On February 25, 2026, Fermi America received a final air permit from the Texas Commission on Environmental Quality (TCEQ) for 6 gigawatts (GW) of gas generation at its Project Matador campus in Amarillo. The application had been filed six months and twenty-five days earlier. Fermi's state selection, attainment-area Texas with existing pipeline infrastructure adjacent to the campus and no mandatory siting board review, produced the fastest documented air permit for a large behind-the-meter (BTM) gas plant in the United States.¹

Forty-five days later, on March 20, 2026, New Mexico State Land Commissioner Stephanie Garcia Richard denied Energy Transfer's application for a right-of-way across 0.63 miles of state trust land in Doña Ana County. That 0.63-mile segment was the critical link in the 17-mile "Green Chile" gas lateral connecting El Paso's pipeline network to Project Jupiter, the OpenAI and Oracle AI data center campus that is part of the $500 billion Stargate program. Commissioner Garcia Richard's decision, on grounds that the pipeline did not serve the best interest of the state's school trust, indefinitely blocked gas supply for up to 2.45 GW of planned data center power capacity.²

No federal agency has authority to reverse her decision. The Bureau of Land Management compressed its own federal environmental review for the Green Chile Lateral from one year to 14 days using emergency permitting authority. It approved the federal land segment in May 2026.³ The New Mexico State Land Office, answerable to the state's voters rather than to Washington, denied reconsideration on July 14, 2026.⁴

This is not a permitting delay. It is the structural condition of BTM gas development in the United States in 2026.

The Escape Route That Has Its Own Gate

Three years ago, data center developers discovered a workaround to the interconnection queue. A private gas plant, private wires, and no connection to the interstate transmission grid meant no ISO/RTO involvement, no interconnection study, and no five-to-eight-year queue position at PJM or MISO. Behind the meter: faster, cheaper, fully under operator control.⁵ As we documented in our analysis of why behind-the-meter won the grid race, the BTM decision was driven by PJM capacity costs that rose 11-fold in two years and interconnection queues stretching to 2030 and beyond.

The BTM build-out that followed is enormous and largely unrealized: 2 GW online today against 101 GW announced, 7 GW under active construction, 57 GW with confirmed equipment orders. The 13 GW base-case target for end of 2027 requires executing the full permitting stack on projects currently in queue. The low scenario is 5 GW, if permitting delays prove systemic.⁶

The gap between 5 GW and 13 GW represents approximately $65 billion in capital deployment difference for a single year. That gap is almost entirely a permitting story.

On June 18, 2026, FERC issued Section 206 show-cause orders to all six RTOs and ISOs, directing them to justify or reform large-load interconnection tariffs by August 17, 2026. The orders address five categories: co-location arrangements, interconnection study processes, cost allocation for network upgrades, retail cost protection (delegated to state commissions), and market participation eligibility.⁷ Our deep dive into the June 18 show-cause orders covers what the RTO responses will and will not address. The orders are significant. For BTM developers, they are irrelevant.

FERC confirmed in its own analysis that BTM arrangements are "unequivocally under state jurisdiction." The agency has no authority over EPA air permits, state land office right-of-way decisions, state siting board approvals, or intrastate gas pipeline construction.⁷ FERC reform fixes the interconnection queue. The BTM permitting queue is five layers deep, and none of those layers answer to FERC.

The Five-Layer Permitting Stack

A 200 MW BTM gas plant running at 70 to 90 percent annual capacity factor is, from a regulatory standpoint, a utility-scale combustion facility. It produces the same NOx, CO, particulate matter, and greenhouse gas emissions profile as a traditional power plant, without the interconnection queue, transmission upgrade, or utility rate case that would normally constrain one. The Clean Air Act applies in full. Every shortcut that BTM developers assumed was available has been tested and, in most cases, closed.

The full permitting stack:

Layer 1: EPA Air Permit (New Source Review/Prevention of Significant Deterioration)

A pre-construction permit required before building any major stationary source. A 200 MW plant running at baseload operation exceeds the 100 tons per year (tpy) NOx major source threshold in virtually every realistic scenario: in nonattainment areas the threshold is 100 tpy; in attainment areas it rises to 250 tpy, but baseload gas turbines exceed even that ceiling. Full Prevention of Significant Deterioration (PSD) review with Best Available Control Technology (BACT) analysis is required. Timeline: 7 months in Texas at best, 18 or more months in Virginia or New Mexico in contested cases.⁸

Layer 2: Title V Operating Permit

A federal operating permit for major sources, applied for concurrently with the NSR/PSD permit. Includes a mandatory 45-day EPA review period after the state issues a draft. Once operational, it imposes continuous emission monitoring, annual compliance certifications, and public records requirements. No major source exemption is available for high-utilization BTM prime power generation.⁸

Layer 3: State Siting Board Approval

State-level review of energy facility location, design, and community impacts. The variation here is extreme. Texas has no mandatory siting requirement for BTM plants not connected to ERCOT, so this layer simply does not exist. Ohio's House Bill 15 (HB15) created a 60-day automatic approval mechanism for BTM facilities on data center developer land; since June 2025, the Ohio Power Siting Board (OPSB) has approved more than 2,000 MW of BTM natural gas generation in the New Albany area alone.⁹, ¹⁰ Virginia's State Corporation Commission reviews facilities of 150 MW or more, with no fixed timeline. California's Energy Commission has exclusive licensing authority for thermal plants of 50 MW or more, a multi-year process.¹¹

Layer 4: Pipeline Right-of-Way

Permission to construct a gas lateral from existing transmission infrastructure to the BTM facility. The pipeline may cross federal land, state trust land, or private land requiring eminent domain. Timeline: 6 to 18 months under standard review; BLM's emergency authority compresses federal segment review to 14 days. The critical risk: state land offices hold independent veto authority over state trust land segments with no federal override mechanism. Western states, including New Mexico, Arizona, Nevada, Colorado, Utah, and Wyoming, control substantial state trust land in precisely the growth corridors where data center development is concentrated.

Layer 5: Local Permits

Zoning, building permits, noise assessments, stormwater management. Typically 30 to 180 days, the fastest layer. Ohio's HB15 removed the local public hearing requirement for expedited BTM facilities. Community opposition can still surface here even after all upper layers are approved.

The Congressional Research Service documented this five-layer structure in December 2025, confirming that none of the layers answer to FERC and that the NSR major source thresholds of 100/250 tpy are statutory, meaning EPA cannot modify them by rulemaking, and no administrative action can change them without an act of Congress.⁸

The Benchmark Cases: Fermi and Project Jupiter

Two case studies define the outer boundaries of what the BTM permitting stack can produce.

Fermi America's Project Matador represents the best achievable outcome under current conditions. The company sited 5,800 acres in the Texas Panhandle, where Amarillo is in attainment for all major National Ambient Air Quality Standards (NAAQS) pollutants. That applies the more lenient 250 tpy major source threshold and the PSD review standard rather than the stricter nonattainment NSR standard. No separate state siting board review exists for BTM plants not connected to ERCOT. Energy Transfer's existing pipeline infrastructure runs adjacent to the campus, eliminating the pipeline ROW layer entirely. No third-party challenges forced a contested case hearing. Result: 6.8 months from filing to final TCEQ permit for 6 GW of gas generation.¹, ¹²

Fermi filed a second 5 GW permit application with TCEQ on March 27, 2026, the same day Energy Transfer was preparing its reconsideration request for the Green Chile Lateral following the initial March 20 denial.¹³

Project Jupiter defines the failure mode. The OpenAI/Oracle campus in Doña Ana County, New Mexico sits in an attainment area, so the air permit threshold and standard are, in principle, no worse than Texas. The constraint was the pipeline route. The 17-mile Green Chile Lateral from El Paso's pipeline network crossed three jurisdictions: federal land (BLM), private land, and 0.63 miles of state trust land controlled by Commissioner Garcia Richard.

Commissioner Garcia Richard's March 20 decision cited minimal trust revenue ($31,900 one-time plus $43,400 per year in business lease payments), greenhouse gas emissions from the sole-purpose pipeline, and campus water consumption in an arid state. BLM approved the federal segment on emergency authority in May. Energy Transfer sought reconsideration. On July 14, 2026, Commissioner Garcia Richard denied it, writing that the project "undoubtedly" benefits investors, developers, and tenants but does not provide significant benefits for state lands, and that even the proposed fuel cell technology pivot "would still produce high amounts of air pollution."⁴

Energy Transfer has approximately until August 13, 2026 to file an administrative review petition. No alternative pipeline route has been announced.

The contrast with FERC's June 18 show-cause orders is precise. FERC directed all six RTOs and ISOs to justify or reform large-load interconnection tariffs by August 17, 2026, addressing co-location arrangements, interconnection study processes, cost allocation, and market participation.⁷ In the same proceeding, FERC confirmed that BTM arrangements are "unequivocally under state jurisdiction." Air permits, state land ROW decisions, state siting board approvals, and intrastate gas pipeline construction sit entirely outside FERC's statutory reach.⁷ Two regulatory processes will resolve their August deadlines simultaneously. One can fix the queue that BTM was designed to bypass. The other cannot reach the queue that replaced it.

The January 2026 Rule That Closed the Last Loophole

Before the permitting stack problem became fully apparent, some developers believed there was a regulatory escape hatch within BTM itself. The theory: classify mobile, trailer-mounted gas turbines as "temporary" equipment or "nonroad engines" — neither category triggers the NSR/PSD major source permitting framework. xAI tested this at scale at its Southaven, Mississippi facility, deploying 27 trailer-mounted gas turbines (growing toward 59 across Colossus 1 and 2) as primary, continuous power for AI training. Potential NOx emissions: 1,700 to 2,000 tons per year, which would make xAI the single largest industrial NOx source in the 11-county Memphis metropolitan area. Memphis is in nonattainment for both particulate matter and ozone.¹⁴

The EPA closed this pathway on January 15, 2026. New Source Performance Standards for Stationary Combustion Turbines, 40 CFR Part 60, Subpart KKKKa, apply to any turbine commenced after December 13, 2024. For large turbines (base load heat input exceeding 850 MMBtu per hour) operating at high utilization (12-month capacity factor above 45%), the Best System of Emission Reduction (BSER) is selective catalytic reduction (SCR) achieving 5 ppm NOx at 15% oxygen. A BTM prime-power plant cannot be classified as low-utilization or as a true backup generator. Mobile turbines used as stationary power sources are regulated as stationary sources regardless of trailer-mounting. The temporary classification cannot be refreshed by swapping units.¹⁵, ¹⁶

In April 2026, the NAACP filed a citizen suit against xAI for Clean Air Act violations. In June 2026, the DOJ moved to dismiss the suit, citing xAI's Grok model as critical to U.S. military operations, without disputing that the turbines lack required permits.¹⁷ If the argument succeeds, it creates a potential enforcement carve-out for existing unpermitted AI infrastructure but establishes no exemption for future projects. The loophole is closed prospectively. Enforcement remains a live wildcard for existing installations.

The full analysis continues below for paid subscribers, including detailed economic analysis, state-by-state strategic playbook, regulatory reform map, and investment landscape.

The Economics: What Delay Actually Costs

Read more

The average time from interconnection request to commercial operation for a US power project hit 55 months in 2024, up from 22 months in 2008.² For large data center loads in constrained markets, the number is worse: projects entering service in PJM during 2025 averaged more than seven years from application to operation.⁸ A new data center takes 12 to 24 months to build. The grid connection takes three times as long.

That asymmetry has a price. At $2.6 million per MW per year in colocation revenue -- the 2025 global average -- every year of interconnection delay on a 31 MW facility costs roughly $80 million in revenue that never materializes.¹⁶ The data center isn't built yet. The racks aren't filled. The contracts aren't signed.

In October 2025, Calibrant Energy and Aligned Data Centers announced they had found a way around that math. They deployed a 31 MW / 62 megawatt-hour battery energy storage system (BESS) at Aligned's new campus in Hillsboro, Oregon -- not as backup power, not as grid arbitrage, but as a deliberate instrument to accelerate interconnection with Portland General Electric (PGE). Calibrant's press release called it "the first time in the US that a battery system is purpose-built and sited to accelerate interconnection for a large-scale data center."¹

The result, per Aligned CEO Andrew Schaap: the facility comes online "years earlier than would be possible with traditional utility upgrades."¹

One deal does not make a trend. Is this a pioneering move in an emerging category, or an unusually creative one-off? How does the economics hold up using correct scale-segmented cost figures? And what do the counter-signals -- gas still dominating behind-the-meter strategy, regulatory frameworks still being codified, FEOC supply-chain uncertainty -- say about how fast this can scale?

This article builds the case for both sides. The honest verdict, supported by the best available public evidence: emerging trend, strong analytic conviction, thin named-deal pipeline.

The Aligned/Calibrant Deal: What Actually Happened

The Calibrant BESS is front-of-meter (FTM), placed on the grid side of the meter, not inside the data center fence. PGE manages it via its Dispatchable Standby Generation (DSG) program.¹ The data center itself is 72 MW -- the 31 MW battery represents approximately 43% of peak load capacity. The system enables Aligned to ramp from initial power to full-rated load as grid capacity becomes available, with the battery providing flexibility to the utility during that ramp period.

The mechanism: the BESS discharges during peak demand periods, reducing the net draw on a constrained transmission node. By showing that the load can be curtailed or self-served during peak constraint windows, the operator allows PGE to approve the full-rated interconnection without waiting for transmission capacity expansion. The battery is not providing backup power. It is providing grid flexibility -- a service to the utility, managed by the utility, that earns the data center earlier access to the grid.

Andrew Schaap framed this precisely: "With this BESS, we're converting our load from a potential grid liability into a dynamic grid asset, providing the regional utility with the tools needed to accelerate our ramp."¹

Phil Martin (Calibrant CEO) was equally direct: "Rather than the false choice between waiting years for system upgrades or having to go off grid entirely, we're working with leading data center providers like Aligned to use distributed energy solutions to facilitate and accelerate grid interconnection."¹

The Aligned/Calibrant project sits within a broader PGE and GridCARE program. GridCARE's DeFlex platform combines generative AI demand forecasting with flexible resources -- batteries and onsite generation -- to model and approve data center interconnection at constrained nodes without waiting for transmission builds. The program has already enabled 80 MW of incremental data center capacity in Hillsboro in 2026, with a 400+ MW pipeline through 2029.⁹ Multiple operators are participating. This shifts the Aligned deal from "one creative company" to "one anchor deal in a utility-managed program."

One additional data point on who made this decision: Aligned Data Centers is being acquired by a BlackRock Global Infrastructure Partners and Artificial Intelligence Infrastructure Partnership (AIP) entity. AIP's co-founders are Microsoft, NVIDIA, and MGX.²² The world's largest infrastructure investor and the dominant AI chip and cloud companies chose grid-adjacent BESS as the interconnection solution. That is not a startup experiment.

The Queue Problem Is Structural, Not Temporary

The interconnection delay that makes the Aligned strategy valuable is not going away. Lawrence Berkeley National Laboratory's 2025 "Queued Up" report documents the trend: average queue time hit 55 months in 2024, up from 36 months in 2015 and 22 months in 2008.² The total active US interconnection queue has grown to approximately 2,290 gigawatts.² Over 700 GW withdrew from queues in 2024 alone -- operators walking away after years of waiting.

For large loads, the numbers are harsher. ERCOT received 198 GW of large-load interconnection applications in a single quarter in 2026, with the roughly 86 GW under active review alone equal to ERCOT's entire current peak load.¹¹ In constrained markets like Northern Virginia and Silicon Valley, interconnection queues extend to seven years or longer.¹⁷ As we documented in our analysis of the grid bottleneck reshaping the data center construction map, the Northern Virginia queue alone stretched to seven years in 2025 -- the direct cause of the first construction slowdown since 2020.

A data center goes from groundbreaking to powered infrastructure in 12 to 24 months.²⁰ The grid connection takes three times as long. "Speed to power is the biggest prerequisite" for US data center developers. Every operator wants faster interconnection. The question is which instrument achieves it.

A December 2025 whitepaper from Princeton University ZERO Lab, Camus Energy, and encoord quantifies one alternative: a 500 MW data center using flexible grid connections (FGC) combined with a bring-your-own-capacity (BYOC) structure reaches full operation three to five years faster than traditional interconnection.¹⁰ Applied to real PJM transmission data, the analysis puts the cost savings at $78 million per gigawatt of new demand for FGC alone; the full FGC plus BYOC combination saves $404 million per GW.¹⁰

The NPV Reframe: What the Battery Is Actually Worth

The standard framing for BESS investment centers on energy arbitrage: buy cheap, sell expensive. For the Aligned model, that framing is almost irrelevant. The value driver is the net present value (NPV) of powering racks two to three years sooner.

Getting the math right requires scale-correct cost figures, not residential or blended-market pricing.

The Aligned/Calibrant system is a large, front-of-meter, utility-coordinated BESS -- a category where pricing is materially lower than commercial or residential systems. NREL's 2025 updated cost projections benchmark utility-scale 4-hour systems at approximately $210/kWh.²³ Wood Mackenzie's H1 2025 US utility-scale energy storage pricing report puts post-tariff large FTM systems at $230 to $320/kWh.²⁶ The comparable figure for a small commercial C&I system runs $250 to $450/kWh; residential installations are $700 to $1,300/kWh.²⁵ The scale premium is large and well-documented.

For the 62 MWh Aligned/Calibrant system, this produces a capex range of approximately $13 to $20 million, with a mid-point around $17 million.²³, ²⁶ Not $25 million (what initial research suggested using C&I-scale pricing), and not residential pricing.

Against that capex, the revenue opportunity from earlier interconnection:

The 2025 global average colocation rate was approximately $217 per kilowatt per month, or $2.6 million per MW per year, per datacenterHawk's 2026 pricing guide.¹⁶ Constrained markets like Northern Virginia and Silicon Valley run $250/kW per month or above.¹⁶ Landgate's analysis of 5 MW data center facilities puts the range at $1.1 to $2.8 million per MW per year depending on utilization and market.¹⁵

For Aligned's 72 MW facility, two years of earlier commercial operation at the $2.6M/MW/year global average, discounted at 10%, produces approximately $140 million in present value.¹⁵, ¹⁶ Against battery capex of $13 to $20 million, the NPV ratio is 7 to 10 times. The battery does not need to earn a single dollar in energy arbitrage. The interconnection acceleration alone generates NPV multiples of the hardware cost.

The conservative scenario holds just as well. At $1.5 million per MW per year and just one year of earlier operation, the 72 MW facility gains $108 million in revenue against a $17 million battery investment. The math is durable across scenarios because the battery is cheap relative to the revenue at stake.

One market analyst captured the dynamic: "Interconnection delays push delivery to the right -- you're not just delaying a build, you're delaying the revenue curve."¹⁵

Read more

On April 2, 2026, the Minnesota Public Utilities Commission handed Xcel Energy a landmark first-round win. The commission approved Capacity*Connect (200 megawatts of utility-owned, front-of-meter battery storage sited at businesses across Xcel's Minnesota service territory) at a total program cost of $430 million, or roughly $2,150 per kilowatt (kW).¹, ² The customer-owned aggregator alternative was deferred, pending pilot data.³

The decision immediately drew a challenge from a coalition of solar and aggregator interests: MnSEIA, SEIA, the Coalition for Community Solar Access (CCSA), Sunrun, Vote Solar, and others. Their central exhibit was not a hypothetical. It was Xcel's own Active Virtual Power Plant (AVPP) program in Colorado, mandated under Colorado Senate Bill (SB) 24-218, with third-party aggregators enrolling customer-owned behind-the-meter (BTM) resources, which delivers comparable dispatchable capacity at $624 per kilowatt.⁴, ⁵

Same utility. Two programs. A 3.45x cost gap.

Correction and clarification (August 26, 2026): After publication, Pier Lafarge, CEO of Sparkfund (Xcel's deployment partner for Capacity*Connect), provided a detailed response that prompted three changes to this piece.

1. The original version put Capacity*Connect's 20-year cumulative equity return at approximately $434 million. That figure did not account for depreciation of the rate base; a utility earns its return on the net rate base, which declines as the asset depreciates. The corrected figure is approximately $218 million.

2. The headline comparison sets Minnesota's all-in, 20-year program cost ($2,150 per kW, including administration, operations, maintenance, augmentation, and warranty) against Colorado's capacity-only, 5-year payment ($624 per kW, which excludes Colorado program administration and the hardware costs borne privately by enrolled customers). Normalized for term and scope, the gap is materially smaller than 3.45x.

3. We initially foregrounded the aggregator coalition's benefit-cost estimate of 0.96 (costs slightly exceed benefits). Xcel's own filed cost-benefit analysis reached a narrow net-positive result, crediting ratepayers with the grid benefits that utility ownership socializes across all customers. The benefit-cost ratio is genuinely contested and close to break-even.

The risk-allocation and competitive-procurement questions this piece raises still stand, but they belong on normalized numbers. Our thanks to Pier for the rigor.

Correction (August 29, 2026): A reader flagged that this piece described IRA Section 25D as running “through 2032.” It does not. The One Big Beautiful Bill Act (Public Law 119-21, July 4, 2025) terminated the residential credit for expenditures made after December 31, 2025. We have corrected the six passages that treated it as an ongoing subsidy.

This one cuts against our own argument, so we will state it directly. Section 48E, the credit that flows to utility-owned storage like Capacity*Connect, survived and runs through 2033. Section 25D, the credit that lowered the cost of the customer-owned hardware aggregators enroll, is gone. Federal tax policy now favors the utility ownership model on new build, which is the opposite of the asymmetry this piece originally described. What survives is narrower and, we think, still correct: the more than five million residential solar installations completed in the United States before 2026 were built with roughly 30 percent of their cost absorbed federally, through Section 25D for customer-owned systems and the commercial credit for third-party-owned ones. That fleet keeps the subsidy permanently, and an aggregator can contract against it without raising a dollar of capital to build it. The cost comparison at the center of this piece is unaffected, because it measures what each utility pays per kilowatt of dispatchable capacity, not what the customer paid for the hardware.

Our thanks to the reader who caught it.

The coalition did not invent a Minnesota counter-proposal. The $624 per kilowatt figure originates from Xcel's Colorado regulatory filings in Docket 24A-0547E; the aggregators introduced it into Minnesota's Docket 25-378 as a benchmark for what Xcel had already demonstrated it could deliver in a different regulatory environment.⁴, ⁶ The argument was pointed: if Xcel can run a 125 MW aggregator-led virtual power plant (VPP) in Colorado for less than a third of the per-kilowatt cost it proposes in Minnesota, what exactly is the Minnesota ratepayer paying for?

The PUC's answer, on close reading, is more nuanced than a simple utility win. The commission approved Capacity*Connect as a pilot, imposed four specific data-collection conditions with a November 2027 deadline, and explicitly noted that lessons from Xcel's Colorado program should be evaluated for Minnesota application. What looked like a regulatory door closing was, on the terms of the order itself, a door with a built-in expiration date.

This is Part 3 of Avanza's VPP Summer Series. Part 1 examined the feeder-first distribution optimization layer. Part 2 covered hyperscaler behind-the-meter capacity contracts and PJM's wholesale market mechanics. Part 3 is the state-PUC layer: the fight over who owns the battery at the distribution node, who earns the return, and who controls the grid's distributed future at the retail level.

The stakes are concrete. The US VPP market reached approximately 37.5 gigawatts (GW) of flexible BTM capacity in 2025.⁷ The Department of Energy (DOE) projects 80-160 GW by 2030, with potential grid cost savings of $10 billion per year — a market we first mapped in our foundational VPP analysis.⁷ Minnesota just set the first precedent for the utility-ownership model. The reasoning behind that precedent is the instruction manual for what every other state PUC will face.

