Research · Power & Energy
Why hyperscale power is moving behind the meter, and why most announcements will not deliver on the AI schedule.
Behind-the-meter generation is no longer fringe. It is becoming the pressure-release valve for a grid that cannot always deliver large AI loads on datacenter construction schedules.
The takeaway
The winning architecture in 2026 is a portfolio: grid service where capacity is real, gas or fuel-cell bridge power where air and fuel constraints are manageable, batteries for ride-through, nuclear restarts or co-located existing nuclear where tariff treatment is explicit, and advanced nuclear or geothermal as a 2030s clean-firm option. The single biggest risk is not whether any one technology can produce electrons. It is whether generation, transmission deliverability, backup service, fuel supply, permits, public acceptance, and cost allocation can all be made to arrive on the same schedule as the AI load ramp.
01
Datacenter construction cycles are now colliding with power-system cycles.
The U.S. Energy Information Administration expects electricity use to grow 1% in 2026 and 3% in 2027, identifying large computing centers as a driver. EIA also identifies ERCOT and PJM as the highest-growth regions in its 2026 load-growth outlook. PJM reports that its interconnection reform reviewed more than 170 GW in the transition process, with 30 GW still to process in 2026 and a 1–2 year study-process target for new projects.
The market is responding rationally. Hyperscalers are trying to buy their way out of queue risk, transmission congestion, and capacity scarcity. But public-record precedents show that self-supply still becomes a public infrastructure question once the load wants backup service from the grid, relies on transmission infrastructure, emits locally, or seeks nuclear licensing.
FERC's rejection of the Susquehanna co-location amended interconnection service agreement is the doctrine-forming case: "In this order, we reject the Amended ISA." That language has reframed every subsequent co-location pitch in the sector.
02
Five things the public record makes clear about BTM in 2026.
01
BTM is a schedule hedge, not a replacement for utility planning
Onsite and co-located generation can shorten some parts of the energization path, especially when the alternative is waiting years for new transmission or resource adequacy. But BTM does not remove the need to solve backup service, fault duty, protection schemes, telemetry, reserve obligations, cost allocation, and local emissions. Northern Virginia illustrates the distinction: physical proximity to high-voltage lines is not the same thing as deliverable capacity, which is why Virginia's SCC has already moved to insulate ratepayers from infrastructure costs tied to large loads.
02
Gas and fuel cells are the near-term bridge, and the highest-reputation-risk technologies
Gas turbines, reciprocating engines, and fuel cells are the only generation classes that can plausibly be deployed on the one-to-three-year timeline demanded by many AI loads. EIA reports 6.3 GW of planned new U.S. natural-gas-fired capacity in 2026 and 18.7 GW of combined-cycle capacity planned by 2028. But gas is not politically invisible: EPA treats stationary combustion turbines and engines as regulated stationary sources, and the xAI Colossus / Memphis–Southaven controversy shows how fast "temporary" speed-to-power generation can become an environmental-justice and permitting flashpoint.
03
Existing nuclear beats new nuclear for this decade
The credible clean-firm nuclear pathway for the late 2020s is not a first-of-a-kind SMR. It is an existing nuclear site, an existing workforce, an existing transmission footprint, and a bankable offtaker. The Microsoft–Constellation Three Mile Island / Crane Clean Energy Center transaction is the clearest template, federal material describes the unit at 837 MW and expected to go online in 2028, with NRC approval needed to restore the operating licensing basis. That restart model is materially more plausible for this decade than first-of-a-kind advanced reactors.
04
Co-location does not escape FERC, PJM, or state-commission politics
The AWS–Talen Susquehanna case is the cautionary template. The commercial logic was obvious, connect a datacenter campus to a large existing nuclear station and contract for clean firm power. The regulatory problem was also obvious: if the load relies on grid-connected infrastructure, backup service, reserve capacity, or network facilities, other market participants and customers will ask who pays. Pennsylvania has since moved to a model large-load tariff framework. The direction of travel nationally: co-located load will be allowed where it is transparent, studied, reliable, and cost-caused, not as a private exemption from the grid.
05
FOAK economics remain the binding constraint for SMRs
The NuScale–UAMPS cancellation is the reference-class warning: projected VOYGR LCOE rose to $89/MWh before cancellation. DOE's advanced nuclear commercialization work frames new nuclear economics around overnight capital costs that can span $7,000/kW to $20,000/kW. For hyperscalers, that means SMRs can be strategically valuable as long-term clean-firm options, but they should not be underwriting near-term AI capacity ramps. A high-capex nuclear asset with a six-to-ten-year development path is extremely sensitive to schedule slip; in an illustrative 100 MW case at $8,000/kW capex, 90% capacity factor, $120/MWh offtake, and a 7.75% nominal discount rate, project value remains negative, and a two-year FOAK overrun pushes the economics materially worse.
03
The grid is no longer keeping pace with load announcements.
