interface, not necessarily the turbine. The scarce, durable assets are interconnection positions, powered land with a signed ESA, substations, transformer and turbine slot reservations, air permits, gas supply contracts and the financing wrapper. TeraWulf’s dual 345 kV interconnection, ERCOT Batch Zero queue positions, Malaysia’s DCTF approvals and the Green Lane Pathway are all examples of queue positions being monetised as assets. For an Asia Pacific manager structuring JVs, club deals and funds, this is the most natural ownership layer.
3. Hybrid is the bankable default; islandable is the exception. Underwrite the grid tie as an option, not as a design failure. The most robust structure we found is a bridge asset with a defined second life: engines that revert to backup duty when permanent grid service arrives, or BESS that transitions from prime-power firming to peak shaving and demand-response revenue. Regulators increasingly require exactly this — Ireland’s matching-generation rule, NSW’s proposed 25% demand reduction for up to two hours and 4-hour storage equal to 25% of contracted generation capacity.
4. Contract length is the hidden mismatch. Fifteen-to-twenty-year take-or-pay leases are being written against three-to-seven-year bridge assets and four-to-six-year GPU depreciation lives. A credible underwriting model needs an explicit replacement reserve, an overhaul schedule and a refinancing case for year 10 — the point at which RaboResearch sees the economic crossover from engines to combined cycle.
5. Regulatory and social licence are first-order variables. FERC cost allocation, EPA temporary-turbine rules, state moratoria, CRU connection policy, water tariffs, noise litigation and community opposition have each changed or delayed a project in our sample. These belong in the base case, not the risk appendix.
6. Price the carbon honestly. Plugin Capital’s dual mandate — financial resilience plus a low-carbon transition — sits in direct tension with a gas-heavy bridge. We do not resolve that tension by relabelling gas as “transition”. We resolve it by requiring: a measurable 24/7 carbon-free matching plan, a defined retirement or conversion date for bridge assets, a funded CCS/fuel-cell/geothermal pathway where feasible, and an explicit carbon price in the model. The IEA’s own projection that fossil fuels still supply ~40% of data centre electricity in 2035 is a description of the baseline, not a licence to reproduce it.
7. Size and structure matter more for mid-market APAC projects. Below ~100 MW, self-supply economics generally do not work; the RaboResearch threshold is 500 MW–1 GW with a 20–25 year commitment and a balance sheet able to absorb a stranded asset. For 20–100 MW projects in Malaysia, Indonesia, Thailand and Vietnam, the realistic route is a shared or clustered microgrid across multiple offtakers — a club deal or fund structure that aggregates load to reach scale, shares redundancy, and spreads equipment and permitting cost. That is precisely the structuring capability a real assets manager can add and a single-tenant developer cannot.
VIII. The risks we are watching
- Demand could flatten. The IEA’s own sensitivity cases (Lift-Off, High Efficiency, Headwinds) show wide uncertainty; cheaper inference, model efficiency and slower AI adoption would reduce the scarcity premium that makes a $165/MWh bridge rational. The 2026 IEA update still projects AI-focused data centre electricity roughly tripling between 2025 and 2030, but that is a projection, not an observed outcome.
- Grid reform could actually work. 1,200–1,600 GW of advanced-stage projects may be unlocked through non-firm connections and grid-enhancing technologies; FERC Order 2023 reforms are too new to evaluate. A faster queue removes the scarcity premium from bridge assets and strands those bought at 2026 prices.
- The regulatory pendulum. Co-location rules, cost-allocation decisions, emissions rules and local moratoria can convert a BTM structure into an FTM one — or kill it — after capital is committed.
- OEM capacity expansion. GE Vernova is targeting ~24 GW/yr by 2028 from ~20 GW; Chinese transformer capacity is expanding; Doosan Enerbility has entered the US data-centre turbine market. A supply response compresses lead times and equipment premia.
- Stranded-asset history repeating. The 1997–2002 merchant gas turbine cycle ended in write-downs. A bridge that becomes permanent has a different cost curve from the one in the model.
- Tenant credit concentration. Many 2026 projects are anchored to a single tenant. Our August note flagged this as the first risk in modular DC; it applies with equal force to a project-financed power plant.
- Data quality. “Designed to scale to 10 GW” is not installed capacity; a queue volume is not a pipeline; a PPA is not physical hourly delivery; “100% renewable” is not 24/7 carbon-free. We have kept these distinctions throughout and flagged every figure we could not confirm.