The Fight That Got Here

The Capacity*Connect proposal arrived in October 2025. Xcel announced 200 MW of 1-3 MW utility-owned battery systems sited at businesses and nonprofits, deployed by 2028, with Sparkfund as the deployment services partner.⁸ The batteries connect at primary voltage (4-25 kilovolts), dispatch via Xcel's supervisory control and data acquisition (SCADA) system, and land on the rate base as fully capitalized assets earning Xcel's authorized return on equity (ROE) of 9.60 percent.⁹

The aggregator coalition filed its formal challenge on February 19, 2026.⁴ The core argument: Xcel had already demonstrated, in its own Colorado service territory, that aggregator-managed BTM resources could deliver dispatchable VPP capacity at $624 per kilowatt, 71 percent cheaper than what it was proposing to build and own in Minnesota. The coalition called Capacity*Connect "the only approved distributed storage program in the country with a cost-benefit ratio below one."⁶ The MnSEIA, SEIA, and CCSA joint statement on April 2 (the day the PUC approved the program) sharpened the indictment: "Competitive markets for energy storage deployment ensure that ratepayers get the best, most affordable deal possible. The Capacity*Connect program takes the exact opposite approach," said Andrew Linhares, SEIA Midwest Director.⁶

The commission approved Capacity*Connect on April 2, 2026. The order did not declare the aggregator model dead. It imposed four conditions that define what happens next.

The Economic Exhibit That Will Not Go Away

The cost gap deserves precise examination, because it will appear in every subsequent VPP rate case in the country.

Capacity*Connect's $2,150 per kilowatt figure represents utility-owned, 1-3 MW front-of-meter batteries at distribution nodes, fully engineered, interconnected, and integrated into Xcel's SCADA. The program cost is 100 percent capitalized; ratepayers fund the asset and fund Xcel's guaranteed equity return for the asset's useful life.⁹

Xcel's Colorado Active VPP is a fundamentally different procurement model. Under SB 24-218, Colorado prohibited Xcel from serving as its own aggregator, a single statutory provision that forced the competitive design.⁵ Third-party aggregators (Itron and Tesla) enroll existing customer-owned batteries, smart thermostats, EV chargers, water heaters, and heat pumps under performance-based capacity contracts. Xcel pays aggregators a service fee when they deliver capacity. No new hardware construction. Zero capital addition to Xcel's Colorado rate base.⁵, ¹⁰

The resulting cost: $624 per kilowatt for 125 MW of dispatchable capacity. As of January 2026, 15-plus MW was actively managed, with 5-plus MW deliverable during peak dispatch events.¹⁰

The 3.45x gap is real, but it requires one important nuance. These programs do not deliver identical products. Capacity*Connect delivers utility-grade primary-voltage infrastructure with exclusive Xcel dispatch authority and a utility performance guarantee.¹¹ The Colorado AVPP delivers aggregated customer resources with performance-based pay for actual dispatch events. The relevant question for ratepayers is not "are these identical products?" but "what does each cost per kilowatt of dispatchable capacity delivered, and who bears the financial risk?" On cost per kilowatt, the gap is 3.45x. On risk allocation, Capacity*Connect places all performance risk on ratepayers: the utility earns its authorized return regardless of whether the program meets projected grid benefits. The Colorado model places enrollment and dispatch performance risk on aggregators, who get paid only for capacity they actually deliver.

The coalition and Vote Solar calculated a benefit-cost (B/C) ratio of approximately 0.96 for Capacity*Connect, meaning projected costs slightly exceed quantified benefits.¹², ¹³, ⁶ That figure is a coalition estimate, not an independent third-party study. Xcel's own filed cost-benefit analysis reached the opposite conclusion, a narrow net positive, counting the ratepayer benefits utility ownership socializes across all customers: avoided distribution upgrades, avoided cost of new entry, capacity value, and ancillary services. The two analyses bracket break-even, and no neutral third-party study exists. What is not in dispute: the Minnesota Attorney General filed separately in December 2025 noting that Xcel never conducted a direct cost-benefit comparison of utility ownership versus customer or third-party ownership.¹²

The question for every other state PUC now considering a similar program: with the benefit-cost case this close to break-even, and contested, what should tip the decision toward utility ownership over a competitive procurement?

Why the PUC Said Yes: The Steelman Case

The commission's approval makes more sense when you understand the structural constraint it was working within.

Federal Energy Regulatory Commission (FERC) Order 2222, issued in September 2020, removes barriers to distributed energy resource (DER) aggregators participating in organized wholesale electricity markets.¹⁴ But Order 2222 requires each regional transmission organization (RTO) to finalize its own compliance tariffs, and those implementation timelines vary enormously. PJM's wholesale capacity market integration went live in June 2026, with the 2028/29 BRA the first auction open to DER aggregations. ISO-New England: energy and ancillary services November 2026, capacity market 2027. The Midcontinent Independent System Operator (MISO), which covers Minnesota, lags its peers with a phased schedule: an initial DER-participation phase targeted for June 2027 and full aggregation market participation by June 2029.¹⁴ We covered the revenue potential these Order 2222 markets unlock for BTM asset owners — up to $120,000 per MW per year in PJM capacity markets — which explains exactly why the wholesale market gap matters so much for MISO-state aggregator economics.

That gap is material. Until MISO finalizes its Order 2222 compliance, Minnesota BTM aggregators have no wholesale capacity market revenue pathway. Aggregator programs in PJM states (New Jersey, Maryland, much of the Mid-Atlantic) can stack utility capacity contract payments with wholesale market revenues. Minnesota aggregators in 2026 cannot. Their entire economics depend on what a state program pays them. Without the wholesale revenue layer, the aggregator financing case is materially thinner in MISO states than in PJM or California ISO (CAISO) states.

This is also why Order 2222's cousin, FERC Docket RM26-4, is largely irrelevant to this fight. FERC issued Section 206 show-cause orders to all six RTOs on June 18, 2026, with responses due August 17, 2026, covering five reform categories including "flexible load and BTM service" definitions.¹⁵ RM26-4's primary focus is hyperscale data center load integration at 20 MW or larger, a different jurisdictional track that does not modify Order 2222, does not govern distribution-level battery ownership, and does not preempt state PUC decisions on rate-base treatment. Conflating RM26-4 with the utility-vs-aggregator ownership fight is a category error.

The PUC's logic, read on its own terms: approve the utility pilot, collect actual performance data from Capacity*Connect's first phase, use that data to evaluate whether the aggregator alternative makes sense in a MISO state without a wholesale revenue path, and revisit the question in 2027. That is a "we need evidence" decision, not a "utility wins forever" decision.

The aggregator coalition disputes even this reading. In their April 2 statement, MnSEIA, SEIA, and CCSA characterized Capacity*Connect as a "closed pilot" with "no meaningful pathway to collect comparative data from third-party developers or competitive market participants."⁶ The tension: the PUC ordered data collection and independent evaluation, but the exclusive-utility-dispatch design means the only data collected will be from Xcel's own program, not from a head-to-head comparison with aggregator-managed resources. Both characterizations are on the record.

The full analysis continues below for paid subscribers: the rate-base math behind the utility's incentive, the IRA advantage compounding for customer-owned assets, the 2027 reopener chessboard, the state-by-state competitive landscape, and strategic recommendations.

Read more

Three years ago, a 100 MW data center faced a PJM capacity obligation of about $1.05 million per year. Annoying, but manageable. Today that same data center faces a $12.2 million annual obligation — and a 1 gigawatt (GW) hyperscaler deployment faces $121.7 million.¹ That 11.5-fold increase did not happen gradually. It happened suddenly, across three consecutive auction cycles in which PJM's Base Residual Auction (BRA) cleared at or near the FERC-approved price cap, culminating in the 2027/28 delivery year price of $333.44 per megawatt-day (MW-day).¹

At $29/MW-day, self-procuring wholesale grid capacity through a fleet of residential thermostats and commercial batteries was an interesting experiment. At $333/MW-day, it became a strategic imperative. As we documented in our analysis of PJM's capacity market repricing and its implications for data center load flexibility, this price escalation has moved demand-side participation from a theoretical opportunity to a financial necessity for any hyperscaler operating at scale in PJM territory.

On June 2, 2026, Google and Voltus announced the deal that proved it: a 100 MW, three-year Bring Your Own Capacity (BYOC) agreement covering PJM delivery beginning in 2027 — the first hyperscaler-aggregator wholesale capacity deal in U.S. regional transmission organization (RTO) history.² Google funds a distributed energy resource (DER) fleet of residential and commercial sites across PJM's footprint. Voltus aggregates and dispatches that fleet. The resulting accredited capacity flows to Dominion Energy, Google's load-serving entity (LSE), satisfying Google's capacity obligation through a private bilateral contract, not through PJM's standard auction process.

This is not a demand response program. It is not a sustainability initiative. It is a hyperscaler using virtual power plant (VPP) technology to self-procure wholesale grid capacity and cap its own exposure to a market that has repriced 10 times in three delivery years.

The significance of this deal is not the 100 megawatts (MW). It is who is paying.

The Buyer Has Changed

For the first five years that FERC Order 2222 existed on paper, the demand-side actor in every VPP aggregation deal was a utility. A utility faced a capacity obligation, contracted with an aggregator to manage customer demand, and the aggregator enrolled households and businesses into demand response programs. The utility was the buyer. The utility set the terms. The utility determined the revenue.

The Google/Voltus deal inverts this structure entirely.

Alphabet, not Dominion Energy, is writing the check. Google has a PJM capacity obligation because it draws enormous quantities of power from Dominion's load zone in northern Virginia. Rather than paying Dominion's standard capacity charge — which flows from the BRA clearing price — Google contracted directly with Voltus to deliver accredited Unforced Capacity (UCAP) to Dominion on its behalf. Dominion's obligation is satisfied. Google's cost is the contracted Voltus fee rather than the spot capacity charge.

The utility is not eliminated from the equation. Dominion is still the LSE, still the entity that must hold sufficient capacity, and still the settlement counterparty with PJM. But Dominion is no longer the revenue source for the aggregator. Google is.

That structural shift carries four compounding implications that make this deal more significant than its 100 MW scale.

The first is credit quality. Alphabet carries an AA+ credit rating and over $95 billion in cash on its balance sheet.² A three-year fixed-price contract with Alphabet is a materially different financial instrument than a utility program contract subject to annual renewal. Aggregators can underwrite longer terms, higher commitment levels, and deeper capital investment in DER enrollment when the counterparty is a Fortune 50 corporation with essentially zero default risk.

The second is the disintermediation of utility VPP programs. Utilities including Eversource, Dominion, and ComEd have built demand response and VPP programs over the past decade, enrolling residential and commercial customers. Those programs paid customers to reduce load during utility-defined peak events. BYOC routes those same customers — or new ones who didn't qualify for or want utility programs — into a corporate-funded aggregation where the event triggers are PJM Performance Assessment Intervals (PAIs), not utility peak windows. The hyperscalers that utility VPP programs coveted as large load participants now enroll in parallel channels that bypass the utility revenue model entirely. Our foundational analysis of the VPP market documented how only 19.5% of distributed energy resources were enrolled in VPP programs under the utility-buyer model — the corporate-buyer structure Google just demonstrated may be the mechanism that changes that figure.

The third implication is aggregator revenue certainty. Traditional aggregation depends on winning PJM capacity auction clearings, which can reprice dramatically between auctions. FERC's new price collar set a $325/MW-day cap and a $175/MW-day floor for the 2028/29 auction — but when that auction cleared on July 14, 2026, it hit the $325 cap, not the floor, extending the near-cap price environment rather than relieving it.⁴ A contracted Voltus-to-Google fee insulates Voltus from that repricing exposure. The aggregator that closes hyperscaler BYOC deals today locks in the current price environment in long-term contracts and avoids the next repricing cycle.

The fourth implication is the White House math. In March 2026, seven technology companies — Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI — signed the White House Ratepayer Protection Pledge, committing in principle to "build, bring, or buy" new generation to satisfy data center energy demands and bear the full costs of grid infrastructure they require.⁵ That is seven signatures and one executed deal. Google has translated a policy commitment into a specific commercial mechanism. The other six remain in the "advocate for interconnection reform" posture. The gap between signing a pledge and structuring a bilateral capacity procurement deal is where Voltus is building its competitive position right now.

Why This Deal Happened in June 2026

Three conditions converged simultaneously that made this deal possible in this window.

The price signal crossed the rational-action threshold. At $333/MW-day, a hyperscaler with 3 GW of PJM-territory data center capacity faces a capacity obligation approaching $365 million per year. Conservative BYOC estimates suggest deals that deliver accredited capacity at 20 to 35% below the BRA clearing price would generate $70 to $125 million in annual savings at that scale. When the savings opportunity is measured in nine figures, the transaction costs of structuring a bilateral aggregation deal become trivial.

The regulatory pathway became available. FERC approved the DER Aggregator Participation Service Agreement in PJM — which had been pending as of April 28, 2026 — enabling the bilateral contract structure through which Voltus transfers accredited UCAP to Dominion without routing through the standard BRA process.³ Without this mechanism, the Google deal could not settle properly within PJM's administrative framework. This was a legal enablement, not a market innovation.

Voltus had the product, the supply chain, and the track record. The nine months from September 2025 — when Voltus launched its BYOC product in partnership with data center developer Cloverleaf Infrastructure⁶ — to the June 2, 2026 Google announcement were not sales months. They were capability-building months.

The Voltus/Octopus Energy partnership in February 2026, which brought residential device aggregations from smart thermostats, home batteries, and electric vehicles (EVs) across PJM, MISO, New York, and California into Voltus's supply chain,⁷ addressed the most binding near-term constraint: only 0.4% of registered demand response at PJM is residential. The BYOC market at hyperscaler scale requires residential DER supply that no single aggregator had previously assembled.

When Google went to market looking for a counterparty who could absorb the accreditation risk and structure a contract that looked like every other power purchase agreement (PPA) they had ever signed, Voltus was — in their own words from Latitude Media coverage — "the only provider in the region up for absorbing the accreditation risk and able to provide a contract that looked like every other PPA we've ever signed."⁸

That framing matters. Hyperscalers have sophisticated procurement desks that negotiate 3-year fixed-price agreements for server hardware, cloud bandwidth, and construction services. A PPA-like capacity contract fits that procurement model exactly. An aggregator that can speak the language of corporate procurement — fixed price, defined term, performance guarantee, creditworthy settlement — wins deals that technically capable aggregators fluent only in utility program language will lose.

The full analysis continues below for paid subscribers, including the precise mechanics of how BYOC works (and what it is not), the ELCC upgrade that made it more capital-efficient, the $18.3 million non-performance question, the economics at every scale from 100 MW to 1 GW, the regulatory constraints that cap scaling below 500 MW, the competitive landscape analysis, the policy trajectory from 27 states advancing mandates to Ireland's regulatory endpoint, and strategic implications by stakeholder type.

Read more

There is a feeder in the South Bay Area of California that has a problem you cannot see on any system operator dashboard. CAISO's grid is fine. The system peak is within tolerance. By every traditional metric, the grid is working.

But at 9:47 PM on a Tuesday in November, that feeder's transformer is running at 103% of its rated thermal capacity. Twenty-three EV owners came home from work, plugged in, and added to the load profile of a residential circuit engineered in 1998, when the heaviest load was a central air conditioner running on a summer afternoon. Nobody dispatched anything to fix it. No demand response event was called. The overload will repeat tomorrow night, and the night after, and every night until something gives — either the transformer, or the utility's capital plan.

This is the real grid reliability problem of 2026. Not the system peak on four hot August afternoons. The feeder, every night, in your neighborhood.

The traditional demand response (DR) paradigm was built for the system operator's problem: curtail load across an entire service territory when summer heat drives demand to its annual maximum. It generates roughly 100 event-hours per year. The other 8,660 hours of the year, the incentive to flex is zero. Meanwhile, a landmark 2024 study in the Proceedings of the National Academy of Sciences found that 50% of California's 5,000+ distribution feeders will face EV-driven overload by 2035 — overload that peaks at 9-11 PM in residential neighborhoods, hours outside the 2-7 PM CAISO system peak window.¹ The cost to upgrade those feeders with conventional infrastructure: $6 to $20 billion.¹ As we documented in our deep dive on EV charging and distribution infrastructure, Norway avoided that entire capital outlay at 94% EV penetration through managed charging — the feeder-first logic applied at national scale.

The constraint has moved. The tools have not kept up.

Feeder-first virtual power plants (VPPs) are the architectural response to that mismatch. Rather than dispatching assets across an entire service territory in response to a system-level signal, feeder-first programs target specific neighborhoods and circuits, dispatching continuously across all 8,760 hours to relieve local thermal and voltage constraints before they become equipment failures or capital expenditures. This is not an incremental improvement on the old DR model. It is a different problem statement.

The economics already validate the concept in the markets that have tried it. Con Edison's Brooklyn-Queens Demand Management (BQDM) program deferred $1 billion to $1.2 billion in traditional distribution infrastructure using just 52 MW of geo-targeted demand management resources at a program cost of $69.9 million, a leverage ratio of 14:1 to 17:1 on every dollar spent.² New York's locational pricing signal for feeder relief, $140.76 per kilowatt per year, published monthly by substation and feeder zone, is the only deployed example of its kind in the United States.³

The question is whether the policy, technology, and business models exist to unlock this everywhere, at the speed the feeder overload trajectory demands.

The Old Model and Its Limits

Demand response was designed for a simpler era of grid stress. In the 1990s and early 2000s, the grid's binding constraint was straightforward: summer afternoon peaks driven by cooling load, predictable by season and weather, coincident across the entire service territory. The DR program design reflected that reality — enroll large commercial and industrial (C&I) customers, pay them to curtail load when the system operator called an event, reset, repeat next summer.

It worked because the assumption at its foundation was true: distribution feeder stress occurred at the same time as transmission and generation stress. The thermal peak on the feeder and the capacity peak on the system were coincident. You didn't need to target individual circuits because the problem was everywhere at once.

That assumption has been eroding for years. It is now structurally false.

Three converging load types have created constraints that are local, non-coincident with system peak, and running year-round:

Electric vehicle charging creates residential feeder peaks in the 9-11 PM window — after residents return home, after solar generation has ramped down, well after the system's late-afternoon peak has cleared. California's residential EV charging peaks in this window; CAISO's system peak is a 2-7 PM event driven by cooling load as solar ramps down.¹ These are different problems occurring at different times on different physical assets.

Data centers create large, rapid load additions at specific points of interconnection on distribution circuits. A 5 MW edge computing facility at the end of a suburban feeder changes the thermal load profile of that circuit fundamentally. When that data center's workload spikes — or when it fails and a neighboring facility absorbs its load — the feeder constraint appears in minutes, not the hours that traditional DR programs are designed to respond within.

Whole-home electrification — heat pumps replacing gas furnaces, induction ranges, heat-pump water heaters — converts gas demand into electric load on residential circuits that were not sized for it. A neighborhood where 40% of homes have replaced their furnaces with heat pumps carries a winter heating peak on its residential feeders that the distribution planner from 2005 never modeled.

The National Renewable Energy Laboratory's (NREL) Advanced Hosting Capacity Analysis (AHCA) framework documents this divergence in detail: time-varying hosting capacity analysis shows that adjacent feeders in the same substation can differ in hosting capacity by a factor of 10, and a feeder can be simultaneously unconstrained at system peak and severely constrained at 10 PM on a mild Tuesday.⁴ Static hosting capacity assessments — which most utilities still use — treat both conditions as equivalent and systematically underestimate feeder stress.

The result is a program design that misses the problem it was built to solve. Traditional DR dispatch events called by CAISO to manage the system's 2-7 PM summer peak do nothing to relieve the transformer running at 103% of capacity in that South Bay neighborhood at 9:47 PM. The customers enrolled in that program are not being dispatched when and where the constraint exists. The grid value is uncaptured.

Where the Feeder Constraint Actually Lives

The PNAS study by Lin et al. provides the most quantified picture of what's coming.¹ Across more than 5,000 California distribution feeders, under projected EV adoption curves:

• 7% of feeders face overload risk from EV charging today, in 2025

• 27% will face overload risk by 2030

• 50% will face overload risk by 2035

• 67% will face overload risk by 2045

The trajectory is not gradual. The jump from 27% to 50% happens in five years.

Residential feeders require twice as many upgrades as commercial feeders, because overnight residential charging — not workplace charging — carries the dominant load growth. Uncontrolled nighttime residential charging is, by the study's analysis, the most severe EV charging scenario for feeder infrastructure of any modeled option.

The infrastructure cost to address this conventionally: $6 to $20 billion for 25 gigawatts (GW) of needed distribution capacity additions in California alone.¹ The DOE's 2025 Virtual Power Plants Liftoff Update estimates that VPPs at the 80-160 GW national target would save approximately $10 billion annually in avoided generation buildout and deferred infrastructure investment.⁵ California is a microcosm of the national arithmetic: managing feeder constraints through traditional infrastructure costs an order of magnitude more than managing them through aggregated flexible assets. Our original VPP market analysis documented that same leverage in the Brooklyn-Queens program — $200 million in DER investment versus $1 billion in traditional infrastructure, an 80% cost reduction — but that program operated at the distribution substation level, not individual feeders. Feeder-first takes the locational logic one level deeper.

The policy and program apparatus has not caught up with these numbers. California — the most technically advanced VPP state, with the most detailed feeder-level data infrastructure in the country — cut funding to its Demand Side Grid Support (DSGS) program for 2026, the primary driver of 785 MW of 2025 capacity growth. Governor Newsom vetoed AB 740 in October 2025, the bill that would have mandated a statewide VPP deployment roadmap. The most EV-stressed state in the union entered 2026 without a statutory mandate or guaranteed funding for the distribution-level response to its own feeder overload problem.

National VPP enrollment has reached 37.5-38 GW of capacity — impressive growth, but less than 20% of deployed distributed energy resources (DERs) are enrolled in VPP programs at all.⁶ The constraint is not DER availability. It is incentive structure. Existing DR incentives pay for system-peak coincidence; they leave locational feeder value entirely unmeasured and uncompensated. The 80% of deployed DERs not enrolled need a different signal to participate — one denominated in feeder-level constraint relief, not system-level event response.

The full analysis continues below for paid subscribers, including the feeder-first VPP architecture, the economics of locational dispatch, the business model shift that makes it commercially durable, the regulatory landscape across all major ISO territories, and strategic implications for utilities, aggregators, and investors.

Read more

In December 2025, PJM Interconnection held its most expensive capacity auction in history. The grid operator, which serves 67 million people across 13 states and Washington, D.C., cleared at $333.44 per megawatt-day, the federally approved price cap, for the third consecutive year. Total cost: $16.4 billion.¹, ²

That number alone would make this story worth telling. The more important detail sits inside it: $6.5 billion of that $16.4 billion, 40% of the entire auction, was attributable to data center load, according to PJM's own Independent Market Monitor (IMM).¹, ³ Of that $6.5 billion, $6.2 billion came from data centers not yet built.¹

Ratepayers across PJM territory are paying today, in their current electricity bills, for capacity reserved to serve facilities that may never come online. If those data centers are never constructed, the money doesn't get refunded. The grid operator came up 6,625 megawatts short of its own reliability target, the first miss in the capacity market's 18-year history, in the same auction where costs hit their record.², ⁴

Paying more and getting less. That is the story.

The three-year price trajectory makes the scale concrete. Just two auction cycles earlier, the 2024/2025 delivery-year auction cleared at $28.92 per megawatt-day. The 2027/2028 auction cleared at $333.44 per megawatt-day, a 1,053% increase.¹, ⁵ For average households in PJM territory, the NRDC projects bill increases of roughly $70 per month by 2028 compared to pre-surge levels.⁶ In Washington, D.C., Pepco customers saw $10 per month added to bills from the 2025/2026 delivery year alone.⁷

How a Capacity Market Actually Works

To understand why this matters, and why reform is now inevitable, you need to understand what PJM is actually buying in these auctions, and how a number in a load forecast becomes a line item on your electricity bill.

PJM operates the Reliability Pricing Model (RPM), which runs forward auctions called Base Residual Auctions (BRAs). Each BRA procures capacity three years in advance for a specific delivery year — generators commit to be available during peak demand in exchange for capacity payments. The December 2025 auction procured capacity for June 2027 through May 2028.⁸, ⁹

The auction works in five steps.

Step 1: Set the requirement. PJM calculates how much capacity it needs by taking its forecasted peak load and adding an Installed Reserve Margin (IRM). The IRM is calibrated to a reliability standard called Loss of Load Expectation (LOLE), the probability of an outage event. PJM targets no more than a once-in-a-decade extreme weather event.¹⁰ For the 2027/2028 delivery year, that translated to a total requirement of approximately 152,102 megawatts, roughly 127,000 megawatts of forecasted peak load plus a 25,000-megawatt reliability buffer.

Step 2: Convert resources to reliability value. Every resource bidding into the auction converts its nameplate capacity to an Unforced Capacity (UCAP) value via an Effective Load Carrying Capability (ELCC) factor, the statistical probability the resource will be available during peak stress hours.⁹ A gas peaker with a reliable operating history earns more UCAP credit than a solar farm with variable output. Demand response resources cleared 7,299 megawatts of UCAP in the 2027/2028 auction after PJM updated demand response's ELCC factor from 69% to 92%.

Step 3: Run the auction. Generators submit bids. PJM clears where supply meets a downward-sloping demand curve called the Variable Resource Requirement (VRR) curve, which pays more when supply falls short of the reliability requirement and less when it exceeds it. The curve has a hard cap at the Net Cost of New Entry.⁹

Step 4: Hit the cap. When the market cannot satisfy the demand curve below the cap price, the auction clears at the cap. In 2027/2028, that cap was $333.44 per megawatt-day. PJM cleared there for the third straight year and still came up 6,625 megawatts short of its requirement.², ⁴

Step 5: Allocate costs. Load-serving entities, the utilities and retail suppliers who serve end customers, pay based on their Peak Load Contribution (PLC), calculated as each customer's average demand during the five highest-demand hours of the prior summer.¹¹ Those costs pass through to retail electricity bills as monthly capacity charge riders. Every PJM ratepayer pays, residential, commercial, industrial.

The mechanism most people overlook is the one connecting load forecasts to monthly bills: the forecast doesn't just predict what customers will consume. It determines how much capacity PJM is required to buy. If PJM includes speculative data center load in its forecast, it must procure capacity to serve that load, even if those data centers never get built.

PJM's IMM stated this directly in its auction analysis: "The extreme uncertainty in the load forecasts based on uncertainty about the addition of large data center loads is also unique and unprecedented and raises questions about the meaning of clearing a capacity auction based on those forecasts."¹

The Scale of PJM's Data Center Footprint

PJM's problem is geographically concentrated and then geographically generalized. Both things at once.

Northern Virginia, anchored by Amazon's "Data Center Alley," sits in PJM's Dominion Zone. In the 2025/2026 auction, the first year the price spike became visible, the Dominion Zone cleared at $444.26 per megawatt-day versus $269.92 per megawatt-day for PJM as a whole.⁵ That gap was the market's signal that the Northern Virginia supply-demand imbalance was acute. By the 2026/2027 and 2027/2028 auctions, the price cap had spread uniformly across all PJM zones. The zone-specific imbalance had become a systemwide problem.

PJM's 2026 Long-Term Load Forecast projected 32 gigawatts of peak load growth from 2024 to 2030, with 94% of that growth attributed to data centers.¹² The annualized summer peak growth rate reached 3.6% per year, versus the 0.3% rate in PJM's 2021 forecast, before data center proliferation reset the numbers. In the Dominion Zone alone, the 2025 forecast called for more than 20,000 megawatts of data center growth by 2037, versus just 5,700 megawatts in the 2022 forecast, a 3.5x upward revision in three years.¹²

Wood Mackenzie analysis put collective utility forecasts for new large load growth in PJM at 55 gigawatts by 2030 and 100 gigawatts by 2037.¹³

The load forecast revision also contains a self-correcting mechanism that arrived too late. In late 2025, PJM trimmed its summer 2028 peak load forecast by 4.4 gigawatts and its summer 2027 forecast by roughly 4 gigawatts after implementing stricter vetting through the Load Adjustment Request Implementation framework.¹⁴ PJM accepted only 34 gigawatts of the approximately 60 gigawatts submitted by utilities for 2030, a 43% rejection rate. The rationale: developers routinely submit redundant interconnection requests across multiple zones to hedge against transmission constraints, inflating the apparent pipeline. PJM's pre-2025 process had no mechanism to distinguish a fully-financed, shovel-ready project from a speculative site assessment.

But the 2027/2028 auction used the older, higher forecasts before those corrections were applied. The record $16.4 billion cost was locked in before PJM could update its numbers.¹ That is the mechanism of the speculative load trap: the correction lags the commitment.

The $6.2 Billion Asymmetry

Here is the structural injustice that has turned this into a policy crisis rather than just an economics story.

PJM's IMM calculated precisely: including forecast data center load, versus only the load from data centers already operating, increased total RPM revenues by $6,244,523,827, a 61.4% increase above what the auction would have cost without speculative data center load.¹ That number, $6.2 billion, is the financial consequence of one decision: whether speculative demand should drive capacity procurement.

The asymmetry runs in one direction only. Ratepayers pay for capacity reserved to serve speculative load. If the data centers materialize, the capacity was needed. If they don't, the capacity was procured unnecessarily, and the cost has already been collected and will not be refunded. The data center developer, whose unverified request drove up the requirement, faces no financial penalty.

Three mechanisms compound the problem.

The forecasting gap is structural: utilities in data center-dense zones submitted enormous pipeline numbers reflecting their customers' announced project lists, which include everything from fully-financed construction to early-stage site assessments. PJM's pre-2025 vetting process could not distinguish between them.

The interconnection hedge amplifies the problem: data center developers routinely submit redundant requests across multiple utility service territories to protect against transmission constraints in any one zone. Each request contributes to the load forecast that drives the capacity requirement.

The cost pass-through completes the trap: capacity charges flow from ratepayers to utilities to PJM to generators. The NRDC described the destination of those payments precisely: "Virtually none of this rate hike will go to investments in new power plants and infrastructure; instead, almost all of it will be windfall payments to owners of existing power plants, which face no meaningful competition."⁶ As our analysis of the utility rate structure problem documented, this cost pass-through dynamic — where large loads impose costs that residential customers absorb — is a feature of how PJM allocates charges, not a bug.

The primary financial beneficiaries of elevated PJM capacity prices are owners of existing generation, nuclear, natural gas, and coal plants that collect capacity payments without requiring new investment. Constellation Energy and Vistra Energy, two of the largest existing-capacity owners in PJM, have each cited positive capacity market results in investor communications.

Paying More, Getting Less: The Reliability Shortfall

The capacity market is designed to guarantee reliability. That guarantee broke down in December 2025.

PJM's auction procured 134,479 megawatts of UCAP from generation sources. Adding Fixed Resource Requirement areas brought total secured capacity to approximately 145,479 megawatts. The reliability requirement was 152,102 megawatts. The shortfall: 6,625 megawatts.⁴

The achieved reserve margin was 14.8%, against a 20% IRM target. This was the first time since the capacity market launched in 2007 that PJM failed to meet its reliability requirement across the entire region, including areas with their own local capacity plans.², ¹⁵

FERC Chairman Laura Swett called it "very concerning." FERC Commissioner David Rosner called it "unacceptable" and demanded PJM file meaningful rule changes to address the generation interconnection and financing barriers driving the supply gap.¹⁶ The market worked exactly as designed. Supply responded to the highest prices in its 18-year history. It simply could not grow fast enough to meet a demand projection that was partly fictional.

The supply-side constraint amplifying the crisis is PJM's interconnection queue. Average wait time from application to commercial operation grew from under two years in 2008 to over eight years in 2025. In 2025 alone, developers cancelled 38 gigawatts of projects, more than ten times the amount of new generation that successfully connected that year.¹⁷ Network upgrade costs reached $240 per kilowatt in 2020-2022, up from $29 per kilowatt in 2017-2019.¹⁷ The combination — soaring costs and years-long queues — means that even at $333.44 per megawatt-day, the price signal has not produced enough new supply to close the gap. Our February 2026 analysis of the interconnection crisis showed how these same queue dynamics were already forcing hyperscalers toward behind-the-meter generation as the only path to timely power delivery.

The NERC 2026 Summer Reliability Assessment noted that PJM has adequate installed reserves for summer 2026, but projects margins will fall below requirements by 2029 without accelerated supply additions.¹⁸ Bloomberg analysis estimates PJM could face a supply shortfall of up to 60 gigawatts over the next decade if major capacity additions don't materialize.¹⁹

The price signal is working against reliability, not for it. The same forecast that drove costs to a record produced a shortfall. The market is not malfunctioning — it is optimizing correctly against a requirement that is partially based on speculation.

The full analysis continues below for paid subscribers, including the Google/Voltus BYOC escape valve, detailed regulatory reform landscape, two scenarios for how this resolves, and strategic implications by stakeholder — with comprehensive citations.

Read more

$163 billion. That is the cumulative cost PJM ratepayers face from 2028 through 2033 if the current interconnection system continues unchanged — roughly $70 per month added to the average household utility bill by 2028, according to the NRDC.²⁹ It is the downstream consequence of a simple, broken process: data centers need power, the grid cannot connect them fast enough, and the cost of that failure lands on everyone except the operators causing it.

On October 23, 2025, DOE Secretary Chris Wright invoked Section 403 of the DOE Organization Act, a rarely-used statutory tool that compels FERC to act, and directed the commission to fix data center interconnection. FERC issued its Order Regarding Intent to Act on April 16, 2026.² Then, on June 18, 2026, it delivered: tailored show-cause orders under Section 206 of the Federal Power Act to all six regional grid operators it oversees, directing each to either justify why its current large-load tariffs remain just and reasonable or file changes that fix them.³² The proceeding that determines how every data center above 20 MW connects to the U.S. interstate transmission system is now live, and it has a clock attached.

FERC did not pick a winner. It did something the market had not priced: instead of one national rule litigated for two years, it gave each grid operator 60 days to defend or rewrite its large-load tariffs, and teed up five specific reforms spanning cost allocation, co-location, and flexible load.³³ The instrument is faster than a rulemaking and more surgical than a policy statement. The consequence most analysts will miss: it accelerates grid-connected and behind-the-meter data center strategies at the same time. Both bets just became more investable, not one at the other's expense.

The scale of what is at stake

U.S. data center power demand reached 61.8 GW in 2025, a 22% year-over-year increase, and S&P Global projects it rising to 134.4 GW by 2030.¹¹ That trajectory requires connecting hundreds of new facilities to a transmission system whose interconnection queue now holds more than 2,060 GW of backlog and averages nearly five years of wait time, up from under two years in 2008.⁷ Of all projects requesting interconnection between 2000 and 2019, only 19% reached commercial operation by the end of 2024.⁷

Moody's projects $3 trillion in global data center infrastructure investment through 2030.²⁸ Hyperscalers committed more than $350 billion in 2025 and approximately $400 billion in 2026. Every dollar of that capital faces a single bottleneck: grid connection.

The financial pressure is most acute in PJM Interconnection, the grid operator covering 13 states plus the District of Columbia. The December 2025 PJM capacity auction, for the 2027/28 delivery year, cleared at the FERC-approved price cap of $333.44/MW-day, generating $16.4 billion in total capacity costs. Data centers drove $6.5 billion of that, 40% of the total, and the auction still fell 6,623 MW short of PJM's reliability requirement.¹⁰ For context: the 2024/25 auction cleared at $28.92/MW-day. Clearing prices rose more than tenfold in a single year. The line between "data center power demand" and "everyone's utility bill" is direct and quantifiable.

The cost allocation dispute sits at the center of RM26-4. A UCS analysis found that of 130 large load interconnection projects in PJM, only 6 (5%) required the requesting customer to pay their own connection costs directly.⁸ The remaining 95%, representing more than $4 billion in annual grid upgrade costs, were socialized across all ratepayers. The 100% participant funding model proposed under RM26-4 would transfer that burden from households to operators. GridLab calculates that if just 10% of PJM's queued projects had connected in time for the 2026/27 auction, clearing prices would have fallen to $254/MW-day, a $3.5 billion reduction.¹⁹ The system cost of the interconnection bottleneck is not theoretical. It is showing up on utility bills now.

The queue is not the only signal: 56 GW is already building its own power plants

Before examining what FERC just decided, consider what the market decided first. Cleanview identified 46 planned U.S. data centers representing 56 GW of combined behind-the-meter (BTM) capacity, approximately 30% of all tracked planned U.S. data center capacity.²⁰ Ninety percent of those projects were announced in 2025. Nearly 75% are powered by natural gas.

The BTM economics are structural, not speculative. On a 20-year levelized basis, a combined-cycle natural gas BTM solution costs $2.9 billion per 1 GW facility versus $6.2 billion for a solar-plus-storage grid-connected alternative, more than twice as expensive.²¹ The interconnection cost alone drives much of that gap: $24/kW for gas BTM versus $253/kW for solar and $335/kW for offshore wind.²¹ When grid connection costs more and takes longer than building your own plant, developers build their own plants. That is not irrational behavior. It is rational capital allocation responding to a broken process. As we documented in our analysis of data centers building their own power plants, the BTM economics were compelling before RM26-4 was filed — the June 18 show-cause orders determine how much further the gap widens.

The committed capital behind BTM tells the story: Williams Companies has committed $5.1 billion to data center gas-power infrastructure; VoltaGrid, Energy Transfer, and Oracle have committed 2.3 GW of off-grid gas power; Brookfield and Bloom Energy have committed up to $5 billion to solid-oxide fuel cells for data center BTM; Wärtsilä has more than 2.4 GW of U.S. data center BTM orders for 2025-2026.²⁰ Midstream capital expenditure was up 61% in 2025 and is projected up 23% in 2026. This capital is not contingent on FERC's ruling. It is already deployed.

The SPP HILL framework: proof the fast-track works

The critical piece of context for evaluating the June 18 orders is that a fast-track interconnection pathway already exists and is already operational. FERC unanimously approved the Southwest Power Pool (SPP) High Impact Large Load (HILL) framework on January 14, 2026.¹⁵ SPP's HILL process delivers interconnection agreements within 90 days for qualifying loads, cutting the national five-year average to a project-finance-compatible timeline.

The technical requirements for HILL participation define what "flexible" means in practice: hourly load forecast data provided in real time to SPP; remote disconnect capability (the grid operator, not the data center, initiates curtailment); ramp rate limits not to exceed 20 MW per minute; and full compliance with ride-through requirements, meaning the load stays connected through brief voltage and frequency deviations rather than auto-disconnecting.¹⁵ The companion HILLGA process pairs on-site generation with a HILL, delivering interconnection in 150 days with double security deposits, up to $500,000 for requests over 500 MW, and a five-year agreement after which the generator must enter the standard process.

SPP's HILL framework proves 90-day study timelines are achievable. The June 18 show-cause orders push this concept to all six RTOs, directing each to justify its current process or file a comparable fast-track threshold for loads that commit to these technical requirements.

The instrument FERC chose: a 60-day clock, not a two-year rulebook

Markets spent months handicapping two instruments. A Notice of Proposed Rulemaking (NOPR) would have proposed binding national rule text, opened a comment period, and produced an enforceable Final Rule in 12 to 24 months. A policy statement would have set interpretive principles with no binding timeline, leaving each ISO to move at its own pace. FERC chose neither. It issued tailored show-cause orders under Section 206 of the Federal Power Act to PJM, MISO, SPP, CAISO, ISO-NE, and NYISO, plus their transmission owners, each in a docket of its own.³²

The mechanism matters because it inverts the burden of proof. Under Section 206, FERC does not propose a rule and ask whether the industry agrees. It declares the existing tariffs presumptively unjust for large loads and orders each operator to defend them or replace them within 60 days, by August 17, 2026. A resource-adequacy report, detailing how each operator will secure generation to serve new large loads, is due in 30 days.³³ Commission staff reviewed more than 3,500 pages of comments before acting. The result is faster than a NOPR and carries more compulsion than a policy statement.

Each order tees up the same five categories of reform:³³

• Efficient study and transmission-service processes, including alternative transmission technologies

• Preventing cost shifting and requiring transparency into transmission costs

• Accommodating co-location and behind-the-meter generation

• Providing new transmission services for flexible large loads

• A process to study generation serving co-located and electrically proximate loads

Two design choices define the investment consequence. First, FERC held the jurisdictional line exactly where the market expected: the orders guard against cost shifting among transmission customers but leave retail cost allocation, generation siting, and permitting to the states, and they explicitly protect large-load transmission agreements already nearing completion.³³ Second, by tailoring orders to each operator instead of imposing one national rule, FERC preserved the regional divergence that drives the behind-the-meter trend, while pushing every operator toward faster, cost-transparent grid connection.

That combination is why this is not a binary. The five categories advance the grid-connected fast-track and entrench the behind-the-meter economics at once. The framework below maps both pathways, because after June 18 the question is no longer which outcome to position for. It is how fast you can position for both.

The full analysis continues below for paid subscribers, including detailed economic analysis, competitive landscape assessment, regulatory implications, and strategic recommendations with comprehensive citations.

Pathway 1: The grid-connected fast-track, now on the clock

Read more

On May 19, 2026, the Denver City Council voted unanimously to impose a one-year moratorium on new data center zoning permits. Not a rural county commission, not a suburb without political infrastructure — the capital of Colorado, acting as one, stopping the industry cold. Council President Amanda Sandoval apologized from the dais for not blocking the CoreSite facility that triggered the vote when the company sought a $9 million tax incentive 18 months earlier. "My biggest regret on council," she said in public session.²²

Denver wasn't first. Moratoriums had been spreading through county commissions and small city councils for two years. But Denver marked a category change. It proved the opposition movement had grown sophisticated enough to capture urban governance, where the political power to block, regulate, and shame is considerably greater than in any rural jurisdiction. Seattle followed within weeks — three council members introducing an emergency moratorium after 93,270 constituent emails opposed five proposed data center campuses totaling 369 megawatts (MW) of demand.²³

$130 billion in U.S. data center projects were blocked or delayed by community opposition in Q1 2026 alone — a single quarter that roughly matched the total for all of 2025, according to a June 2026 study from Data Center Watch.⁵⁴ Project cancellations quadrupled from 6 in 2024 to 25 in 2025, with 20 more killed in Q1 2026 alone, a record quarterly pace.¹⁷ Active opposition groups more than doubled in a single quarter — from 396 at the end of 2025 to 833 across 49 states by March 2026. "In some cases," the authors wrote, "opposition mobilized before any project was officially filed — the mere rumor of a data center was enough to trigger organized resistance."⁵⁴

Against that backdrop, hyperscalers are collectively planning to spend approaching $1 trillion on data center capital expenditure in 2026 — with roughly $150 billion committed to onsite private power generation alone.²⁷ The industry is betting a trillion dollars on a permitting environment that voted unanimously against it in Denver last month.

This article documents that crisis in full — and then makes the case that the crisis is not inevitable. The evidence is clear: developers who adopt proactive community engagement and binding sustainability commitments before opposition forms have materially better permitting trajectories than those who don't. The differentiated path exists. What the industry does with that information in the next 12-18 months will determine whether it gets regulated differently in every state, or whether it sets its own floor before legislators do it for them.

The Structural Crisis: Five Numbers That Tell the Story

1. Public opinion has inverted

A Gallup poll conducted in March 2026, covering 1,000 adults across all 50 states, found that 71% of Americans oppose an AI data center being built near them — 48% strongly.⁶ Only 27% support new data center siting nearby.

The Virginia numbers are even more instructive because Virginia is the world's data center capital, and the trajectory is stark. In 2023, 69% of Virginians said they were comfortable with a new data center in their community. By April 2026, that number was 35% — a 34-percentage-point collapse in three years.²⁶ Fifty-seven percent of Virginia residents now believe data centers are negatively impacting their home energy bills. Sixty-seven percent want to end data center sales-tax exemptions — the same population that was 61% supportive of those exemptions in 2023.²⁶

Data centers now rank below nuclear plants in community acceptance.⁶ That is not a communications problem. It is a structural opposition crisis.

2. The tax math has exploded

The political bargain at the center of data center development — tax breaks in exchange for jobs and investment — has collapsed numerically.

Georgia estimated its data center tax breaks would cost $327 million in 2023. The actual cost in fiscal year 2026 is $2.5 billion, a 7.6-fold increase in three years.¹⁶ Virginia's data center tax exemptions now cost $1.94 billion per year, and the state legislature is considering ending a $1.6 billion annual sales-tax break.⁷ Texas revised its FY2025 projection from $130 million to $1 billion in just 23 months.²⁵ Nationally, experts estimate data center tax incentive costs have reached $90 billion per year, though 14 states fail to disclose these figures at all — in violation of standard accounting requirements.¹⁶

Good Jobs First executive director Greg LeRoy's assessment: "No form of state spending is more out of control today than data center tax abatements."¹⁶

The problem for developers isn't just that the numbers are large. It's that the jobs promised to justify the breaks haven't materialized at the scale implied. A typical 100 MW hyperscale campus — representing $500 million or more in capital investment — creates 850 construction jobs over 18 months and then employs 100-200 permanent operations staff.⁷ The Brookings Institution's analysis of 770 U.S. data center facilities found that "the standard model of data center development has produced mostly short-term construction jobs in recent years and relatively little long-term, high-value tech activity or large-scale employment."⁷ In colocation counties, tax incentives represent 62% of total investment — the breaks are largest precisely where the employment benefits are smallest.

The rate impact is the other side of this bargain. Our analysis of how large load growth shifts costs onto residential customers found that Virginia's own legislative auditor projects $444 per year in residential bill increases by 2040 — from a population where 57% already believe data centers are raising their bills.

3. The legislative environment has become hostile

More than 300 state data center bills were filed across 30-plus states in just six weeks in early 2026, shifting from incentive-focused legislation to regulatory oversight.⁴³ Twelve states have active moratorium bills in session.¹⁵ New York's Responsible Data Center Development Act passed both chambers of the state legislature on June 5, 2026 and now awaits Governor Hochul's signature — if signed, it will be the first statewide moratorium in U.S. history.⁵ Federal legislation from Sen. Bernie Sanders (I-VT) and Rep. Alexandria Ocasio-Cortez (D-NY) — S.4214 — would impose an immediate federal construction moratorium until Congress addresses energy costs, safety, and worker protections.⁴

Voters in at least five jurisdictions will decide data center ballot measures in 2026.¹⁹ Ohio organizers are pursuing a statewide initiative that would ban data centers above 25 MW.²⁰

4. Water is the issue with the hardest ceiling

Water use is cited in more than 40% of contested data center projects, making it the single most common driver of opposition.¹⁴ U.S. data centers directly consumed 17.4 billion gallons of water in 2023, equivalent to roughly 160,000 American households.² The EPA projects that figure rising to 38-73 billion gallons by 2028.² The largest hyperscale facilities can consume up to 5 million gallons per day on hot days.¹¹

Geographic concentration amplifies the problem. Two-thirds of data centers built or in development since 2022 are located in water-stressed areas — the Colorado River Basin, southern Arizona, Texas.⁴⁰ The Houston Advanced Research Center's January 2026 analysis projects Texas data center water consumption reaching up to 161 billion gallons annually by 2030, against an aquifer recharge rate in the Trinity system of just 4-5%.⁵²

Hood County, Texas crystallizes what this looks like on the ground. Eight proposed data centers spanning 7,600 acres sit above a recharge zone for the Upper Trinity Aquifer, within the watershed feeding Dinosaur Valley State Park. One project alone would consume 1 million gallons of water per day at full buildout. County commissioners attempted moratoriums twice; both failed after developer lawsuits and state legal warnings. The county has been sued twice by developers for attempting to exercise local control.²⁴

This is not an abstract resource question. For communities that draw water from wells above stressed aquifers, opposition to data center siting is a direct response to a direct physical threat.