IX. Conclusion
The microgrid is not an alternative to the grid. It is an option on time. It is bought because a utility connection arrives in 2031 and a GPU cluster depreciates from the day it is delivered, and it is priced — correctly — at a premium to grid power because the alternative is idle silicon.
But the option only pays if the holder is honest about three things: that most “off-grid” projects still keep a wire to the utility; that reliability, fuel, permitting and redundancy are costs that do not disappear when the utility does; and that a bridge asset needs a destination. The projects in our sample that look most robust are not the largest or the greenest. They are the ones with a defined second life — engines that become backup gensets, batteries that become grid assets, interconnection rights that survive the technology they were bought for.
For Asia Pacific investors, the investable surface is not the combustion turbine. It is powered land, queue position, long-lead equipment, flexible bridge capacity, and the capital structure that binds them to an investment-grade offtaker. The electrons are not the product. The energisation date is.
Sources, calibres and limitations
Primary sources (directly relied upon).IEA, Energy and AI (2025) — Executive Summary · IEA, Energy demand from AI · IEA, Electricity 2026 — Grids · UNDP TIDE Centre, The Energy Sprint of the AI Race (regional table reproducing IEA) · LBNL, Queued Up: 2026 Edition · Ember, Global Electricity Review 2026 (PDF) · EIA Electric Power Monthly, Tables 6.07.A / 6.07.B · Uptime Institute, 2025 Global Data Center Survey · Uptime Institute, Annual Outage Analysis 2025 · Uptime Institute, Tier Certification · RaboResearch, The Sprint (May 2026) · GE Vernova FY2025 results · FERC order, ER24-2172 · Talen Energy 8-K, 11 June 2025 · Constellation 8-K, 20 September 2024 · Oracle / Bloom press release · Homer City Redevelopment · US Senate EPW letter on xAI · TNB newsclip, 1 October 2026 · DOE Portsmouth fact sheet · Brookfield Paducah PDF · Johns Hopkins, State Playbook.
Secondary sources (explicitly not primary).The Star (Malaysia) · SDEA on Ireland / Singapore / Germany · New Polis, citing CBRE and JLL · Data Center Dynamics on Amsterdam · POWER Magazine on Talen/AWS · Gas Turbine World on SCDE · Argus Media on Ember · ORF Global Quarterly (IEA data-centre fuel mix) · GridReadiness transformer tracker · Nexus Power transformer guide · AEInsights (IEA 2026 vintage) · Sustainability Online (IEA regional) · The Electricity Hub (20% delay) · C-Suite Newsletter (RaboResearch interview) · E&E News on xAI · World Nuclear News · Alphabet/Intersect release (reproduction) · Cleanview case trackers · Southeast Asia Green Economy Report 2026.
Four explicit limitations of this post.
- Figure vintages are not blended. The regional demand series is the IEA’s 2025 vintage; the 485 TWh (2025) / ~950 TWh (2030) figures are the 2026 update. They are quoted separately and never averaged.
- Capacity factors are US fleet averages, not global or site-specific. The nameplate calculation in Figure 4 is an illustrative annual-energy-parity calculation by the author, deliberately conservative; it is not an engineering design.
- Case-study GW figures are not interchangeable. Campus targets, permitted onsite generation, turbines on order, IT load and contracted pipeline capacity are different measures. We flag conflicts in-text rather than reconciling them silently. x.ai primary pages were Cloudflare-blocked; Amazon GW Ranch, Google Project Tembo and Williams Socrates “fully off-grid” claims rest on secondary or company sources only. FERC EL25-49 and a hyperscaler purchase of an operating thermal plant were not verified and are excluded from load-bearing arguments.
- Status counts in Figure 9 are our own exclusive reclassification of 38 cases, not a source statistic. The underlying status tags overlap and must not be summed. Announced capacity is not installed capacity.
Disclaimer — This article is published by Plugin Capital Research for general information and does not constitute investment advice. All figures, valuations and projections reflect public-source reporting as of the publication date; forward-looking statements involve uncertainty. Private materials were not used. Figures in company announcements (especially GW-scale “designed to scale” claims, IPO-stage disclosures and pre-tenant projects) are not installed capacity. Institutional readers should re-verify the FERC co-location docket, the Google/Project Tembo primary announcement, the Amazon GW Ranch primary and the Williams Socrates OPSB docket before relying on them. Past performance is not indicative of future results.