5. Noise is the grievance that creates plaintiffs

Data centers emit continuous industrial noise from cooling systems, generators, and fans — 24 hours a day, seven days a week, including during sleeping hours. Monthly backup generator load tests produce 80-100 decibels at the property line, comparable to a leaf blower or motorcycle, running 30 minutes to two hours each.¹² Most county noise ordinances were written to address parties and construction, not permanent 24/7 industrial operations.

In April 2026, residents of the Sacred Heart and Back Central neighborhoods in Lowell, Massachusetts filed the first data center lawsuit in the state, challenging the air permit for an expansion of Markley Group's 352,000-square-foot facility with 27 diesel generators. The site ranks in the 97th percentile nationally for nitrogen oxide (NOx) exposure and the 90th percentile for adult asthma rates. Yale Law School's Environmental Justice Law and Advocacy Clinic is representing the plaintiffs.¹⁸

In New Jersey, the first national class-action noise lawsuit targets DataOne USA — a subsidiary of Nebius Group, which signed a $17 billion deal to supply Microsoft AI computing power — for industrial humming that approximately 1,000 residential households within one mile describe as constant and damaging.⁵³

Both cases will produce legal templates that communities across the country will replicate.

The Inflection Point

The Denver and Seattle actions in spring 2026 signal something specific: the opposition movement has crossed a threshold from rural grievance to urban political strategy. Cities have different tools than counties — stronger legislative infrastructure, closer ties to state lawmakers, more media attention, and organized constituent networks. The 93,270 emails that reached Seattle's city leaders in days represent the kind of constituent mobilization that changes careers.

The implication for site selection is severe. Urban and suburban locations — which typically offer the fiber density, power infrastructure, and low-latency positioning that hyperscalers need — now carry moratorium risk that did not exist six months ago. The JLL 2026 Global Data Center Market Outlook now ranks "community support" as the number two site selection criterion, behind only speed to power.²¹ That placement in the JLL framework — institutional real estate analysis used by the firms deploying capital — is a clear signal that the industry's financial infrastructure has absorbed what the technical teams haven't fully priced in.

The strategic question is not whether the opposition will intensify. It will. The question is whether the industry will set its own standards before 300-plus state bills create an incoherent regulatory patchwork, or wait until each state writes its own rules.

The Differentiated Path: Three Pillars

Developers who adopt proactive community engagement and binding sustainability commitments before opposition forms have materially better permitting trajectories than those who don't. The evidence is concrete, and in the case of water, the technical problem is already solved.

Pillar 1: Water efficiency — the gap is 12x and it's closed at the frontier

The industry average water usage effectiveness (WUE) for evaporative cooling is 1.8 liters per kilowatt-hour (L/kWh). AWS is operating at 0.15 L/kWh — 12x better, using closed-loop chip-level cooling deployed at commercial scale. Microsoft went further: in August 2024, it deployed a zero-water cooling design that eliminates evaporative cooling entirely and saves 33 million gallons of water per facility per year. The pilot operates in Phoenix, Arizona, ground zero for data center water opposition in a drought-stressed basin. That deployment is the reference point regulators and communities in constrained markets use when evaluating what is operationally feasible.

The technical problem is solved. The industry has not adopted these approaches as the standard. That gap — between what is achievable and what most developers still propose — is precisely what 300-plus state bills are trying to close by force.

Pillar 2: Engagement before opposition forms — the mechanism is transparency, not money

DBT Data President David Tolson described the approach simply: they committed to sharing full project plans with residents before filing anything — not after. In a community that had already rejected five prior data center applications, this produced a complete reversal from active opposition to community support, driven entirely by process transparency.³⁹

Early engagement prevents the next moratorium. Converting hardened opposition — once community groups have organized and a narrative has calcified — is far more expensive and far less reliable.

Pillar 3: Binding commitments over secrecy — and the standards gap

Lancaster, Pennsylvania's community benefit agreement (CBA) — the most specific data center CBA in the United States by the assessment of Columbia Law's Climate Law Blog — includes a hard 20,000-gallon-per-day water cap and $20.25 million in binding cash commitments.³² Researchers who study it still consider it a work in progress, because the community received the agreement two days before the council vote.

The EU Climate Neutral Data Centre Pact has 100-plus signatories with binding power usage effectiveness (PUE) and renewable energy targets. The U.S. has no equivalent. That gap is the direct cause of 300-plus bills in 30-plus states. When state legislators cannot point to a credible industry-produced floor, they write their own — and each state writes something different.

Below: five documented case studies of what proactive engagement produces in practice, the complete water efficiency frontier, what CBAs require to create durable community acceptance, and how the EU regulatory model forecasts the U.S. trajectory.

What "Good" Actually Looks Like: Five Cases Worth Studying

Read more

On July 10, 2024, a lightning arrestor failed on a 230-kV transmission line somewhere in the Eastern Interconnection.⁴ Grid protection systems operated exactly as designed (nothing unusual there). Faults happen dozens of times per year. But what followed was unlike anything NERC's engineers had seen in 58 years of bulk power system oversight.

Within seconds of the fault, approximately 1,500 megawatts (MW) of load vanished from the grid.⁴ Industrial loads typically take minutes to shed, following operator responses to process conditions. Motor-driven equipment takes seconds, decelerating through stored rotational energy. This shed in milliseconds, faster than a human blink. Grid frequency overshot to 60.047 hertz (Hz). Voltage surged to 1.07 per unit. System operators scrambled to remove shunt capacitor banks to restore balance.⁴

No utility equipment tripped. No operator commanded a curtailment. Every megawatt of that 1,500 MW shed itself, automatically and in unison, because data center uninterruptible power supply (UPS) systems across a regional footprint simultaneously detected the voltage disturbance and switched to battery backup. Each one executing its protection logic in under four milliseconds.

NERC's official summary of the event: "The electric grid has not historically experienced simultaneous load losses of this magnitude in response to a fault on the system."⁴

That sentence, written in a January 2025 incident review, is the origin of a regulatory cascade that culminated on May 4, 2026, when NERC issued only the third Level 3 Essential Actions Alert in its 58-year history.¹ The subject: computational loads. The audience: every Transmission Planner, Transmission Owner, Balancing Authority, and grid reliability coordinator in North America. The message: the grid has a new category of threat, it is growing at a pace that outstrips existing reliability standards, and immediate action is mandatory.

Level 3 is NERC's highest alert tier, not a warning or a recommendation. Both previous Level 3 Alerts preceded binding mandatory Reliability Standards.¹³ The compliance clock is running.

What the Coverage Gets Wrong

August 3, 2026 is not a compliance implementation deadline. It is a reporting deadline.⁹

By August 3, registered entities must submit responses to 33 structured questions documenting the status of their activities relative to the seven Essential Actions. Failure to respond is a Rule 810 violation, enforceable by FERC.² But failure to have fully implemented the Essential Actions by August 3 is not independently enforceable. Not yet.

The real enforcement teeth arrive when NERC Project 2026-02's bridge standard completes. NERC's drafting team has targeted the bridge standard filing for December 31, 2026, with FERC approval making the standards binding in 2027 or 2028.⁶ At that point, the Essential Actions become Reliability Standards with the full weight of NERC's penalty framework behind them, up to $1 million per day per violation.¹² NERC's highest assessed penalty to date: $1.54 million per day.¹²

Three things hang on this distinction. Operators who read August 3 as an implementation deadline and scramble for compliance theater will miss the more important 2026-2028 standards development process. The 33-question responses form the first-ever public accounting of how far the industry actually is from compliant architecture; NERC will publish an anonymized aggregate report. And that compliance gap directly informs Project 2026-02 stringency: the further the industry falls from compliance in August, the more aggressive the eventual standard.

The Scale That Made the Alert Inevitable

The physics of the July 2024 event would have been concerning at any scale. At today's scale, they are an emergency.

U.S. data centers drew approximately 50 gigawatts (GW) in 2024.¹⁷ BloombergNEF's December 2025 forecast (a 36% upward revision from the projection issued just seven months earlier, the largest single-cycle revision in the history of that forecast) projects 106 GW by 2035, representing 8.6% of all U.S. electricity demand.¹⁷ Globally, capacity stands at approximately 103 GW today and is expected to nearly double to 200 GW by 2030, requiring up to $3 trillion in investment.¹⁶

NERC's January 2026 Long-Term Reliability Assessment projects summer peak demand across the bulk power system will grow 224 GW over the next 10 years, a 69% increase over the 2024 forecast driven primarily by data centers.¹⁰

Two geographies carry the highest regulatory exposure:

ERCOT (Texas): 45 GW of data center interconnections forecast by 2030, with 23 GW specifically attributable to data centers. In the first quarter of 2026 alone, 198 GW of large load applied for interconnection in ERCOT, with 86 GW under active review.¹⁶

PJM (Mid-Atlantic/Midwest): Projected to add 30 GW of data center load between 2025 and 2030.¹⁰ Dominion's service territory, home to approximately 35% of global colocation supply in Northern Virginia, is the single highest-concentration zone in North America and the location of documented oscillation events.³

At this scale, a normally-cleared transmission fault is no longer a routine event. It is a potential grid stability crisis. The buildout driving this exposure, now exceeding $150 billion in committed onsite and behind-the-meter capacity across five simultaneous technology buckets, is the physical substrate beneath every compliance obligation this alert creates.

Why Computational Loads Are Categorically Different

Most coverage treats the July 2024 event as a large-load story. It isn't. It's the first documented encounter between bulk power system fault physics and a load category with no precedent in reliability engineering.

Four characteristics make computational loads categorically different from every prior large-load type:⁵

Voltage sensitivity at millisecond timescales. Data center UPS systems follow the IEC 62040-3 standard and the ITI/CBEMA curve, which specify that a voltage sag below approximately 0.7 to 0.8 per unit triggers automatic transfer to battery within one to three electrical cycles (17 to 50 milliseconds). A typical transmission fault causes voltage to dip 0.25 to 0.40 per unit for 40 to 60 milliseconds. These timescales were engineered independently, one for IT equipment protection and one for grid fault clearing, but they are identical. The fault and the UPS protection response operate on the same clock. The UPS always wins.

Synchronized simultaneity. Unlike industrial loads with varying sensitivities and operator response times, data centers using the same UPS equipment and protection settings trip simultaneously. A regional fault doesn't trip one data center. It trips all of them, within the same fault window. The July 2024 event affected multiple facilities across an entire regional footprint in a matter of seconds.

Zero inertia contribution. Traditional industrial loads include large motors and rotating machinery that store kinetic energy and resist rapid frequency change, what engineers call system inertia. Power-electronic UPS-based data center loads contribute zero inertia. As renewable generation displaces synchronous generators, the grid carries less and less rotational inertia to buffer rapid frequency events. Data centers are simultaneously the largest new load category and the category with zero inertia contribution. That combination produced the physics of the July 2024 frequency spike.

Customer-initiated disconnection. None of the 1,500 MW shed in July 2024 was disconnected by utility equipment or operator command. Every megawatt was shed by customer-side protection systems optimized for IT equipment survival, not grid stability. Grid operators had no visibility into the event, no warning it was coming, and no authority to stop it.

Sub-synchronous oscillation adds a fifth failure mode. GPU training workloads run in synchronized compute cycles that create rhythmic power demand fluctuations, pulsing the grid at sub-synchronous frequencies. NERC's Reliability Guideline documents oscillation events from cryptocurrency mining in ERCOT and from data center operations in Dominion territory, both confirmed on the bulk power system.³ As GPU cluster sizes grow, AI training workloads amplify this risk.

The Strategic Paradox at the Core of This Alert

In the same month NERC issued its most severe alert about data centers destabilizing the grid, NERC's 2026 Summer Reliability Assessment reduced the ERCOT net internal demand forecast by 3.7 GW (4.6% of the total), specifically because "more data centers can be curtailed by grid operators when needed to prevent grid emergencies."¹¹

The liability and the asset are the same physical infrastructure. What separates them is architecture, contracts, and a willingness to engage with grid operators as participants rather than passive loads.

That wedge, between a data center that is a grid liability and one that is a grid asset, is now a regulatory, economic, and investment variable of the first order. Google has already signed 1 GW of data center demand response agreements with five U.S. utilities: Indiana Michigan Power, TVA, Entergy Arkansas, Minnesota Power, and DTE Energy.³¹ As we analyzed in our deep dive on load flexibility's breakthrough moment, integrated resource planning modeling of 1 to 2 GW of data center flexibility produces net present value savings of $300 to $400 million over 2025 to 2050, per NBER Working Paper 34065 on flexible data centers and grid costs.³²

The liability today. The asset tomorrow. The compliance floor in between.

The full analysis continues below for paid subscribers, including the seven Essential Actions decoded, what compliant versus non-compliant architecture looks like, the Computational Load Entity registration framework, FERC RM26-4's site selection implications, and strategic recommendations for operators and investors.

The Seven Essential Actions: Decoded for Operators and Investors

Read more

On May 4, 2026, NERC issued a Level 3 "Essential Actions" Alert -- its highest urgency classification — and the target was not a failing transmission line or an aging nuclear unit.¹⁹ It was data centers.

Grid operators had documented multiple incidents since 2022 in which more than 1,000 megawatts of computational load dropped off the bulk power system in seconds. Not because of grid failures — because data center protection systems tripped automatically, shedding load faster than real-time operators could respond. The frequency spikes that followed were the kind of event that, scaled up, cascades into regional blackouts. NERC's language was unambiguous: uncontrolled data center disconnects represent an "immediate risk" to the bulk power system.¹⁹ Transmission planners, system operators, and balancing authorities have until August 3, 2026 to implement seven specific compliance actions.

This is the grid telling the data center industry: your load behavior is no longer your private business decision.

What makes NERC's alert consequential beyond the compliance deadline is the timing. Six weeks earlier, on March 23, 2026, EPRI launched Flex MOSAIC at CERAWeek — the first standardized taxonomy for data center demand flexibility, built with 65+ stakeholders including Google, NVIDIA, Constellation Energy, National Grid, MISO, and CAISO.¹ The framework classifies large-load flexibility across four dimensions (magnitude, timing, duration, and frequency) and produces five DCFlex Flexibility Classes that transform flexibility from a bespoke bilateral negotiation into a standardized, contractable product.²

The combination is not coincidence. Grid operators need data centers to be predictably dispatchable. Data center operators need a commercial framework to get paid for that dispatchability. Flex MOSAIC is the bridge. For operators who built behind-the-meter (BTM) generation to escape the interconnection queue — which, as Avanza's earlier coverage of the BTM gas buildout documented, totals approximately 50 GW of announced capacity — the timing creates a specific, actionable opportunity that most are not yet positioned to capture.

The Narrative Inversion

Two years of energy coverage has framed data centers as the grid's problem: explosive load growth outpacing transmission buildout, interconnection queues stretching seven years, capacity prices spiking 262% in PJM's 2025 auction.¹² That framing is not wrong — as our analysis of the data center power paradox documented in detail, the structural forces pushing operators toward BTM generation are real and accelerating. But it misses what the current buildout is actually constructing.

The four cloud giants — Google, Microsoft, Amazon, and Meta — tripled their data center capital expenditure to $217 billion in 2024 from $69.4 billion in 2019, and 2026 hyperscaler capital spending will exceed $400 billion.⁵ Most of this spending includes substantial generation investment. Cleanview tracks 46 planned U.S. data centers intending to operate 56 GW of BTM capacity, representing 30% of planned U.S. data center capacity.⁸ Gas turbine orders for U.S. destinations reached 18 GW in the first half of 2025 alone.

What the industry is building, collectively, is the largest distributed generation fleet in U.S. history. Fast-responding, geographically distributed, dispatchable generation sitting adjacent to substations already at capacity. ACEEE estimates 60–200 GW of peak demand reduction potential from load flexibility over the next decade — up to double the most aggressive projections of total U.S. data center capacity by 2030.⁸ Only 6% of U.S. energy consumers currently participate in retail demand response programs.⁸

The gap between what is technically possible and what is commercially realized is enormous. The global demand response management system market will reach $25.92 billion by 2030 at 15.8% CAGR. The automated demand response segment was $2.77 billion in 2025, projected at $5.43 billion by 2032. Against a backdrop where only 6% of U.S. energy consumers participate in any retail demand response program, the data center sector's combined scale, controllability, and geographic concentration make it the single most actionable market for closing that gap fast.⁸

EPRI Flex MOSAIC exists to close it. But the story of what it actually does — and what stands between 50 GW of BTM generation and its first dispatched flexibility event — begins with understanding why that gap has persisted.

What Flex MOSAIC Actually Is

The framework's genius is structural rather than technical. Before Flex MOSAIC, a utility negotiating flexibility with a data center negotiated everything from scratch: What load can you shed? How fast? For how long? How often? Can you commit to that contractually? Each negotiation was custom, each contract was bespoke, and the result was a market that couldn't scale.

Flex MOSAIC creates five DCFlex Flexibility Classes defined by four measurable dimensions: magnitude (10–40% of facility load), timing (response latency from seconds to hours), duration (minutes to days), and frequency (event recurrence).¹ A Class 1 facility provides seconds-to-minutes response for frequency regulation. A Class 5 facility modulates load for hours, enabling economic dispatch and peak shaving. Each class carries specific, verifiable performance requirements — creating a credential that utilities can write standard contract language around.

This is the same conceptual shift NERC reliability standards made for generation assets decades ago. Once classified and quantifiable, flexibility can be priced, compared across vendors, aggregated across sites, and eventually mandated. For data center operators, Flex MOSAIC certification is the difference between a unique bilateral deal and a market product that can be securitized.

The proof of concept is already documented. In December 2025, Emerald AI and National Grid ran a live demonstration at the Nebius AI Factory near London — 96 NVIDIA Blackwell Ultra GPUs against real National Grid dispatch signals, not simulations. The result: 30% power reduction achieved in under 40 seconds from signal receipt, 22 live grid events, 100% compliance, zero SLA violations.⁴ High-priority workloads — inference, real-time applications — continued uninterrupted throughout. The 256-GPU Phoenix, Arizona demonstration at Oracle Cloud Infrastructure added further rigor: 33 separate experiments, 212 AI training jobs, two real utility peak events, and a simulator prediction accuracy of 4.52% root mean square error.⁶

One operator has already moved past demonstration into commercial contract. Google reached 1 GW of aggregate demand response capacity across long-term utility agreements with Entergy Arkansas, Minnesota Power, DTE Energy, Indiana Michigan Power, and Tennessee Valley Authority.⁷ The DTE Energy contract alone projects nearly $1.7 billion in positive affordability benefits for existing customers — with Google covering its full electricity costs and infrastructure while ratepayers benefit from avoided peak capacity investment.⁷ EPRI Flex MOSAIC is designed to make this commercial template accessible to operators without Google's legal budget and bilateral negotiating leverage.

AI workloads absorb significant power reductions without affecting the compute jobs that matter — zero SLA violations across 22 live events in London and 33 field experiments in Phoenix establishes that baseline. Google's 1 GW demonstrates the revenue model. The question is no longer whether it works. It is whether operators have built the infrastructure to participate.

Here is the central paradox: approximately 50 GW of BTM generation announced in 2025 physically cannot participate in the flexibility markets Flex MOSAIC is designed to serve. The barrier is not technical. It is regulatory. And it determines whether an operator's BTM generator is worth $60,000–$80,000 per MW per year in grid services revenue — or worth nothing beyond avoided grid bills.

That barrier explains why the most impressive flexibility demonstration in current literature happened in London rather than Northern Virginia, Dallas, or Chicago. It also determines which technology decisions, made right now at the site selection and permitting stage, are reversible — and which are not.

The EPA Barrier: Why the BTM Fleet Is Legally Sidelined

Read more

Every infrastructure vendor in the data center market is telling the same story right now: the era of owning your infrastructure is ending. Move to OpEx. Convert your capital expenditures into monthly service fees. Transfer the risk. Let experts handle operations. Focus on your core business.

The pitch resonates because the economics are real. Aggregate hyperscaler capital expenditure for 2026 approaches $725 billion — Microsoft deployed $34.9 billion in a single quarter in 2025, AWS $28.3 billion in the same period, Google $23.95 billion.¹⁰ The spend required to operate modern AI data center infrastructure at scale is beyond the capital position of most enterprises. For organizations without the financial depth, infrastructure operating expertise, or institutional scale that hyperscalers deploy, the OpEx model — risk transfer, embedded expertise, guaranteed outcomes — solves a genuine problem. The industry is moving this way for legitimate reasons.

Who it actually works for is the harder question, and where most conversations in this industry fall short. Earlier this month, I moderated a panel at Data Center Expo titled "From CapEx to OpEx: Re-Engineering the Commercial Model for Modern Data Centres." Three practitioners sat across the table: Martin Renkis, Executive Director of Data Center Infrastructure at Johnson Controls, who runs a Cooling-as-a-Service business with $6.7 billion in contractual infrastructure commitments. Jeff Wood, Vice President North America at Uptime Institute, two decades advising enterprises on operational risk, reliability standards, and what certifications are actually worth. And Ganesh Mani, Senior Director of Engineering at GEICO, who inherited a $300 million annual cloud bill, did the actual math, and went the opposite direction.

The panel produced no winner. What it produced was a concrete framework for understanding when each model is right — and what organizational conditions determine the answer.

What the enterprise buyer who ran the full math found

GEICO's story illustrates the in-house path clearly — it's the most rigorously documented public case from the enterprise buyer side, and the conditions that made it work are specific.

By 2021, GEICO had moved 80% of its workloads to public cloud, spread across eight different cloud providers, and ran a cloud bill north of $300 million per year.⁵ Then GEICO's engineering leadership did something the vendor ecosystem doesn't encourage: they built a fully-loaded total cost of ownership model from scratch. The analysis covered server acquisition, rack power and cooling, colocation space, lifecycle operations, support, software licensing impacts, and depreciation across a multi-year horizon.⁹

The conclusion was stark. By deploying owned Open Compute Project hardware — standardized, non-proprietary server architecture originally developed by Facebook — in colocation facilities using the ORv3 rack architecture, GEICO achieved more than 50% reduction in cost per compute core and more than 60% reduction in cost per gigabyte of storage, versus both prior on-premises systems and equivalent public cloud pricing.⁹

They deployed 1,000+ OCP servers across two colocation facilities, targeting 3,000 servers and repatriating at least 50% of cloud workloads by 2028–2029.⁹

The savings are documented. The Open Compute Project Foundation — a nonprofit whose members include Google, Microsoft, Meta, and the enterprise engineering leaders who built the infrastructure these companies run — published GEICO's analysis methodology in full. These are not marketing materials.

The same infrastructure decision, build versus buy, produces radically different outcomes depending on who executes it, how they're set up, and what they're optimizing for. GEICO had three things most enterprises don't: the operational talent to run what they were building, the capital to fund it, and the institutional experience to know what questions to ask before committing. Those three conditions are the real story.

Why the OpEx model exists and who it genuinely serves

OpEx and Cooling-as-a-Service models solve a specific problem, and that problem is real. The organizations these models serve are mid-market enterprises, regional operators, and infrastructure developers who have land access, power contracts, and committed offtake agreements — but lack the institutional depth to operate complex data center infrastructure at the unit economics large-scale operators achieve.

Renkis put the distinction plainly: hyperscalers don't need financing. They have capital and want execution. The OpEx pitch is for everyone else — companies that need capital flexibility, embedded operational expertise, and the performance guarantee structure a long-term service agreement provides.

Johnson Controls' Cooling-as-a-Service model is the clearest articulation of this structure. JCI finances, owns, and operates cooling infrastructure inside customer facilities. The customer pays a monthly service fee tied to financially guaranteed performance KPIs, with explicit penalties for misses. Average contract length is 12 years, with terms running up to 50 years — JCI claims a 99.98% success rate on financially guaranteed outcomes over 40 years of operation.

The market is large. The data center cooling market reached $22.1 billion in 2024 and will grow to $56.2 billion by 2030.¹² The chiller segment stands at $4.8 billion in 2025, with the hyperscale segment holding approximately 40% share.¹³ JCI has also invested in Accelsius, a liquid cooling company whose two-phase direct-to-chip technology claims 35% operational expenditure savings versus single-phase systems.2, 29

The broader Data Center as a Service market — colocation, managed services, and cloud together — reached $127.4 billion in 2025, with projections reaching $694 billion by 2034 at a 20.7% compound annual growth rate.¹¹ Whether you accept those exact figures or the materially different estimates from other research firms, the direction is consistent: the market for outsourced infrastructure management is large and growing.

Building and operating a modern AI data center — at 100 kilowatts per rack, with liquid cooling, with diesel backup managed against non-attainment zone air quality thresholds, with ESG reporting obligations and hyperscaler tenant demands — is not something most enterprise IT organizations can handle in-house. The OpEx model bundles expertise with capital. That bundle has real value.

What the panel pushed on — and what Mani's seat at the table made concrete — was the organizational prerequisite that sits in front of the financial analysis.

The governance prerequisite

The GEICO repatriation story is not primarily a story about cost. It's a story about organizational capability — and this is where most CapEx-versus-OpEx discussions go wrong, because capability doesn't fit neatly into a financial model.

On stage at Data Center Expo, Mani was direct: "Taking this in-house is not just a question of cost in TCO. You need governance structures." The panel returned to this point from multiple directions. Governance readiness is the prerequisite, not the outcome. An organization that takes infrastructure in-house without building the procurement capabilities, FinOps discipline, legal infrastructure, and operational maturity to execute will not simply pay higher capital costs. It will pay CapEx costs plus the cost of operating dysfunction simultaneously.

The OCP Foundation's white paper on GEICO's deployment makes this explicit: adopting open hardware "is more than a technical decision — it is an operational and cultural transformation that demands maturity across procurement, planning, legal, and engineering."⁹ That sentence is the most important one in the document. It's also the most likely to be skipped by anyone approaching the case as a simple financial argument.

What does governance readiness look like in practice? Mani described a multi-layer approach. GEICO built a chargeback and showback model — engineering teams are accountable for their consumption costs in real time, not at quarterly review. They developed FinOps discipline before moving the workloads. They ran TCO analysis that layered infrastructure, infrastructure services, platform, and application: four distinct cost layers that most enterprise TCO models collapse into one or two, producing understated total cost in both directions.

This methodology made the GEICO numbers credible. When their analysis showed 50% savings per compute core, those savings were calculated against a fully-loaded baseline covering server acquisition through software licensing through depreciation.⁹ Most organizations running a comparable analysis don't include all of those layers — which is exactly why both the "cloud is cheaper" and "on-prem is cheaper" arguments find adherents while describing the same set of costs.

The 37signals case provides a useful data point at the smaller end. The company behind Basecamp and HEY ran $3.2 million per year in AWS and Google Cloud costs. They invested approximately $600,000 to $700,000 in Dell server hardware, and their first full year of on-premises operation produced approximately $2 million in net savings, with the hardware investment fully recouped.¹⁴ Five-year projected savings exceed $10 million.¹⁵ No additional headcount required — the same technical team that managed cloud workloads managed the on-premises infrastructure.

37signals is a software company with deep technical capability and a well-defined workload profile. They had the governance prerequisites before making the decision. That's the point.

Organizations that fail at this transition share one characteristic: they ran the financial analysis first, then discovered the governance requirements after they had already committed.

TCO done right: what the four-layer model reveals

Mani's cost framework illustrates why conventional enterprise TCO models produce misleading conclusions.

Most financial analyses of build-versus-buy decisions treat "infrastructure cost" as a single line item.

Maybe they break it into hardware and software. More sophisticated analyses add operations. What they almost never include is the full stack: infrastructure, infrastructure services, platform, and application — each with distinct cost drivers that interact with the choice of commercial model differently.

Infrastructure: The physical layer — servers, storage, networking, power, cooling, physical space. This is where the headline CapEx-versus-OpEx comparison stops. At this layer, the analysis often does favor OpEx for organizations that lack the purchasing volume to buy OCP-class hardware at hyperscaler unit costs.

Infrastructure services: The operational layer — monitoring, maintenance, patching, hardware lifecycle management, incident response. This is where the expertise premium lives. Outsourced partners embed this cost in the service fee. In-house operators must build it. Organizations that underestimate this layer overestimate the savings from repatriation.

Platform: The abstraction layer — containerization, orchestration, data services, security, compliance tooling. This is where the cloud conversation gets complicated. Public cloud platforms bundle an enormous amount of platform-layer value into pricing that looks expensive at the compute level but is difficult to replicate cheaply in-house. Organizations repatriating compute workloads must assess what platform services they're also repatriating — and which they plan to rebuild.

Application: The workload layer — what runs on the infrastructure, at what utilization rate, with what variability profile. This is where the economic case for repatriation either holds or falls apart. Predictable, high-volume workloads with stable demand profiles are candidates for owned infrastructure. Variable, burst, or experimental workloads requiring scale-on-demand are not.

GEICO's workload profile — 200,000+ compute cores running 30,000+ container and VM instances at a predictable insurance company utilization pattern — suits owned infrastructure.⁵ The demand is stable. The workloads are largely defined. The utilization is high enough to justify the fixed asset investment.

Mani's risk asymmetry argument on stage mattered for exactly this reason. He distinguished the failure modes of each model:

• CapEx bad forecasting: Wasted capital. A one-time, durable hit. You built more than you needed and take the write-down.

• OpEx bad forecasting: Wasted money in real time. A continuous bleed. If your AI workloads scale faster than anticipated and your OpEx model wasn't priced for that scale, you pay overage costs against a committed run rate you can't easily renegotiate.

The OpEx model's flexibility is real, but it's priced. The premium for that flexibility is embedded in the service fee. Organizations that genuinely need flexibility — because their workload profile is variable or their demand trajectory is uncertain — pay for real value. Organizations with stable, predictable demand pay for optionality they don't need.

When outsourcing genuinely wins

Read more

On April 30, 2026, ENTSO-E — the body that coordinates Europe's transmission system operators — published a warning that reframed a crisis most analysts were already tracking. The concern was no longer that data centers couldn't get power. The new concern was that data centers might consume so much of Europe's available generation headroom that grid operators would be forced to curtail renewable energy penetration across the continent.²⁷

That shift in framing captures how far this story has moved. We passed "European data centers face a grid problem" roughly two years ago. We are now at "European data centers have built workarounds so substantial that the workarounds themselves are reshaping the grid."

The numbers behind the original crisis are stark. FLAP-D — the Frankfurt, London, Amsterdam, Paris, and Dublin cluster that accounts for the majority of European data center capacity — runs grid connection queues averaging 7 to 10 years against a data center construction window of 18 to 24 months.¹ Ireland has €5.8 billion in stranded data center investment: fully permitted projects on purchased land that cannot reach commercial operation because the grid cannot connect them.⁶ The Netherlands enacted a national decree in January 2024 prohibiting hyperscale data centers across the entire country except two remote northern municipalities.³² The UK's contracted demand queue tripled from 41 gigawatts (GW) to 125 GW in seven months before emergency reform.³¹

In February 2026, AWS's head of energy markets and regulation for EMEA went on record: "There's a misalignment. We want to expand and grow within two years. These delays are challenging our growth aspirations."⁸ The world's largest cloud operator had committed €15.7 billion to Aragon, Spain — a secondary market selected precisely because it had available grid capacity. The FLAP-D queue was not an acceptable timeline.

European operators did not wait for policy reform. They engineered around the constraint. They did not do it the same way. They fragmented into three structurally distinct deployment patterns — each defined by which regulator's permit authority it triggers and bypasses, each executing in a different geographic zone, each creating a different competitive moat.

Understanding those three patterns is the analytical work this article does. The grid crisis is context. The response patterns are the analysis.

The Framework: Three Patterns, Three Regulatory Wedges

The European grid constraint problem has a common cause — infrastructure that cannot absorb the pace of AI-driven demand growth — but it has produced three structurally different operator responses. Each response exploits a gap between the regulator whose authority it triggers and the regulator whose authority it bypasses.

Pattern 1: Behind-the-Meter (BTM) Combustion. Burn fuel on-site to generate electricity. Bypasses the grid connection queue because no new grid capacity is needed for the behind-the-meter portion. Triggers air permitting requirements, EU Emissions Trading System (ETS) carbon pricing on combustion emissions, and member-state environmental frameworks. Dominant today in Ireland. Emerging in Germany via utility partnerships that supplement existing grid connections.

Pattern 2: Co-located Renewable Self-Consumption. Site the data center adjacent to dedicated wind, solar, hydropower, or biogas generation. Structure supply under member-state self-consumption frameworks. Bypasses both air permitting and the grid connection queue for the renewable portion. Falls under renewable energy policy rather than fossil generation policy. Dominant in Iberia and the Nordics, where geography permits hyperscale co-location at renewable sites.

Pattern 3: Data Center as Grid Asset. Stay grid-connected but use the facility's uninterruptible power supply (UPS) battery fleet to provide ancillary services — fast frequency response, voltage support, capacity — to the transmission system operator (TSO) under the TSO's grid services market. Triggers neither air permitting nor new generation connections. Transforms the data center from net load problem to net grid contributor. Operational at scale in Ireland. Emerging in the UK.

These patterns are not alternatives that operators choose between on preference. They are products of the regulatory environment that a given geography offers. An operator building in Aragon, Spain cannot choose Pattern 1 without the air permits that Aragon's planning regime would grant. An operator building in Dublin cannot deploy Pattern 2 at hyperscale without geographic access to co-located renewable generation. The permit determines the pattern. The geography determines the permit.

The three patterns are not converging. They are hardening as geographic competitive positions, locking capital into specific architectures faster than any European-level policy reform can respond.

The Scale of the Bet

Before examining each pattern, the market context establishes what is at stake. European data centers consumed approximately 96 terawatt-hours (TWh) of electricity in 2024 — about 3% of EU electricity demand.¹⁸ Ember Energy projects this rising to 168 TWh by 2030, a 75% increase in six years.³ Ember Energy's June 2025 analysis extends the trajectory to 150% growth by 2035, reaching 5.7% of total EU electricity consumption.³ For context: data center demand growth between 2024 and 2030 (72 TWh additional) exceeds the projected growth from electric vehicles (67 TWh).³

The European Data Centre Association (EUDCA) has quantified the investment required to meet this demand: €176 billion cumulatively from 2026 through 2031, running approximately €25–26 billion per year for large AI-dedicated scale colocation facilities alone.² The compound annual growth rate (CAGR) for European IT power demand through 2031 sits at 17%.² Grid readiness — not capital availability — is the binding constraint.

The capital is not staying in FLAP-D to wait for the grid. It is redirecting to secondary markets where Pattern 2 executes at hyperscale. Ember Energy projects secondary European markets — Nordics, Iberia, Central and Eastern Europe — will grow 110% between 2024 and 2030 versus 55% for the traditional hubs.³ By 2035, half of European data center capacity will sit outside those hubs.³ Spain absorbed AWS's €15.7 billion Aragon commitment plus Microsoft's $7.16 billion Aragon investment and $2.1 billion Madrid commitment.¹⁰ The Nordics attracted $15 billion in hyperscaler commitments in 2025 alone — Microsoft, CoreWeave, and Brookfield — anchored by Stargate Norway, the OpenAI + Nscale + Aker partnership targeting 230 megawatts (MW) on 100% Norwegian hydropower.¹²

This is not geographic diversification. It is capital flight to places where one of the three patterns can execute.

Pattern 1: The Irish BTM Buildout

Ireland's grid constraint was not abstract for operators. The Commission for Regulation of Utilities (CRU) imposed a de facto moratorium on new data center connections to Dublin's grid beginning in 2021, triggered by data centers reaching 22% of Ireland's metered electricity demand — a share projected to reach 30% by 2030.⁶ CRU's December 2025 decision formally ended the moratorium but replaced it with a requirement that every new connection above 10 megavolt-amperes (MVA) must provide matching dispatchable generation or storage — on-site or proximate — that participates in the Irish Single Electricity Market (SEM).⁶, ⁷ The moratorium became a mandate. The mandate is Pattern 1.

Three operators are executing this pattern in Ireland today, collectively deploying over 600 MW of operator-provided generation:

EdgeConneX Grange Castle (Lucan, County Dublin) — approximately 100 MW gas generation, operational 2023. South Dublin County Council approved three gas-powered generation plants in 2022, with 20 to 21 generator units per plant per planning records. The EdgeConneX Dublin data sheet records 170+ MW of power banked across the campus, supporting two existing and up to five planned data centers. A November 2024 investigation documented 28 emergency generators that have emitted approximately 130,000 tonnes of CO₂ since 2017. Gas Networks Ireland's "Gas to Grange" infrastructure project supplies the fuel. EdgeConneX secured these permits before Irish regulatory tightening. Based on the planning record, the site has not subsequently obtained permission for an additional data center building that would put matched IT load at the same scale as the gas generation — the power infrastructure is substantially in place, but the proportional data center capacity is not fully built out. An Coimisiún Pleanála's August 2025 approval for future generators at the site imposed a condition requiring those generators to run on renewable fuel — signaling that the air permit pathway used in 2022 is closing for future natural gas.

Pure DC Dublin (DUB01, Ballycoolin) — 110 MW Wärtsilä microgrid, announced March 2026. Pure Data Centres Group partnered with AVK-SEG on a €1 billion project building three energy centres of up to 30 MW each, plus 20 MW of battery energy storage.⁴ The system operates in full island mode — it is the primary power source for the campus, not a backup layer. Primary fuel is natural gas; hydrotreated vegetable oil (HVO) provides seamless switching backup; biomethane has been trialed; the system is hydrogen-blend ready with minor modifications. Energy Centres 1 and 2 are operational by end of 2026; EC3 follows.⁴

Pure DC's marketing describes this as "Europe's first microgrid data center." The framing is accurate in one specific sense: it is the first project explicitly branded a microgrid at this scale, and the first designed from inception for full grid independence as the primary operating mode. EdgeConneX Grange Castle and Echelon DUB10 predate it for primary on-site gas generation in Ireland. The "microgrid" term is marketing differentiation, not architectural novelty. The practical consequence of this mislabeling: the actual volume of Pattern 1 deployment in Ireland is substantially larger than the headlines suggest, because ECX has been executing this architecture since 2023 without the microgrid label.

Echelon DUB10 (Clondalkin, Dublin) — 120 MW on-site energy center. Echelon Data Centres' Clondalkin facility includes a 110 kV grid substation and an on-site energy center with planning permission for 100 MVA of gas-fired generation. Global Energy Monitor's power station database records the capacity at 120 MW. Echelon raised €1.7 billion from Morgan Stanley in 2025 to fund Irish expansion, with DUB10 and DUB20 as primary targets.

The regulatory wedge for all three is identical: air permits granted before or during Irish tightening, EU ETS exposure on combustion, and a CRU connection framework that codifies on-site generation as the only viable path. What differentiates Pattern 1 from the alternatives is that it executes immediately wherever air permits are available — but faces a deteriorating carbon economics trajectory as EU ETS prices rise toward €100 per tonne of CO₂ by 2027 and analysts forecast €126 per tonne by 2030.³⁴

Read more

This is Part 2 of the Ratepayer Pledge Series. Part 1, "The $444-a-Year Lie," examined how data center operators are shifting grid upgrade costs onto residential ratepayers. Part 2 asks the harder question: can the infrastructure deliver the power these deals actually promise?

In roughly ninety days — January through mid-April 2026 — hyperscalers and their power partners announced more than 19.2 gigawatts of named behind-the-meter (BTM) natural gas deals. Then the announcements kept coming. By early May, the tally of named, publicly confirmed gas power commitments had crossed 35 gigawatts: SoftBank's 9.2 gigawatt Stargate campus in Ohio, Microsoft and Chevron's 2.5 gigawatt West Texas plant, Oracle's dual-technology strategy spanning 5.25 gigawatts across Bloom Energy fuel cells and VoltaGrid engines, Pacifico Energy's 7.65 gigawatt West Texas private grid, Meta's 7.5 gigawatt Louisiana buildout — all announced within a single fiscal quarter.¹, ¹⁶, ¹⁸, ¹⁹

This is not a trend. It is a category decision. The data center industry has collectively concluded that the electric grid cannot deliver power fast enough, reliably enough, or cheaply enough to support the AI infrastructure buildout. So it is building its own generation — at a scale that, if fully executed, adds the equivalent of 13 to 17 percent of the entire U.S. electric sector's gas consumption in one cycle.²⁸

The question is not whether gas is the right fuel. It is whether the infrastructure to deliver that gas — pipelines, engines, turbines — can meet the timeline the deals imply.

The answer is decisively geographic. Texas deals will execute on time. Ohio has the pipeline and regulatory infrastructure to proceed. Northern Virginia, home to the world's densest data center corridor, faces a compound constraint — pipeline capacity, air quality permits, and federal co-location costs — that no data center's planning horizon can resolve.

Geography is destiny. And the market is already pricing the difference.

The Announcement Wave

The speed of capital commitment is without precedent in modern U.S. energy history. Global Energy Monitor reports the United States nearly tripled its gas-fired capacity in development in 2025, reaching approximately 252 gigawatts total — with more than one-third of that, roughly 84 gigawatts, slated to directly power data centers on-site.²⁰ Texas alone accounts for 80.6 gigawatts of planned buildout, with nearly half (40 gigawatts) intended to serve data centers directly. The full buildout, if completed, represents capital investment exceeding $416 billion nationally.

The specific deals driving this wave reveal how serious the commitment is:

SoftBank Stargate Ohio — 9.2 GW, $33.3 billion. At the former Portsmouth Gaseous Diffusion Plant in Piketon, Ohio, SoftBank's SB Energy is developing a 10 gigawatt data center campus backed by $33.3 billion in investment, including Japanese government-sourced capital. The generation fleet: 9.2 gigawatts of natural gas. AEP Ohio committed $4.2 billion in transmission infrastructure to support the campus. Initial power target: 2028.¹⁸, ¹⁹

Pacifico Energy GW Ranch — 7.65 GW, West Texas. Pacifico Energy secured the largest single gas-fired air permit in U.S. history from the Texas Commission on Environmental Quality (TCEQ) in January 2026 for a 7.65 gigawatt gas-fired, 1.8 gigawatt battery storage, and 750 megawatt solar private-grid campus in Pecos County, West Texas. Phase 1 delivers 1 gigawatt in 2027; full buildout exceeds 5 gigawatts by 2031. Pacifico is explicitly avoiding ERCOT interconnection to bypass the interconnection queue entirely — a private utility selling power directly to data center tenants.⁴³

Meta Hyperion / Entergy Louisiana — 7.5 GW, $27 billion. Meta tripled its initial plans for its Hyperion AI campus in Richland Parish, Louisiana — now 10 gas-fired plants delivering 7.5 gigawatts under a 15-year contract with Entergy.⁴⁴, ⁵¹ Louisiana regulators fast-tracked approval in April 2026. Meta is also funding up to 2.5 gigawatts of renewables and battery storage as part of the deal. Unlike the Texas and Ohio BTM models, this is utility-developed generation: Entergy owns the plants, and data center power costs flow through utility cost-of-service accounting — with ratepayer backstop concerns already flagged.

Microsoft / Chevron / Engine No. 1 — 2.5 GW, $7 billion. Microsoft and Chevron entered an exclusivity agreement for a 2.5 gigawatt plant in Pecos, West Texas, with an option to expand to 5 gigawatts. First power target: 2027. The deal's economic foundation: Waha Hub gas prices were negative 84 of 97 days in 2026 through mid-April. At a 7,000 BTU/kWh heat rate, Waha at negative $2.50 relative to Henry Hub cuts the fuel cost from approximately $27.65 per megawatt-hour to approximately $10.15 per megawatt-hour — a 63 percent reduction.¹⁶, ²³

Oracle — 5.25 GW across two technology strategies. Oracle's power procurement strategy is itself a geographic hedge. In Texas, Oracle contracted VoltaGrid for 2.3 gigawatts of modular reciprocating engine capacity deployed across multiple sites — the largest single reciprocating engine deployment in U.S. data center history, using INNIO Jenbacher engines in VoltaGrid's 25 megawatt QPac platform modules with firm gas supply from Energy Transfer.⁶ Nationally, Oracle contracted Bloom Energy for up to 2.8 gigawatts of solid oxide fuel cells (SOFCs) — the largest fuel cell order in U.S. history — with Bloom named as the sole power provider for Oracle's 2.45 gigawatt grid-independent AI campus in New Mexico, in addition to the broader master services agreement covering multiple U.S. sites.⁴, ⁴⁵

Wärtsilä 34SG Ohio — 412 MW, 40 engines. In April 2026, Wärtsilä received an order for 40 of its 34SG power-optimized gas engines (412 megawatts nameplate) for a hyperscale data center in Ohio — the first-ever data center deployment of the 34SG model, and Wärtsilä's fourth U.S. data center order. Commercial operation target: early 2028, approximately 20 months from order to power.¹

As we documented in our earlier analysis of the $42 billion onsite generation boom, the gas turbine supply chain was already under severe pressure before this latest announcement wave — and the reciprocating engine manufacturers filling the gap faced their own deployment constraints.

The aggregate demand math is straightforward. The broader BTM pipeline — 56 gigawatts of behind-the-meter capacity, 75 percent gas — translates at a 7,200 BTU/kWh heat rate and a 70 to 80 percent operational profile to **4.9 to 5.6 billion cubic feet per day of new natural gas demand.**²⁸ For context: U.S. total natural gas consumption set a record at 92 billion cubic feet per day in 2025. The electric power sector consumes 35 to 37 billion cubic feet per day. The BTM pipeline, if fully materialized, adds the equivalent of 13 to 17 percent of current electric sector gas consumption in a single construction cycle.

That demand lands on a pipeline system already running at records — and expanding not primarily to serve domestic data centers, but to supply liquefied natural gas (LNG) export terminals on the Gulf Coast. U.S. LNG export capacity is expanding from 11.4 billion cubic feet per day to 24.4 billion cubic feet per day by 2028 — competing for the same firm pipeline transport on the same interstate infrastructure, with no federal policy mechanism coordinating between them.³⁹

Geography Is Destiny

The BTM gas market is not a single national market. It is three distinct markets with radically different infrastructure realities operating under the same branding. Understanding which deals execute requires understanding which geography they sit in.

Texas: Enabled. Deep pipeline capacity, near-zero fuel cost at the Waha Hub, favorable permitting under TCEQ, and ERCOT's exemption from FERC co-location rules combine to make West Texas the lowest-barrier BTM gas geography in the country. The Microsoft/Chevron deal and Pacifico's GW Ranch are both anchored here. VoltaGrid's Oracle deployment runs on firm gas contracted through Energy Transfer's pipeline and storage network. The structural cost advantage is not marginal — at scale, a 2.5 gigawatt plant running at 85 percent capacity factor consumes approximately 330 million cubic feet per day, and the Waha basis discount on that volume represents hundreds of millions of dollars annually versus a comparable project in a Henry Hub-priced market.

Ohio: Developing. SoftBank's Stargate campus represents an attempt to develop a new gas infrastructure corridor at hyperscale, benefiting from Appalachian Basin gas access via Marcellus and Utica shale, Ohio House Bill 15's 90-day approval process, and $4.2 billion in AEP Ohio transmission investment. Ohio approved 2,000 megawatts of BTM power in 2025 alone — more than the prior 20 years combined — and received applications for an additional 2,755 megawatts of traditional gas facilities. At 9.2 gigawatts, the Stargate campus alone consumes approximately 1.5 billion cubic feet per day of natural gas, roughly 5 percent of current total U.S. production from a single site in rural Ohio.³⁷, ³⁸

Louisiana: Utility-Developed. Meta's Hyperion campus adds a fourth geography — one with abundant Henry Hub-corridor gas supply, fast regulatory approval, and a different ownership structure. Entergy owns the plants; Meta secures the power through a long-term power purchase agreement. This is not BTM in the strict sense — it is utility-developed generation purpose-built for a single data center anchor customer, with costs flowing through the regulated utility model rather than purely off the grid. The outcome (large-scale gas-powered AI campus) looks similar. The economics (ratepayer exposure vs. merchant risk) are fundamentally different.

Northern Virginia: Constrained. Loudoun County alone has more data center floor space than any other county globally. But the pipelines serving that corridor — Transco principal among them — are running near capacity on firm transport subscriptions. A data center's proximity to the Transco mainline does not equal access to firm capacity. That distinction is the central technical fact of the Northern Virginia constraint.

The pipeline expansion targeting this corridor is Williams' Transco Power Express project. In Q1 2026 earnings, Williams confirmed the project had been rescoped down from 950 million cubic feet per day to 750 million cubic feet per day — a 200 million cubic feet per day reduction attributed to "customer-driven scope adjustments."⁴⁶ The constraint thesis is not weakening. It is hardening: even Williams, the pipeline operator that stands to profit most from Northern Virginia gas demand, scaled back its rescue pipeline.

Power Express remains on track for 2030 in-service — pending a FERC application that has not yet been filed. Even at the original 950 million cubic feet per day, it was insufficient to support the density of data center gas demand concentrated in Loudoun, Fairfax, Prince William, and Arlington counties. At 750 million cubic feet per day, it is less so.

Data centers are being built now — 18 to 24 months from groundbreaking to power. The pipeline arrives in 2030. That 4-to-5-year gap is the operating condition for every Northern Virginia BTM gas decision being made today.

The Pipeline Math That Doesn't Close

Read more

In September 2025, we published "How AI Data Centers Are Driving a $42 Billion Onsite Generation Boom" — an analysis that named the opportunity, mapped the technology, and argued that onsite power would become a structural feature of hyperscale infrastructure, not a backup workaround.

Nine months later, the buildout has accelerated beyond any published forecast.

The $42 billion figure captured generator-equipment market projections for one technology bucket in one segment of the market. What no forecast captured was the simultaneous activation of all five technology buckets at once: the nuclear procurement pipeline, the fuel cell mega-contracts, the co-located renewable energy parks, the private equity infrastructure vehicles formalizing this as an asset class. Add those in, and the onsite generation buildout represents roughly 60 GW of committed capacity — approximately $150 billion in power infrastructure investment at current installed costs — with no sign of deceleration.

Here is what changed in nine months.

GE Vernova's gas turbine backlog stands at 100 gigawatts — more than three times the total US electricity generating capacity added in all of 2024.¹ One supplier, one technology, backlogged past 2028. The order pace is still accelerating. Then add the Oracle/Bloom Energy fuel cell deal at 2.8 GW — the largest single onsite power supply agreement in market history.² Add 9.8 GW in nuclear procurement across 13 hyperscaler deals.³¹ Add Wärtsilä booking 790 MW of reciprocating engine (RICE) capacity in a single Texas order.⁴ Add the US battery energy storage system (BESS) market setting consecutive quarterly records at 9.7 GWh (Q1 2026) and 10 GWh (Q2 2026).³³

Every technology bucket is deploying at gigawatt scale simultaneously. That has never happened before.

The 35 GW Gap That Activated Five Markets at Once

The market-creation event is a single number: a 35 GW gap between announced data center capacity and available grid power supply through 2030.⁶ That gap — equal to six times New York City's annual energy consumption — explains why the buildout spans five technologies rather than one. No single technology closes a 35 GW shortfall on the timelines hyperscalers require. The grid cannot solve it; interconnection queues run three to seven years in most major markets. So the industry activated every available alternative simultaneously.

The pace of attitude change tells the story. In January 2026, 38% of hyperscaler and colocation decision-makers expected onsite power to be their primary power source by 2030 — up from 13% just seven months earlier. One-third now expect to operate data centers running entirely on onsite generation.⁶ McKinsey projects 25-33% of all incremental data center demand through 2030 will come from behind-the-meter sources, as cited by Bloom Energy.¹⁹

Those are not aspirational numbers from a vendor survey. They are procurement commitments backed by signed contracts, construction permits, and capital already deployed.

The four major hyperscalers — Amazon, Google, Meta, and Microsoft — will collectively spend $725 billion in capital expenditures in 2026, up 77% year-over-year.⁸ More than 60% of announced AI data center project capital goes to power infrastructure, not chips or compute. The IEA projects 15-27 GW of onsite natural gas alone powering US data centers by 2030 — from gas turbines only, before accounting for reciprocating engines, fuel cells, or nuclear.⁷

The Parallel Grid Is Already Under Construction

The thesis we advanced in September 2025 — that hyperscalers were building toward private grid independence — has crossed from concept to construction reality. Our comprehensive analysis of why data centers abandoned grid interconnection documented how PJM capacity costs and 3-7 year interconnection queues made BTM generation the rational economic choice; what has changed since is the scale of execution.

The Nscale Monarch Compute Campus in Mason County, West Virginia, is designed for full island operation independent of the local grid. It draws 2 GW from Caterpillar G3500 series natural gas generators, serves Microsoft as anchor tenant with 1.35 GW of committed capacity, and scales beyond 8 GW on a 2,250-acre state-certified AI microgrid.²⁵ Meta's Hyperion campus in Richland Parish, Louisiana — 8,000 acres, 7+ GW of dedicated onsite generation across 10 planned power plants, $200 billion projected total investment — is the largest announced AI infrastructure project in history, with no grid dependency in the design.¹⁶

Oracle's 2.8 GW fuel cell pipeline with Bloom Energy covers US locations where grid interconnection timelines made grid dependency operationally unworkable.² xAI's Colossus supercomputer complex in Memphis targets 2 GW of onsite capacity with 1.1 GW fully operating by Q2 2027.¹¹

These are not backup strategies. They are primary power strategies — a parallel private power infrastructure built from scratch over 18 months.

The strategic picture is unambiguous: the data center industry has assembled the largest private power infrastructure program in US history, with every technology bucket now deploying at GW scale simultaneously. This is no longer a supply chain workaround. It is a structural shift in how power infrastructure works in America.

Five Technology Buckets, Five Different Markets

The market's defining characteristic is differentiation — not just in technology, but in deployment speed, power density, emissions profile, and regulatory exposure. The research behind this analysis mapped all five in detail. Here is the frame.

Gas turbines are the dominant platform — GE Vernova's 100 GW backlog makes that concrete — but slots are sold out through 2026-2027, 90%+ booked through 2028.¹, ¹² Developers who cannot get a near-term slot are not waiting. They are reaching for the next available alternative.

Reciprocating internal combustion engines (RICE) are filling that gap. Wärtsilä, INNIO, and Caterpillar have collectively absorbed GW-scale orders that would have gone to turbines if slots were available. RICE engines offer 12-18 month lead times versus 5+ years for heavy-frame turbines, synchronize in 30 seconds, and commission unit-by-unit — qualities that match AI workload variability in ways large turbines cannot replicate.

Fuel cells and solid oxide fuel cells (SOFC) have become the solution for the most constrained market segment: urban campuses and non-attainment zones where gas combustion faces strict air permit review. Bloom Energy's 2.8 GW Oracle deal is the headline, but the deeper story is regulatory: approximately 46% of planned US data center sites sit in ozone non-attainment zones where SOFC is the only onsite thermal technology that deploys without triggering Best Available Control Technology (BACT) or Lowest Achievable Emission Rate (LAER) review.⁴⁷

Nuclear has produced 9.8 GW in procurement commitments across 13 deals — every major hyperscaler has signed at least one.³¹ But the timeline is a 2030-2035 story, not a 2027 story. The one exception is Three Mile Island Unit 1 (835 MW restart, targeted for 2027). TerraPower broke ground on the first Gen IV advanced reactor in the US in 40 years on April 23, 2026 — targeting 2030.²⁸ Full commercial SMR fleets are realistically 2032-2035. Nuclear procurement today is forward contracting on carbon-free baseload power, not near-term capacity.

Solar and BESS hybrids serve as the carbon strategy layer atop firm power — not 24/7 baseload on their own, but a cost and carbon optimizer that changes the economics meaningfully when co-located. The C&I BESS segment is growing 55% year-over-year, and the Section 48E investment tax credit (ITC) for solar and wind construction that begins after approximately July 4, 2026 terminates — a deadline driving a construction start rush right now.⁵, ³⁸

The xAI Memphis Story: What Happens When You Move Faster Than the Permitting System

No case study captures the speed and the hard limits of this buildout more concisely than xAI's Memphis Colossus deployment.

xAI deployed approximately 35 Solaris Energy Infrastructure turbines — roughly 400 MW total — at its Colossus AI supercomputer facility before Memphis grid connections were secured. TechCrunch and Data Center Dynamics reported that xAI operated nearly 50 gas turbines in violation of Shelby County permit limits in a non-attainment zone for ozone.¹¹, ⁴⁸ Phase I grid connection (150 MW from MLGW/TVA) came May 1, 2025 — months after turbines were already running.

This was not a paperwork problem. Non-attainment zone constraints are a hard regulatory ceiling on gas turbine deployment in urban markets. In non-attainment zones, LAER requirements can force NOx emissions below 5 ppm — a standard that makes conventional gas turbine deployment at primary generation utilization rates effectively prohibited without extraordinary mitigation.

xAI's workaround was operationally creative: it repurposed a Mississippi power plant across the state border and drew transmission back to the Memphis site. Colossus 2 is physically located in a different state to escape the ozone non-attainment zone that made Colossus 1 a regulatory problem. The company's total target remains 2 GW, with 1.1 GW fully operating by Q2 2027.¹¹

The cross-border siting workaround will become a template for other urban market buildouts. But it is expensive, time-consuming, and not available everywhere. The lesson for developers is clear: air quality attainment status must be the first variable in siting decisions, not an afterthought. Technology selection follows from the permitting map, not the other way around.

The full analysis continues below for paid subscribers: per-technology deep dives with unit economics and named project details, the non-attainment zone dynamic that has built a structural competitive moat for Bloom Energy, the complete nuclear timeline reality check, four regulatory developments that are actively repricing each bucket right now, and the strategic implications by technology and player.

Five Technology Buckets: A Definitive Market Map

Read more

In October 2025, Google signed a power purchase agreement no one had done before: electricity from a natural gas plant with carbon capture and storage as a contractual requirement, not an optional add-on. The plant doesn’t exist yet. Final investment decision (FID) is expected in Q2 2026. Commercial operations are projected for 2029-2030. But the Broadwing Energy deal in Decatur, Illinois is already rewriting the commercial logic of an entire industry.

Fourteen months after that announcement, five named organizations have committed to 6+ gigawatts of natural gas generation with integrated carbon capture and storage specifically for the data center market.¹ A sector that didn’t exist as a market category in September 2025 now has a pipeline larger than most US states’ total utility-scale solar queues. ExxonMobil — which had never built a power plant for external customers — is in advanced development on a 1.5+ GW natural gas facility with more than 90% carbon capture, exclusively for data centers.² Chevron, Engine No. 1, and GE Vernova formed a joint venture targeting up to 4 GW of behind-the-meter (BTM) power for the same customers, designed from day one with the option to integrate carbon capture and storage (CCS) at greater than 90% capture.³

This is what a market category creation event looks like from the inside.

Subscribe now

The technology existed. The policy support — Section 45Q at $85 per ton for geological storage — existed. The industrial infrastructure: Mitsubishi Power turbines, Mitsubishi Heavy Industries’ (MHI) capture technology, ADM’s Class VI injection wells. What didn’t exist was a commercial template: a way to finance a gas power plant with CCS as a bundled product rather than a compliance obligation, anchored by a creditworthy corporate buyer who treated the carbon performance as integral to the purchase. Google’s PPA created that template. Everything since has been the market responding to it.

This article is Part 2 of Avanza’s CCS series. The $160 Billion Last Mile covers the full technology landscape — post-combustion amine capture, the CO2 concentration challenge for gas turbines, cryogenic and membrane alternatives, and the EPA Class VI permitting bottleneck that determines how fast storage can scale. Here, the focus is the market and commercial thesis: why this demand is structural, how the OEMs are positioned, what the commercial model means for project finance, and what comes next.

The Broadwing Deal: A Deconstruction

Most announcements in the energy transition are directional commitments with fuzzy implementation. The Broadwing deal is different — it is a contract with specific commercial, technical, and operational requirements that define what a gas + CCS data center power project must actually deliver.

The parameters matter because they establish the standard. Broadwing Energy, developed by Low Carbon Infrastructure and financed by I Squared Capital (a global infrastructure private equity manager with more than $35 billion in assets under management), is building a 400+ MW natural gas cogeneration facility at ADM’s grain processing campus in Decatur, Illinois.⁴, ⁵ Google purchases the majority of its electricity output through a long-term PPA. The technical requirements embedded in that PPA set the floor for the entire emerging market:

Generation technology: Mitsubishi Power M501JAC gas turbine — a proven H-class machine operating at high efficiency, selected for compatibility with MHI’s capture system.⁴

Capture technology: MHI KM CDR Process using the KS-21 advanced solvent, deployed in 18 commercial plants globally. The contractual performance target: greater than 90% CO2 removal.⁶, ⁴

Storage: ADM’s EPA Class VI approved injection wells at the Decatur site — the longest-operating Class VI wells in the nation, with over a decade of injection history from ADM’s own ethanol plant CO2 operations.⁷ More than one mile underground, into a saline formation with the best-documented performance history in the country.

The accounting mechanism: Google worked with Broadwing and external validators to develop a new CCS-specific Energy Attribute Certificate (EAC) — a tradeable instrument that credits the captured CO2 against Google’s Scope 2 greenhouse gas reporting. The Carbon Capture Coalition called the announcement “an important milestone in the deployment of natural gas plus CCS” and specifically highlighted the storage permanence credentials of the ADM wells.⁷

The FID timeline: Q2 2026. Commercial operations: 2029-2030.

The pricing structure: Google did not disclose the PPA price per megawatt-hour. Kate Brandt, Google’s Chief Sustainability Officer, stated: “We believe that CCS can be competitive at scale with other generation technologies, and certainly other low carbon or zero carbon generation technologies.”⁸ The project economics are inferable: a conventional firm power PPA for a new natural gas combined cycle (NGCC) plant prices at $60-80/MWh in current markets. Natural gas + CCS adds approximately $15-25/MWh to generation cost at DOE/NETL baseline capture economics of $61 per tonne.⁹ The 45Q geological storage credit at $85 per tonne — applied to roughly 90% of 0.35-0.40 tonnes CO2 per MWh of NGCC generation — returns approximately $27-31/MWh to the project. Google is paying a modest power price premium above market gas generation, with 45Q roughly covering the incremental CCS cost.¹⁰

The commercial innovation is the structure, not the price. By bundling CCS costs into the power price rather than treating them as a separate carbon product, Broadwing finances like a solar or wind project — a long-term PPA anchoring project finance debt, with a creditworthy corporate offtake counterparty. Lenders underwrite to the contract. Equity earns upside from 45Q credits and plant performance. The “CCS risk” disappears into the power generation business, which project finance lenders have underwritten for decades.⁸

This structure changes who can invest. It opens CCS projects to the full universe of infrastructure debt capital — pension funds, insurance companies, infrastructure debt funds — that cannot underwrite a standalone “carbon removal” business but will absolutely underwrite a long-term natural gas power supply contract to Google.

The Structural Demand: Why This Isn’t a One-Off

The Broadwing deal is significant not because Google is unusual but because Google is representative. Every major hyperscaler operating at the AI frontier faces the same structural tension — and none can resolve it without CCS or equivalent technology at scale.

The tension is straightforward: AI data centers require 24/7, dispatchable power at the megawatt scale. Intermittent renewables, however cheap, cannot deliver that without storage systems that don’t yet exist at data center scale. Natural gas is the only currently available technology that is fast to interconnect, available everywhere, and reliably dispatchable. But all four major hyperscalers have made public commitments that put gas generation on a direct collision course with their Scope 2 reporting obligations.

Google’s 24/7 carbon-free energy (CFE) target for 2030 requires every megawatt-hour of grid power matched to a carbon-free generation source at the same hour it is consumed. Gas generation, even from a new and efficient plant, fails this test without CCS.⁴ Microsoft’s commitment to carbon negativity by 2030 and near-zero Scope 1 and Scope 2 emissions makes on-site gas combustion a direct liability — yet Microsoft VP of Energy Bobby Hollis stated in March 2025 that Microsoft is “open to using natural gas with carbon capture technology to power AI data centers” if it is “commercially viable and cost competitive.”¹¹ Amazon’s total emissions reached 68.25 million metric tonnes in 2024 — a 6% increase year-over-year and 33% above 2019 levels — while the company simultaneously commits to net zero by 2040.¹² Meta is deploying H-class natural gas turbines at its Hyperion multi-gigawatt AI data center in rural Louisiana right now, with a Net Zero 2030 commitment those turbines directly undermine.¹³

As we documented in our analysis of hyperscaler climate disclosure gaps, the structural problem is not bad faith — it is a disclosure framework that preserves low-carbon public narratives while the physical power system becomes materially more fossil-dependent. CCS-enabled PPAs are the mechanism that closes that gap.

The scale of the gas buildout makes resolution unavoidable. EPRI projects 6.6 to 13.7 GW of incremental natural gas capacity additions per year from 2025 through 2030, driven by data center demand.¹³ BCG quantifies an 80 GW gap between projected round-the-clock data center power demand and available generating capacity in the US by 2030.¹⁴ Bain projects that more than 25% of new data center facilities above 500 MW will have BTM power by 2030, up from 1% today.¹³

The sustainability math is equally stark. Carbon Direct’s 2025 analysis found that if Google, Meta, Amazon, and Microsoft collectively achieved their 2030 clean power commitments using only natural gas combined cycle plants with carbon capture and storage, the scenario would require $70-80 billion in capital expenditure and over 105 million metric tonnes per annum (Mtpa) of CO2 sequestration.¹⁵ That is a ceiling scenario. The real market is a fraction of it — hyperscalers will use a mix of nuclear, renewables, storage, and gas + CCS to meet their targets. But even 20-30% of the projected hyperscaler gas portfolio procured through CCS-enabled PPAs represents 20-30 Mtpa of CO2 sequestration and $14-24 billion in capital deployment by 2030.¹⁵

To put that in context: total global CCS capacity was 51 Mtpa at year-end 2024, accumulated over three decades of commercial deployment.¹⁶ The data center sector adds a comparable quantity in five years if the PPA template scales.

The full analysis continues below for paid subscribers, including detailed OEM competitive positioning, the Northern Lights commercial model and its US implications, the behind-the-meter regulatory breakthrough, 45Q political durability, market sizing, and strategic investment implications with comprehensive citations.

The OEM Race: Three Strategies for the Same Market

Read more

Carbon dioxide emissions are not a monolith. An ethanol fermentation plant exhales a stream that is more than 95% pure CO2, practically demanding to be captured. A natural gas combined cycle (NGCC) power plant produces exhaust at roughly 3–4% CO2 — a fraction of the concentration, requiring far more energy and equipment to process the same amount of carbon. The open atmosphere sits at 0.04%, a dilution so extreme that capturing a tonne of CO2 from it requires processing an enormous volume of air.

That gradient — from 95% to 0.04% — is the most important variable in all of carbon capture economics. It determines cost. It determines technology choice. It determines which projects get built and which stay on whiteboards. It is why the global carbon capture and storage (CCS) industry is bifurcating sharply between approaches that work at reasonable cost today and approaches that remain expensive but strategically necessary for the long term.

The scale of what is at stake is large. The International Energy Agency’s (IEA) Net Zero Scenario requires roughly 1.2 billion metric tonnes of CO2 captured and stored per year by 2030, rising to 6.2 billion tonnes by 2050.¹ Current operating capacity globally stands at 51 million metric tonnes per year — across 50 commercial projects.¹³ The gap between where we are and where we need to be is roughly 120x over 25 years. That is not a rounding error. That is an industry that needs to be built.

The IEA estimates that closing that gap requires $160 billion in cumulative investment through 2030 alone — a 10x increase from the prior decade.¹ The US Energy Information Administration (EIA), in its 2025 Annual Energy Outlook, projects US CO2 capture peaking at just under 71 million metric tonnes in 2039 — less than 2% of the country’s overall energy emissions, even with 45Q tax credits at $85 per tonne driving deployment.³

CCS will not solve the climate problem alone. But it is necessary for a category of emissions that cannot be decarbonized any other way — and the technology choices being made right now will determine which projects are economic, which are not, and which industries emerge from the energy transition with a credible path forward.

This is Part 1 of a two-part series. This article maps the technology landscape, explains the concentration gradient that drives all CCS economics, and examines where the industry’s honest track record stands. Part 2 will cover the industrial acquisition wave, the commercial models emerging from it, and why a deal signed in October 2025 between Google and a Decatur, Illinois power developer signals that CCS is about to find its largest new customer: the data center.

Subscribe now

The Last Mile, Explained

The phrase “last mile” in CCS is not rhetorical. It describes a technical reality that applies across the decarbonization landscape: most emissions can be addressed by switching fuels, deploying renewables, or improving efficiency. A shrinking but significant category cannot.

Cement production emits roughly 2.4 billion tonnes of CO2 per year globally — approximately two-thirds of which comes from the calcination reaction that converts limestone to clinker.² This process CO2 is chemically unavoidable: it is the product of the reaction itself, not a byproduct of burning fuel. Electrification cannot eliminate it. Fuel switching cannot avoid it. Only carbon capture removes it.

Steel production, chemical manufacturing, and several other industrial processes share a similar structure. The IEA estimates that carbon capture and utilization/storage (CCUS) delivers 38% of emissions reductions in the chemicals sector and roughly 15% in cement and iron and steel under its Sustainable Development Scenario.² These are not sectors where CCS competes with renewable electricity — they are sectors where CCS is the only option.

The IPCC’s Sixth Assessment Report (AR6) makes this point in aggregate. All scenarios limiting warming to below 2°C rely on some degree of carbon dioxide removal. Even the most renewable-heavy 1.5°C pathway requires more than 3 billion tonnes captured annually by 2050.⁴ The IPCC’s median scenario for low-overshoot 1.5°C pathways requires 196 billion tonnes cumulative stored through 2050.⁴

That number is not achievable through direct air capture at current costs, nor through a single technology. It requires scaling every viable approach — industrial point-source capture, natural gas generation with CCS, bioenergy with CCS, and eventually DAC — simultaneously. The question is not whether CCS matters. The question is which technologies get there fastest, at what cost, and under what conditions they work.

The answer starts with a chart you won’t see discussed enough.

The Concentration Gradient Is Everything

NETL’s 2022 study of industrial CCS costs across nine sectors is one of the most useful datasets in the field.⁵ It separates sources into two categories that explain almost everything:

High-purity streams (greater than 95% CO2): Ethanol fermentation, ethylene oxide production, ammonia synthesis, natural gas processing. Average capture cost: $17.50 per tonne. The economics are favorable today, in many cases without 45Q credits.

Low-purity streams (less than 50% CO2): Cement, steel, gas-to-liquids. Average capture cost: $62.00 per tonne. These require policy support — specifically the 45Q credit at $85 per tonne for geological storage — to be economically viable.

NGCC power plants sit at the challenging end of the dilute-stream category. At 3–4% CO2 in the exhaust, capturing 90% requires a $61-per-tonne process based on NETL’s updated 2023 baseline.⁶ That is marginal economics with 45Q, which is why the first large-scale NGCC project with integrated CCS required a Google power purchase agreement (PPA) to anchor the financing — the credits alone were not sufficient. As we examined in our analysis of hyperscaler carbon accounting and the structural disclosure gap between climate commitments and physical power infrastructure, the data center sector’s simultaneous expansion of fossil generation and CCS contracting reflects a deep tension in how large technology companies are managing scope emissions obligations.

DAC sits at the extreme. The Stratos project (1PointFive/Occidental, West Texas) is targeting $400–$500 per tonne at current scale.¹⁶ The Harvard Belfer Center’s analysis of long-term cost trajectories projects $341 per tonne for liquid solvent systems and $374 per tonne for solid sorbent systems at 1 billion tonnes of cumulative deployment — the frequently cited $100-per-tonne DAC target is not achievable at any near-term scale.¹⁷

The concentration gradient is therefore the deployment sequence. High-purity industrial streams first. Then coal combustion (10–14% CO2). Then gas combustion (3–4%). Then DAC. Each step down the concentration curve requires either better technology, higher subsidies, or a customer willing to pay a premium — and usually all three.

Five Approaches to the Same Problem

The CCS technology landscape is commonly simplified to “amine scrubbing” — a characterization that captures the dominant commercial approach but obscures the dynamic competitive environment forming beneath it. Five distinct capture approaches exist, each with specific strengths, limitations, and trajectories. Understanding them is prerequisite to evaluating which projects will succeed.

Amine Absorption: The Established Standard

Amine absorption — specifically liquid-solvent scrubbing using amines such as monoethanolamine (MEA) or proprietary advanced formulations — is the TRL 9 benchmark of CCS technology. It is how 18 of the world’s commercial-scale post-combustion CCS plants operate, including those using Mitsubishi Heavy Industries’ (MHI) KM CDR Process.¹⁸ It works. It is commercially proven. And it carries a set of operational challenges that are motivating every alternative technology in this space.

The mechanism: CO2-containing flue gas contacts the amine solvent in an absorber tower, where CO2 binds to the amine at low temperatures. The CO2-laden solvent is then heated to 120–150°C in a stripper column, releasing concentrated CO2. The regenerated solvent cycles back to the absorber. The captured CO2 is compressed for transport or injection.

Performance at commercial scale: MHI’s KM CDR Process guarantees 90% CO2 removal efficiency, with purity of 99.9%.¹⁸ The energy penalty — the parasitic load imposed by operating the capture system — runs 15–30% of a plant’s net electrical output. For a 550-megawatt plant capturing at full design rate, that equates to 125 or more megawatts of additional generation capacity required just to power the capture process.¹²

The drawbacks deserve direct treatment. MEA, the most common benchmark solvent, degrades through two pathways: oxidative degradation from oxygen in the flue gas, and thermal degradation at high regeneration temperatures. The degradation products include nitrosamines and nitramines, compounds classified as potentially carcinogenic.²⁹ Measured solvent loss at Technology Centre Mongstad in Norway reached 1.5 kilograms of MEA per tonne of CO2 captured — significantly above the theoretical range of 0.01 to 0.8 kilograms.²⁹ A life-cycle analysis found that adding MEA-based CCS increases freshwater toxicity impact by a factor of ten from MEA emissions alone.

Physical footprint is a constraint that project developers consistently underestimate. The absorber column for a 500-megawatt plant at 90% capture can exceed 150 meters in height. This is not a trivial retrofit on a plant not designed to accommodate it.

Water consumption increases 38–90% per megawatt-hour when adding CCS, depending on cooling configuration.¹² In water-scarce regions, this is a genuine siting barrier.

Why does amine scrubbing dominate despite these drawbacks? TRL 9 maturity. An established vendor ecosystem — MHI, Shell CANSOLV (used in Shell’s Quest project), BASF/Linde, Fluor Econamine — provides engineering certainty and proven performance guarantees. For dilute streams in the 3–14% CO2 range, amine scrubbing achieves higher CO2 recovery at lower energy cost than any competing alternative currently at commercial scale. The toxicity and footprint issues are manageable. The capital cost is known. Lenders will finance it.

For the next 5–7 years, amine scrubbing will win most large-scale project competitions on industrial and power sources. The question is whether the challengers can close the gap.

Cryogenic Separation: High Performance, Pre-Commercial

Cryogenic carbon capture (CCC) exploits the physical behavior of CO2 at low temperatures: CO2 desublimes (transitions directly from gas to solid) when cooled below approximately -78°C, allowing it to be separated from flue gas and collected as a frost-like solid on heat exchanger surfaces. The CO2 is then melted, compressed, and collected.

The technology was developed by Sustainable Energy Solutions (SES), whose assets were acquired by Chart Industries. DOE/NETL-funded technoeconomic analysis found the CCC process achieves 30–50% lower overall energy and economic costs than competing processes involving air separation units or solvents, with a parasitic load below 17% of plant output — roughly half the energy demand of equivalent amine scrubbing.²⁶ Projected cost: under $45 per tonne, well below amine scrubbing benchmarks at comparable conditions.²⁶

An additional advantage: cryogenic separation simultaneously removes co-contaminants — sulfur oxides (SOx), nitrogen oxides (NOx), and mercury — from the flue gas, potentially reducing downstream treatment requirements.²⁷³⁰

The limitation is concentration sensitivity. Energy requirements for cryogenic capture range from 2.4 to 5.2 gigajoules per tonne of CO2, varying significantly with inlet CO2 concentration.²⁷ At 20% CO2 concentration, the process requires approximately 350 kilowatt-hours per tonne. At 10%, that rises to 550–600 kilowatt-hours per tonne.²⁷ For NGCC exhaust at 3–4% CO2, the energy penalty becomes steep — cryogenic is not the preferred approach for dilute gas streams. It suits medium-to-high concentration industrial sources (cement, steel, waste-to-energy) far better.

Commercial status: TRL 5–7. Field pilot testing completed at Laramie, Wyoming. No full commercial-scale installations as of 2025. Chart Industries is pursuing commercialization.³⁰

Membrane Separation: Lowest Capex, Highest Power for Dilute Streams

Polymer membranes exploit differential permeability — CO2 passes through the membrane material faster than nitrogen or other gases due to higher solubility and diffusivity. Multi-stage membrane cascades can achieve high CO2 recovery and purity from flue gas streams.

The leading commercial technology is MTR Carbon Capture’s Polaris membrane, which is approximately 20 times more permeable than prior commercial CO2-selective membranes. A patented selective recycle process concentrates CO2 in the flue gas before membrane separation, reducing energy and capital cost. In October 2024, MTR announced completion of the world’s largest membrane-based carbon capture plant at the Wyoming Integrated Test Center — 150 tonnes per day from Basin Electric’s Dry Fork coal power plant, achieving 90% capture rate and producing 99.9%-plus pure liquid CO2.²⁸³² This is the first commercial-scale membrane CCS plant to enter operation.

The economic profile is specific: membrane systems offer the lowest capital expenditure among post-combustion separation technologies because the design is mechanically simple — no regeneration columns, no liquid solvent inventory, no heat exchangers for solvent cycling. The capital cost advantage is real.

The energy cost disadvantage is also real. Separation relies on pressure differential as the driving force, and for dilute flue gas at 3–14% CO2, creating sufficient driving force requires significant compression or vacuum work. For NGCC applications, membrane system costs run $50–$100 per tonne — above amine benchmarks at equivalent CO2 concentrations. At higher concentration sources like coal flue gas or natural gas processing, cost competitiveness improves substantially.

The practical implication: membrane separation is well-suited to high-concentration streams, natural gas processing plants, and applications where capital cost minimization is the primary constraint. It is not currently the preferred approach for dilute NGCC exhaust at 3–4% CO2.

Commercial status: TRL 7. MTR’s Wyoming plant represents the frontier.

Adsorption: Three Sub-Types, One Commercial Breakthrough

Adsorption-based capture uses solid materials that bind CO2 selectively, then release it when conditions change. Three distinct sub-types have different performance profiles and application niches.

Pressure swing adsorption (PSA) and vacuum pressure swing adsorption (VPSA) cycle solid adsorbents — typically zeolite 13X or activated carbon molecular sieves — between CO2-loaded and regenerated states using pressure as the driving force. CO2 adsorbs at elevated pressure; desorbs when pressure drops. PSA is TRL 6–7 for pre-combustion and high-concentration applications, where it is already commercially proven for hydrogen purification. Application to dilute post-combustion flue gas (3–14% CO2) faces selectivity challenges at low CO2 partial pressures. Operating costs are roughly comparable to MEA amine at moderate concentration conditions.

Metal-organic frameworks (MOFs) represent the most strategically interesting advancement in the adsorption category. MOFs are crystalline porous materials constructed from metal ions linked by organic ligands, creating extraordinarily high surface area and tunable pore chemistry. The key commercial breakthrough: BASF announced in October 2023 that it became the first company to produce MOFs at scale of “several hundred tons per year” for CCS applications.³¹ The specific MOF is CALF-20 (Calgary Framework 20), developed in partnership with Canadian CCS company Svante. CALF-20 captures up to 95% of CO2 from industrial point sources using rapid thermal swing adsorption with low-pressure steam. A demonstration at a cement plant in Richmond, Canada captured approximately 1 tonne of CO2 per day from flue gas. Svante claims the CALF-20-based system can capture CO2 at less than half the capital cost of existing amine solutions — an aggressive claim that lacks commercial-scale validation but reflects the real cost structure advantages of solid adsorbents over liquid solvent systems.³¹

Current TRL: 5–6. The production-scale MOF supply chain is in place. Commercial-scale validation is not yet complete.

Thermal swing adsorption (TSA) uses temperature rather than pressure as the regeneration driving force: CO2 adsorbs when the sorbent bed is cool, desorbs when the bed is heated. This creates a specific advantage for engine and gas compression applications: the waste heat in engine exhaust — at 450–550°C for reciprocating engines, far above the 80–200°C needed for most sorbent regeneration — powers the desorption cycle at minimal additional energy cost.

Caterpillar’s Oil & Gas division developed a three-bed TSA system using molecular sieves and ran a 1,000-hour pilot at an active gas compression site in Northwest Texas using a Cat G3606 natural gas compression engine rated at 1,875 horsepower.³⁵ The pilot captured up to 20 tonnes of CO2 per day. Projected commercial-scale performance: greater than 95% capture rate, under $40 per tonne, capable of 100-plus tonnes per day.³⁵ Regeneration requires no additional electricity — the energy comes from exhaust heat that would otherwise be rejected to atmosphere. Caterpillar’s positioning here mirrors its broader strategy in the data center market, where — as we documented in our analysis of the $42 billion onsite generation boom — the company is building carbon management capabilities into the same engine product lines it sells to oil and gas operators and industrial customers.

This is distinct from Caterpillar’s 2021 acquisition of CarbonPoint Solutions, whose Semi-Closed Cycle (SCC) technology concentrates exhaust CO2 from 3–6% to up to 25% before amine capture, targeting distributed power generation at 1–25 megawatts.²³ The TSA approach is for gas compression equipment. Both leverage the unique thermal properties of engine exhaust; both target markets that traditional amine scrubbing at power-plant scale does not efficiently serve.

The Storage Question: What Does “Permanent” Actually Mean?

Carbon captured from an industrial stack has no climate value unless it stays captured. The permanence question — how confident are we that stored CO2 will remain stored? — gets too little attention in most CCS analysis. It should not.

Geological Storage: The 10,000-Year Case

Deep saline aquifers hold the largest identified geological storage potential globally and operate through four trapping mechanisms on overlapping timescales:

Structural trapping is immediate: a confining cap rock prevents upward CO2 migration. This is the primary mechanism at most operating projects.

Residual trapping operates over years to decades: CO2 becomes immobilized in pore spaces as a discontinuous phase, too fragmented to migrate.

Solubility trapping operates over decades to centuries: CO2 dissolves into formation brine, becoming denser than the surrounding fluid and sinking rather than rising. After roughly 1,000 years, this mechanism dominates.

Mineral trapping — the most permanent mechanism — operates on timescales of 500 to 10,000 years: dissolved CO2 reacts with dissolved metals and minerals to form carbonate rock, permanently incorporating the carbon into the geological matrix.

Sleipner, the Equinor CCS project on the Norwegian continental shelf that began injection in 1996, provides the most extensive long-term monitoring dataset in existence. Over 16 million tonnes of CO2 had been stored in the Utsira Formation by 2020 — and continuous injection across 28-plus years of operation has produced zero detected leakage.² Sleipner is the empirical foundation for geological storage permanence claims, and the data it has generated over nearly three decades makes it the most important monitoring dataset in the field.

The primary leakage risk for geological storage is not cap rock failure — it is existing wells. Abandoned wellbores from prior drilling activities provide migration pathways if well integrity is not maintained. This drives the critical importance of site characterization and the bottleneck discussed below: the EPA Class VI well permitting program.

CO2-Enhanced Oil Recovery: The Permanence Controversy

CO2-enhanced oil recovery (CO2-EOR) injects CO2 into mature oil reservoirs to improve production. One-third to one-half of the injected CO2 is permanently trapped in the reservoir; the remainder is produced with oil, separated, and re-injected in a closed loop.¹¹ Whether CO2-EOR represents carbon storage depends heavily on accounting boundaries.

The Clean Air Task Force’s lifecycle analysis found that CO2-EOR operations initially appear carbon-negative in a short-term accounting window — the CO2 injected for storage exceeds the CO2 produced by combustion of the additional oil recovered. But after roughly 10 years, EOR operations typically turn net carbon-positive when full lifecycle emissions — including combustion of the produced oil — are included.¹⁹ ISO Standard 27916:2019 attempts to standardize CO2-EOR accounting but excludes full lifecycle combustion emissions, a significant methodological gap.

Under the One Big Beautiful Bill Act (OBBBA), signed July 4, 2025, the 45Q credit for EOR utilization was equalized with the geological storage rate at $85 per tonne — up from $60 per tonne under the original IRA rate.¹⁴ This removes the prior policy differential that had disadvantaged geological storage relative to EOR. Whether EOR-based CCS delivers genuine climate benefit remains contested; what is clear is that the financial incentive structure now treats both equally.

The Concrete Option: Mineralization at Scale

CarbonCure Technologies has deployed more than 750 carbon capture systems in 30-plus countries using a different permanence pathway: injecting captured CO2 into fresh concrete during mixing, where it reacts immediately with calcium silicate hydrates in the cement paste to form nano-sized limestone particles permanently embedded in the concrete matrix. Even if the concrete is demolished, the carbon remains mineralized.

The theoretical scale potential if 10% of global construction materials incorporated CO2 mineralization: 1.6 billion tonnes of CO2 per year — roughly 4.5% of global emissions. Whether that scale is achievable depends on supply chains and cement industry adoption, but the permanence is unambiguous and the deployment pathway is fully commercial today.

The Food-Grade Paradox

Atmospheric CO2 concentration is rising. Commercial CO2 supply is simultaneously declining.

These two facts describe completely decoupled systems. Commercial CO2 supply — the liquid CO2 used for beverage carbonation, food refrigeration, modified atmosphere packaging, and numerous industrial applications — depends on production byproducts from specific industrial processes: primarily ammonia synthesis and ethanol fermentation. Over 60% of global liquid CO2 supply comes from these byproduct streams.³⁶ As European ammonia plants shut down due to high natural gas prices and US ethanol production fluctuates, commercial CO2 supply constricts.

Olivon Advisors estimated that US merchant CO2 supply could fall 33% over the next decade while demand grows approximately 2% per year.³⁶ During shortage periods in 2024–2025, prices reached $600–$1,000 per tonne for some buyers — against a normal bulk rate of $100–$120 per tonne.³⁶

Food-grade CO2 requires purity of 99.9%-plus by volume, with strict limits on specific impurities including acetaldehyde, benzene, and sulfur compounds. Ethanol fermentation CCS — the ADM model in Decatur, Illinois — already produces near-food-grade CO2 as a process byproduct, since the fermentation stream contains no combustion impurities. With standard drying, deodorization, and liquefaction, the food-grade spec is achievable at modest cost premium.

Amine capture from combustion flue gas faces a harder path. MEA degradation products — particularly nitrosamines — must be completely removed before food-grade qualification, requiring additional activated carbon treatment steps. Cryogenic capture and MTR’s membrane process produce 99.9%-plus pure CO2 streams that are potentially compatible with food-grade qualification after standard purification.

The strategic implication: CCS projects co-located with ethanol plants, biogas facilities, or other high-purity sources have a second offtake channel — food-grade CO2 sales at $100–$120 per tonne — that supplements 45Q credit revenue and reduces policy dependency. The incremental cost of upgrading from pipeline-grade to food-grade CO2 adds an estimated $10–$25 per tonne, and the resulting product commands a price premium well above the 45Q credit for geological storage. For a narrow but real category of projects, food-grade CO2 offtake makes the economics work without subsidies.

The technology taxonomy and storage permanence framework above is the reference layer — the five capture approaches, their honest tradeoffs, and what “permanent” actually means across geological storage, CO2-EOR, mineralization, and food-grade utilization. Paid subscribers get the harder analysis: where these projects have actually performed versus design specs, the subsidy math that most CCS coverage sanitizes, which economics work without government support, and the OEM acquisition wave that signals where the real commercial scaling is happening.

Where the Technology Pencils Out: Project Economics, Honestly Assessed

Read more

In February 2024, the Southern Nevada Water Authority (SNWA) enacted the most aggressive data center water policy in the United States: an outright ban on evaporative cooling in all new commercial construction across Las Vegas, Henderson, and the surrounding valley. The rationale was sound. Twenty-three Southern Nevada data centers consumed 716 million gallons of water in 2024, every drop permanently lost from the Colorado River system.⁷ Google’s Henderson campus alone accounted for 352 million gallons — water that cannot be returned to Lake Mead, cannot earn return-flow credits, and cannot be recovered for the 40 million people downstream who depend on the river.^7,19

The SNWA’s instinct was correct. Its accounting was incomplete.

Air-cooled data centers report a Water Usage Effectiveness (WUE) of zero — the industry metric that regulators, sustainability officers, and ESG analysts treat as meaning “water-free.”²³ It means nothing of the sort. Air-cooled chillers consume 2 to 3 times more electricity than evaporative-cooled equivalents for the same information technology load. That electricity must come from somewhere. In Nevada, it comes primarily from natural gas combined-cycle (NGCC) power plants — plants that cool their condensers with water. When a data center switches from evaporative cooling to air cooling, its WUE drops to zero. Its upstream water burden rises by millions of gallons per year. The water does not disappear. It moves — from a visible cooling tower on the data center’s property to an invisible cooling tower at a power plant, somewhere else, outside the accounting boundary of every current regulatory framework.

This is not a fringe claim. At the ASHRAE Winter Conference held in Las Vegas in February 2026 — two miles from SNWA headquarters — three independent presenters reached the same conclusion. Ken Mortensen of SPX Cooling Technologies told a room of engineers and operators that “when power generation is considered, all cooling technologies consume both energy and water resources.”³ The industry already knows. Policy does not yet require measuring it.

Subscribe now

The Stakes: $600 Billion in Capital and 6 Gigawatts of New Demand

Nevada is not a corner case. It is the test case for how water-stressed jurisdictions will manage data center growth across the entire Sun Belt — and the magnitude of what is at stake makes the accounting gap consequential at a scale that demands attention.

The “big five” hyperscalers — Amazon, Alphabet/Google, Microsoft, Meta, and Oracle — are forecast to commit more than $600 billion in capital expenditure in 2026 alone, a 36% increase from 2025.²⁵ Amazon alone committed $200 billion for 2026 — a figure that stunned markets when disclosed during Q4 2024 earnings.²⁵ Every dollar of that capex produces a data center that will consume electricity for 20 to 30 years.

Nevada has become a premier destination for that spending. Microsoft, Google, Apple, Switch, EdgeCore, and Novva have all committed to the Nevada market.¹³ Vantage Data Centers announced a $3 billion, 224 megawatt (MW) campus in Storey County — first building opens in Q2 2026.⁹ Northern Nevada data center projects have collectively requested nearly 6 gigawatts (GW) of electricity capacity over the next decade.¹⁰ NV Energy’s chief executive has stated publicly that Northern Nevada’s grid could quadruple. That grid expansion requires power plants. Power plants require water. The feedback loop runs in the wrong direction from SNWA’s stated goal.

U.S. data center electricity consumption reached approximately 176 terawatt-hours (TWh) in 2023, representing 4.4% of total U.S. electricity demand.¹ Lawrence Berkeley National Laboratory (LBNL) projects this will reach 325 to 580 TWh by 2028 — a potential tripling in five years. The global hyperscale data center market stands at $205 billion in 2026 and is projected to reach $596 billion by 2031 at a 23.7% compound annual growth rate (CAGR).¹¹ S&P Global’s Sustainable1 unit finds that 43% of data centers globally are already exposed to high water stress.²⁶ Nevada is simply the first jurisdiction to convert that stress into an equipment ban — making it the policy laboratory for an experiment that will repeat in Arizona, Texas, and California within this decade. As we documented in our analysis of how hyperscaler self-generation shifts costs to residential ratepayers, the hidden system-level consequences of large infrastructure decisions rarely appear in the policy accounting that governs them.

What WUE Actually Measures (and What It Doesn’t)

Water Usage Effectiveness — defined as annual site water usage divided by IT equipment energy — is the data center industry’s primary water metric. Industry average WUE is 1.8 liters per kilowatt-hour (L/kWh). Best-in-class facilities achieve as low as 0.26 L/kWh.²³ Air-cooled facilities report 0 L/kWh. That figure is treated as proof of water efficiency.

It is a measurement of on-site water performance. Nothing more.

The European Union implemented the world’s most advanced mandatory data center water policy framework in September 2024, requiring annual WUE reporting for large facilities under the Energy Efficiency Directive.²² The EU’s own framework documentation acknowledges what its metric cannot see: WUE “does not include water embedded in purchased electricity.”²² The world’s most sophisticated water regulator has named the problem. No regulator has yet required measuring it.

When a data center switches from evaporative cooling to air cooling, its WUE drops to zero — and its electricity consumption simultaneously rises by a factor of two to three. The WUE metric records a perfect score. The physical system consumes more water. These two facts are not in contradiction; they are a consequence of WUE measuring only one accounting boundary. The industry term for this is an efficiency metric with an incomplete system boundary. The practical consequence: SNWA’s ban, designed to eliminate data center water consumption, is increasing it.

February 2026: The Industry Validates the Gap

The ASHRAE Winter Conference in Las Vegas in early February 2026 produced the clearest documented evidence that the cooling equipment industry, legal advisors, and academic researchers have independently internalized the systems-level accounting problem — while regulators have not.

Seminar 6, titled “Evaporative Cooling in the Southwest,” brought together three presenters who each arrived at the same thesis from different directions.³ Mortensen of SPX Cooling Technologies framed the issue in terms of what the industry had historically missed: cooling tower bans reduce visible water use at the cost of dramatically increased electricity consumption, which imposes its own water burden on the grid. His learning objective for conference attendees was to “Acknowledge the Water/Energy Resource Balance inherent in cooling technologies” — language that signals an industry consensus forming around what regulators have not yet required anyone to report.

Michael Nerozzi of Baltimore Air Coil took the policy critique further, flagging that Nevada’s ordinance definitions are “overly broad and do not properly account for the water use and associated benefits of varying forms of evaporative cooling equipment available.”³ Adam Green of Baker Donelson, representing building owners and operators navigating compliance, characterized the systems mandated by the ban as “exponentially less efficient” and documented the practical impacts on operators adapting to the new regulatory regime.³

Separately — presented in Milano III/IV at Caesars Palace on February 4, 2026, two miles from the SNWA offices enforcing the ban — Pooya Navid of Nortek Data Center Cooling Solutions presented ASHRAE Paper LV-26-C185: the first peer-reviewed simulation framework explicitly modeling energy-water tradeoffs for artificial intelligence (AI)-driven data centers across U.S. climate zones.⁴ The paper compares air-cooled chillers, water-cooled chillers, evaporative systems, hybrid coolers, and dry coolers across the full range of U.S. climate conditions. Its explicit learning objective: “Compare energy-water trade-offs across climates and supply-water setpoints.”⁴

This is the industry publishing its homework. The SNWA ordinance was enacted without this analysis. Navid’s paper demonstrates what a systems-level evaluation would have shown.

The Upstream Water Math: Doing What Regulators Won’t

Read more

Alphabet’s $4.75 billion acquisition of Intersect Power, closed in March 2026¹, is the loud signal that the PPA is no longer fit for purpose. By bringing one of the nation’s most aggressive clean energy developers in-house, Google isn’t just buying solar panels; it is vertically integrating into the energy supply chain to bypass a grid that is no longer fast enough or reliable enough to support its $240 billion Cloud backlog².

This move represents a fundamental shift in the hyperscale power playbook: the transition from being a “customer” of the grid to becoming a “co-operator” of it. With Alphabet’s 2026 capital expenditure (CapEx) guidance reaching a staggering $175 billion to $185 billion², the company is building its own private utility infrastructure. This is the rise of the “Energy Park”—a co-located, behind-the-meter (BTM) model designed to scale compute at the speed of silicon, not the speed of the Federal Energy Regulatory Commission (FERC).

The PPA Era’s Terminal Phase: Why the Old Model Broke

The PPA model relied on two assumptions that are now obsolete: that the grid has sufficient capacity to deliver power wherever it’s needed, and that interconnection queues are manageable.

In 2026, the reality is a “Queue Crisis.” Data from the Lawrence Berkeley National Laboratory (LBNL) reveals a staggering 2.3 terawatts (TW) of capacity sitting in interconnection queues as of the end of 2024, with median wait times exceeding four years³. For a hyperscaler, a four-year wait for power is an existential threat to revenue. If you cannot plug in your chips, you cannot monetize your AI models. Our February 2026 analysis of the data center power supply crisis documented how grid interconnection delays of 3-7 years in PJM have already pushed BTM economics past the tipping point for large industrial loads.

The Load Growth Inflection

We are witnessing the most significant reversal in U.S. electricity demand in 15 years. After a decade of flat load, the Energy Information Administration (EIA) projects total U.S. electricity demand to grow to 4,283 billion kWh in 2026⁴. This trend will accelerate, with the EIA’s Annual Energy Outlook projecting that data centers will drive commercial sector growth to 818 billion kWh in high-demand scenarios⁵. Data centers are the dominant driver of this growth, with the commercial sector’s electricity consumption rising at 5% annually⁴.

According to the National Renewable Energy Laboratory (NREL), data centers will account for 46% of all new U.S. load growth through 2030⁶. The U.S. data center fleet capacity will hit 51-62 GW by the end of 2026, a massive jump from the ~31 GW seen just three years ago⁶⁹.

In this environment, signing a PPA for a wind farm in Iowa to “offset” a data center in Northern Virginia is no longer enough. The physical reality of the grid—congested lines, transformer shortages, and regulatory friction—has caught up to the digital ambitions of Big Tech.

The AI Efficiency Paradox

The shift is also driven by the sheer power density of new AI hardware. Google’s seventh-generation Tensor Processing Unit (TPU), codenamed “Ironwood,” launched in April 2025⁸. While Ironwood delivers twice the performance per watt compared to its predecessor, it requires significantly higher rack power density—up to 0.85 kW per chip⁸. According to the Department of Energy (DOE), AI data centers now exhibit 10x to 20x higher power density than traditional facilities, creating localized thermal and electrical demands that traditional grid connections cannot meet⁹.

Traditional grid-tied data centers struggle to accommodate these densities without massive upgrades to local distribution infrastructure. By owning the energy developer, Google aligns its modular data center construction with the solar and storage build-out, creating “Energy Parks” that are optimized for AI from the ground up.

The Rise of the Energy Park: Speed as the New Moat

The primary competitive advantage in 2026 is no longer the lowest Levelized Cost of Energy (LCOE). It is “Speed to Power.”

The “Energy Park” model co-locates hyperscale data centers with gigawatt-scale solar and long-duration battery energy storage systems (BESS). These facilities operate primarily behind-the-meter (BTM), using the grid as a secondary reliability layer rather than the primary source of electrons. This shift from grid dependence to BTM self-supply is not unique to Google—we tracked its early momentum in our analysis of how AI data centers drove the $42 billion onsite generation boom, when gas turbine supply chain constraints first forced hyperscalers to rethink power procurement from first principles.

This model allows Google to scale compute before a full grid connection is even secured. In March 2026, Phase 1 of the first Google-owned Energy Park entered operations¹⁰. By utilizing 500 MW of BTM solar and 2 GWh of BESS, Google achieved a “construction-to-operation” timeline of just 18 months—bypassing a five-year PJM Interconnection (PJM) queue for a 350 MW grid injection agreement¹⁰.

Vertical Integration as “Safe Harbor”

Owning Intersect Power gives Alphabet direct control over a 10.8 GW pipeline of clean energy assets targeted for delivery by 2028¹¹. More importantly, it secures Alphabet’s place in the global supply chain.

By owning the developer, Alphabet ensures priority access to critical components in a constrained market. Intersect Power is a top-tier customer for First Solar’s Series 7 modules, part of a 90 GW backlog that provides protection from supply chain risks associated with Foreign Entities of Concern (FEOC)¹². Similarly, the acquisition secures access to Tesla’s Megapack capacity, with Tesla targeting 80 GWh of annual production in 2026 to serve exactly these types of co-location use cases¹³.

For Alphabet, the $4.75 billion price tag is a strategic insurance policy. It guarantees that the $240 billion Cloud backlog won’t be held hostage by a transformer shortage or a regulatory delay in a remote PJM substation.

Technical Deep Dive: The Ironwood/Energy Park Architecture

The vertical integration of Alphabet and Intersect Power is a technical optimization of the entire AI stack, from the silicon to the substation.

TPU v7 “Ironwood” and the Architecture of Scale

Google’s custom silicon, the TPU v7 “Ironwood,” is the heartbeat of these new Energy Parks. Ironwood offers 4.6 PetaFLOPS of FP8 compute per chip, making it 30x more power-efficient than the TPUs Google deployed just eight years ago⁸. However, this efficiency comes with a requirement for massive, localized power delivery that traditional grid architectures cannot sustain.

Google organizes these chips into pods of 9,216 units, delivering a combined 42.5 Exaflops of compute⁸. At this scale, the traditional “rack-and-row” power distribution model fails. To manage the thermal load, Google employs third-generation liquid cooling, which allows the facility to sustain peak performance without the massive HVAC overhead that characterizes traditional air-cooled data centers.

The real innovation, however, lies in the networking. Google’s use of Optical Circuit Switching (OCS) at the inter-pod level reduces networking power consumption by up to 40%⁸. By replacing power-hungry electrical switches with light-based switching, Google allocates more of its power budget directly to the chips themselves.

DC Power: Eliminating the Conversion Tax

By owning Intersect Power, Google can design “Energy Parks” where the solar and storage arrays are electrically integrated with the data center pods at a fundamental level. This allows for a “Direct-to-Chip” DC power architecture.

In a traditional data center, power is converted from DC (solar) to AC (grid/inverter), stepped up for transmission, stepped down for distribution, and then converted back from AC to DC for the server. Each of these five conversion steps incurs a “tax” of 3-5% in energy loss. By co-locating generation with the load, Google can utilize high-voltage direct current (HVDC) distribution within the campus, feeding DC power from the solar arrays and BESS units directly into the liquid-cooled racks. This eliminates at least three conversion stages, improving overall facility efficiency by 12-15%—a staggering advantage when operating at the gigawatt scale.

In Phase 1 of the Google-Intersect flagship project, the 500 MW solar array and 2 GWh BESS serve as the primary power source for a modular data center campus¹⁰. This BTM configuration allows for “islanding” during periods of grid stress or high pricing, turning the data center into a massive flexible load that can provide reliability services back to the grid when it’s not training models. This is not just a data center; it’s a “Parallel Grid” designed for the AI era.

The full analysis continues below for paid subscribers, including detailed economic analysis, competitive landscape assessment, regulatory implications, and strategic recommendations with comprehensive citations.

Economic Analysis: The 3-5x Cost Advantage

Read more

That does not make hyperscaler climate commitments meaningless. It exposes a harder problem than most industry coverage admits. The real issue is not that these companies found a way to route gas emissions into Scope 3. In many cases, owned on-site generation is Scope 1 and electricity purchased from a third-party dedicated plant is Scope 2 under existing greenhouse gas (GHG) accounting rules.¹⁵, ¹⁷ The stronger critique is more uncomfortable: current accounting and disclosure conventions preserve a low-carbon public narrative even as the physical power system serving AI load becomes materially more fossil-dependent.

That distinction matters. It separates an article that accuses Microsoft and Meta of accounting abuse from one that identifies a structural disclosure gap in how corporate decarbonization is communicated. The second argument is harder to dismiss. It does not require proving fraud. It requires showing that public climate language is cleaner than the underlying energy reality.

The Parallel Power System Is Already Being Built

The buildout is easy to underestimate because no clean federal dataset yet exists for all U.S. data-center-specific gas generation in development. The U.S. Energy Information Administration (EIA) only launched a pilot survey on data center energy use on March 25, 2026, covering site characteristics, energy sources, and electricity consumption in selected regions.¹⁹ The buildout is moving faster than the statistics.

State records already show the direction of travel. In its 2025 year-in-review, the Public Utilities Commission of Ohio (PUCO) reported that the Ohio Power Siting Board approved four behind-the-meter natural gas plants at central Ohio data center sites totaling 736 MW during 2025.²³ On February 3, 2026, the Ohio Power Siting Board approved the 350 MW Apollo Power Generation Facility in Wood County for an adjacent data center, with battery storage attached.²² Ohio alone carries an official, cited floor of 1,086 MW of data-center-linked behind-the-meter gas generation from the approvals reviewed here.²², ²³

This is not hypothetical future demand. It is physical infrastructure already under approval.

Microsoft’s case is the clearest new example of the market’s evolution from on-site generation to dedicated power procurement. Chevron confirmed on March 31, 2026 that it, Microsoft, and Engine No. 1 had entered an exclusivity agreement tied to power generation and electricity offtake.⁴ Chevron framed the model as co-located, behind-the-meter generation serving AI load.⁴ Reuters reporting described the proposed West Texas plant as initially generating 2.5 gigawatts (GW), costing roughly $7 billion, and targeting a 2027 startup.⁶ A joint announcement from Engine No. 1, Chevron, and GE Vernova described plans for up to 4 GW of co-located natural gas power, using GE Vernova gas turbines, designed with flexibility to integrate carbon capture and storage (CCS) later.⁵

This is not a company buying generic grid electricity and balancing it with renewable certificates in the background. Microsoft is moving toward a dedicated fossil-backed power relationship because the power market it depends on cannot deliver enough capacity fast enough. As we detailed in our analysis of why behind-the-meter generation won the race against grid interconnection, grid connection timelines in PJM now run three to seven years — a wait no hyperscaler building for AI can absorb.

Meta’s Ohio case differs in structure but not in strategic meaning. On September 18, 2025, the Ohio Power Siting Board approved the 200 MW Socrates North power plant for Will-Power OH, with PUCO’s public bulletin confirming that Sidecat, a Meta affiliate, would consume the power and that the facility would operate behind the meter.²¹ Williams describes Socrates North and South as two separate 200 MW generation sites — 400 MW combined — targeting service in the third and fourth quarters of 2026.²⁴

Meta is not alone in this architecture. EdgeConneX has proposed a 120 MW Ohio gas plant as the primary power source for a data center.⁸ Crusoe announced a 900 MW AI campus in Abilene to support Microsoft workloads, explicitly framing the development as energy-first AI infrastructure with a dedicated on-site power plant.⁹ xAI’s Memphis project has become the industry’s most visible flashpoint: environmental groups filed notice of intent to sue over alleged unpermitted gas turbine use in February 2026.¹⁰, ³³

Every project has a different structure. The market has converged on a single answer to the AI power bottleneck: if utilities cannot deliver power on hyperscaler timelines, hyperscalers and their infrastructure partners will buy or build the power system themselves.

The Accounting Question Is More Nuanced Than “They Hid It in Scope 3”

Most coverage becomes sloppy here. “They are using gas and claiming it as Scope 3” sounds scandalous, but it is too imprecise to carry the argument.

Under the GHG Protocol, emissions from owned or controlled combustion sources are Scope 1; purchased electricity is Scope 2.¹⁵ Treatment of leased assets depends on organizational boundary choices and operational control.¹⁵ Microsoft’s own sustainability-accounting documentation makes this explicit: owned on-site generation falls under Scope 1, and electricity purchased from an on-site plant the company does not own is treated as purchased generation in Scope 2.¹⁷

The stronger claim is not that hyperscalers found a Scope 3 cheat code. The stronger claim is that different contractual structures preserve clean public claims while shifting where the emissions appear and how they are narrated.

An owned on-site gas plant is difficult to sanitize rhetorically. A third-party dedicated power agreement is cleaner from a messaging standpoint. The fossil combustion still happens. The electrons still arrive. But the company discusses the arrangement through the language of purchased electricity, portfolio-wide renewable matching, carbon-free energy ambition, or residual-emissions management — rather than through the language of direct fossil combustion.

That is fully consistent with current accounting standards. That is precisely the problem.

The accounting framework is not broken. The disclosure framework is. And the difference is what investors, regulators, and the public are not being told.

The full analysis continues below for paid subscribers: how the math of a single 2.5 GW gas deal threatens Microsoft’s entire removal portfolio, why “temporary” is a communications claim rather than an engineering commitment, what Meta’s public disclosures don’t say about its Ohio gas infrastructure, how annual renewable matching obscures hourly fossil dependence, and why current AI power disclosures fall below the standard investors should demand.

Microsoft’s Climate Story Gets Harder to Read Once Gas Enters the Picture

Read more

Germany’s largest distributed energy success story isn’t offshore wind or grid-scale battery storage. It’s a $1,500 appliance that plugs into a standard wall outlet.

As of June 2025, Germany had 1,031,298 registered balcony solar installations — up from roughly 500,000 a year earlier, a full doubling in twelve months.¹ Combined, those panels produce nearly 1 gigawatt of electricity. Across that million-plus installation base, the safety incident count for compliant systems stands at zero.²

That track record is now landing in US state legislatures. More than 28 states plus Washington, D.C., have active plug-in solar legislation as of April 2026.⁴ Three have enacted laws. Colorado and Maryland have passed both chambers and are awaiting governor signatures.

Utah passed the first law unanimously in March 2025.⁶ Virginia followed in March 2026 with a 96-0 Senate vote; Governor Spanberger has until April 13 to sign.⁷ Maine became the third enacted law on April 6, when Governor Mills signed LD 1730.²² California’s “Plug Into the Sun Act” cleared its first committee 12-0 and advanced to the Appropriations Committee in April.⁸

In at least five states, utility lobbying has killed or delayed those bills — citing safety concerns that Germany’s million-installation track record directly contradicts.⁵, ²¹

This is not an obscure corner of the energy transition. This is a mass-market consumer technology that bypasses the entire traditional solar installation process — no permits, no inspectors, no utility interconnection agreements, no contractors. It works for renters. It works for apartment dwellers. It costs less than a laptop setup. And utilities understand exactly what that means for their business model.

The question is no longer whether plug-in solar arrives in America. It’s whether the rules get written before or after adoption reaches critical mass.

Germany Built the Proof of Concept

Germany’s balcony solar market deserves careful attention. In 2024 alone, German households registered approximately 435,000 new plug-in solar systems — a 41% year-over-year increase.¹ Between January and April 2025, another 135,000 systems came online, adding 53.4 megawatts of distributed capacity.¹ This is not a hobbyist phenomenon. This is mainstream consumer adoption.

The regulatory architecture that enabled this scale took years to develop but arrived at a remarkably simple endpoint. Under the 2024 “Solar Package 1” reforms, installing a balcony solar system in Germany requires filling out a five-field online registration form.³ The inverter output limit rose from 600 watts to 800 watts, with panel capacity allowed up to 2,000 watts.³ Tenants gained a legal right to install systems that landlords can only block in rare cases.³ No electrician is required for sub-800-watt systems. Systems can operate temporarily without a digital meter upgrade.³

The result is a technology that has genuinely reached consumer product status in Germany. Balcony solar kits sell at hardware stores. Reviews appear on consumer electronics sites. Germany’s official energy market data register — the Marktstammdatenregister — shows the adoption curve bending sharply upward every quarter.¹

The safety record is the most critical data point for understanding the US debate. Academic researchers analyzing the German deployment found zero safety incidents for compliant installations across the entire million-plus base.² The peer-reviewed paper “Barriers to Balcony Solar and Plug-In Distributed Energy Resources in the United States,” published in MDPI Energies in April 2025, concluded that while barriers are real, they are “solvable with appropriate policy and standards development.”²

UL Solutions agreed. On January 8, 2026, UL launched UL 3700 — the “Outline of Investigation for Interactive Plug-In Photovoltaic Equipment and Systems” — the first dedicated US certification framework for plug-in solar.¹⁴ The framework directly addresses the four technical concerns utility opponents had cited: touch safety, breaker masking, bidirectional ground-fault circuit interrupter (GFCI) compatibility, and backfeeding risk.¹⁴ Anti-islanding protection — the system shuts down within one second if the grid goes down, protecting utility workers — is mandatory under UL 3700.¹⁴

The safety objection no longer has a factual foundation. That matters, because the safety objection has been utilities’ primary tool for delaying legislation.

The US Legislative Wave

What is happening in state legislatures across the country right now is unusual in energy policy: broad, fast, bipartisan momentum.

Utah passed House Bill 340 in March 2025 — the first plug-in solar law in the United States — with unanimous bipartisan support. The law allows systems up to 1,200 watts, requires no utility approval or interconnection agreement, mandates National Electrical Code and UL compliance, and includes anti-islanding protection requirements. It also protects homeowners associations and local governments from restricting installations.⁶ Governor Spencer Cox signed it March 25, 2025. Effective May 7, 2025.

Virginia became the second state in March 2026. Senate Bill 250 and House Bill 395 passed 96-0 in the Senate and 93-4 in the House.⁷ The law covers systems up to 1,200 watts and explicitly prevents landlords with four or more units from prohibiting plug-in solar installations — a provision directly targeting the renter market.⁷ Governor Spanberger has until April 13, 2026 to sign. Implementation is delayed until 2027 to allow a stakeholder process, but the legislative direction is clear.

Maine became the third state on April 2, 2026, when its Senate passed LD 1730. Governor Mills signed it into law on April 6.²² Like Utah and Virginia, Maine caps systems at 1,200 watts and bars utilities from requiring interconnection agreements or fees. Maine’s law introduced a notable two-tier structure: systems up to 420 watts can be DIY-installed with no utility notification required; systems from 421 to 1,200 watts require a licensed electrician and 30-day notification to the utility. Maine’s law is also the first to specifically cite UL 3700 by name as a qualifying certification standard. The bill stops short of Virginia’s landlord protections — it imposes no restrictions on landlords prohibiting installations — leaving tenant access as an unresolved gap.

Colorado and Maryland have each passed plug-in solar bills through both legislative chambers and are awaiting governor action. Colorado’s HB 26-1007 passed the state Senate 29-4, establishing a framework similar to Utah and Virginia. The pace — five states through both chambers within fourteen months of Utah’s first-ever enactment — is unusual by energy policy standards.

California’s Senate Bill 868 — the “Plug Into the Sun Act,” introduced by Senator Scott Wiener — cleared the Senate Energy Committee 12-0 in March 2026.⁸ The California bill covers systems from 400 to 1,200 watts and specifically cites the finding that a single 400-watt panel covers approximately 14% of average apartment electricity consumption, saving roughly $250 per year.⁸

New York’s SUNNY Act (Senate Bill 8512), introduced by Senator Liz Krueger, would exempt plug-in solar from interconnection and net metering requirements entirely, with a 1,200-watt export limit.⁹ New Hampshire, Vermont, New Jersey, and Illinois have bills on a credible path to passage.⁴

Bright Saver, the California-based nonprofit driving much of the multi-state coordination, projects that 20 million or more US homes could adopt plug-in solar over the next decade with enabling policies in place.¹¹ Cora Stryker, one of Bright Saver’s co-founders, put it directly: “We are really looking at a movement here.”¹¹

That assessment is supported by the numbers. Approximately one-third of US households rent — over 45 million households — and almost none of them have access to rooftop solar.¹⁰ Traditional solar requires owning a roof, securing permits, hiring a licensed contractor, signing an interconnection agreement, and waiting months for utility approval. Plug-in solar requires none of that. It is the first viable path to self-generated electricity for the renter majority.

“The market framing most people miss is that plug-in solar isn’t competing with rooftop solar. It’s competing with doing nothing,” said Kevin Chou, co-founder of Bright Saver, the nonprofit coordinating the multi-state legislative push. “There are roughly 80 million renter and condo-owner households in the US that have never had any solar option at all. A year ago, zero states had laws on the books. Today five states have passed both chambers. The legislative momentum is real. But the standards process — getting UL 3700 finalized and the NEC updated — is the real bottleneck between enacted legislation and certified products on shelves. That’s the constraint that determines how fast this category develops.”

The Utility Opposition Playbook

In at least five states, utility lobbying has successfully blocked or delayed plug-in solar legislation. Arizona’s House Bill 2843 never came to a vote — a committee chairwoman blocked it after utility concerns were raised in February 2026.²¹ New Mexico, Washington, and Wyoming saw bills delayed by similar utility pressure.⁵

The utilities deploy four arguments. Three have technical responses. One is the real concern.

Argument one: backfeed risk to lineworkers. Utility workers performing maintenance during an outage could be endangered if plug-in solar systems continue energizing lines. The technical response: every UL-certified inverter includes anti-islanding protection that shuts the system down within seconds of detecting a grid outage.¹⁴ Germany’s 1 million deployments have demonstrated zero lineworker incidents from compliant systems.²

Argument two: clustering risk. Multiple unregistered plug-in systems on the same circuit could overload distribution transformers. The technical response: Germany has more than one million systems deployed without a documented clustering incident; individual system caps of 800-1,200 watts make transformer overload from residential clustering implausible.¹

Argument three: absence of safety standards. This argument was defensible before January 2026. UL 3700’s launch substantially narrowed it.¹⁴ But UL 3700 is an Outline of Investigation, not yet a finalized consensus standard, so the standards gap is narrowed, not fully closed.

The real argument: revenue loss. A plug-in solar system is, by design, a device that reduces utility electricity purchases without utility involvement, approval, or compensation. Utilities receive no interconnection fees. They do not control the installation timeline. They do not participate in the customer relationship. The World Resources Institute described it plainly: these systems could “save U.S. consumers billions of dollars a year in electricity costs” by directly offsetting retail-rate electricity consumption.¹⁰

This is disintermediation. Not safety.

The pattern is recognizable. When net metering emerged, utilities fought it with similar arguments and eventually shaped the compensation rates — sometimes reducing them significantly, as California did with its Net Billing Tariff. As we documented in our investigation of the solar cost-shift narrative, utility opposition to rooftop solar consistently deployed manufactured financial arguments — claims that independent analysis found overstated the costs by billions of dollars. When community solar expanded, utilities pushed for program caps and utility administration requirements. The instinct is consistent: if you cannot stop distributed energy, shape it into a form that preserves your revenue model.

The difference with plug-in solar is the scale of simplicity. There is no interconnection to approve. There is no meter to upgrade. There is no inspector to schedule. A consumer buys a kit, plugs it into a wall outlet, and their meter runs slower. The utility’s role is entirely absent — and that is precisely the threat.

“The existing regulatory framework was built for rooftop systems that are five to ten times larger than what we’re talking about,” said Kevin Chou of Bright Saver. “A sub-1,200-watt plug-in system on a dedicated circuit is a fundamentally different animal, and the technology has moved faster than the rules. In California, all three investor-owned utilities have opposed SB 868. But the political dynamics are shifting underneath them. Plug-in solar is affordable, it expands renter access, and it requires zero public subsidies. That combination is very hard for legislators to vote against,regardless of where the utility lobby comes down. The bipartisan margins tell the story — Virginia passed 93 to 4 in the House.”

The full analysis continues below for paid subscribers, including detailed technical analysis of the 240V/120V gap, economic modeling by region, the competitive landscape of the emerging US market, regulatory implications, and strategic recommendations for utilities, investors, and policymakers.

The 240V Problem: Why the US Lags Europe

Read more