AI datacenter load is different from ordinary commercial load. It is large, flat, high-utilization, and schedule-driven. A 500 MW AI campus behaves more like a baseload industrial offtaker than a flexible office park. That creates four simultaneous stresses.
First, the interconnection queue is not designed around loads arriving faster than generation and transmission. PJM's reform effort is significant, but a one-to-two-year study turnaround for new generation projects still does not guarantee deliverable capacity for large load in constrained pockets.
Second, load growth is concentrated in the same regions that already host dense datacenter clusters. EIA identifies ERCOT and PJM as the highest-growth regions in its 2026 load-growth outlook. The public-policy fights in Virginia and Pennsylvania are not abstractions; they are the ratepayer and reliability response to concentrated hyperscale demand.
Third, datacenter construction can outrun grid construction. Turbines, substations, transformers, transmission lines, and gas laterals all have procurement and permitting constraints. The ability to build a shell and procure GPUs faster than a utility can expand transmission creates the incentive to self-supply. Fourth, annual renewable matching is no longer enough for the highest-value AI workloads, hyperscalers increasingly need firm, dispatchable, low-carbon power that matches hourly operations, not annual energy accounting.
Disclosed in the record
04
Ten generation classes, ranked by what they can actually deliver in 2026.
Gas turbines & reciprocating engines · $0.9–2.0M/MW · 1–3 yr
Best role: one-to-three-year bridge power, early load blocks, resilience, and hedging against grid delay. Most deployable class for large onsite generation where gas service, emissions, and local approvals are manageable. Main risk: air permitting and gas deliverability, a prime-power gas campus cannot rely on emergency-generator assumptions.
Fuel cells · $3.0–6.5M/MW · 1–2.5 yr
Best role: modular bridge power where turbines face public resistance or a lower local-emissions profile is valuable. Commercial but costly at hundreds of megawatts; AEP's up-to-1-GW Bloom Energy agreement is the strongest public hyperscale signal. Main risk: delivered cost and fuel exposure.
Existing light-water reactor restarts · $1.5–6.0M/MW · 3–7 yr
Best role: medium-term clean-firm supply for large hyperscale contracts. Credible where an existing nuclear asset, transmission interconnection, workforce, and safety case can be restored, Crane Clean Energy Center / Three Mile Island Unit 1 is the reference case. Main risk: latent restart scope; licensing-basis restoration, workforce, equipment condition, outage readiness, and market/tariff arrangements can each move the schedule.
NuScale & light-water SMRs · $9–20M/MW · 8–15+ yr
Best role: 2030s clean-firm option where licensing, orderbook, and financing are already advanced. NuScale has the strongest U.S. design-approval reference among SMR vendors, but design approval lowers one category of licensing risk. It does not provide site permission, construction execution certainty, or commercial cost certainty. Main risk: FOAK economics and customer attrition before commercial operation.
Holtec SMR-300 · $7–20M/MW · 7–15+ yr
Best role: long-dated clean-firm option, especially at existing nuclear sites. NRC has stated it is reviewing Holtec SMR application material with a detailed technical review schedule not to exceed 18 months once accepted for processing. That is an important process milestone, not a near-term supply guarantee. Main risk: vendor-specific commercial cost, licensing execution, and EPC wrap remain insufficiently public for investment-grade underwriting.
X-energy Xe-100 · $7–20M/MW · 7–15+ yr
Best role: early advanced-reactor platform with strategic hyperscale relevance. DOE reports NRC accepted Dow's construction-permit application for an X-energy project and describes Xe-100 as a high-temperature gas-cooled reactor using TRISO fuel. X-energy's Energy Northwest pathway is important because utility-region sponsorship improves credibility. Main risk: TRISO fuel readiness, licensing execution, and first-project construction productivity.
Kairos Hermes / KP-FHR · $8–22M/MW · 7–15+ yr
Best role: technology-demonstration path that may mature into 2030s commercial supply. NRC issued a construction permit for the Hermes test reactor in December 2023; DOE has described commercial KP-FHR deployment as a 2030s proposition. Main risk: translating a test-reactor licensing path into commercial datacenter-grade supply.
TerraPower Natrium · $7–20M/MW · 8–15+ yr
Best role: strategic advanced-nuclear proof point; long-term clean-firm supply with load-following value. NRC approved a construction permit for the Natrium project in Kemmerer, Wyoming, with safety review completed December 2025. Wyoming's PSC role also matters, large power facilities and transmission can require state certificates. Main risk: FOAK sodium fast-reactor execution, operating authorization, fuel readiness, and state certificate / cost-treatment issues.
BESS + renewable hybrids · $4–10M/MW-firm · 2–6 yr
Best role: ride-through, peak shaving, emissions strategy, limited self-supply, demand response, and grid-support value. Highly deployable as a component. But a solar-plus-storage campus is not automatically firm 24/7 AI power, the amount of overbuild and storage required to serve flat load across seasons can make the apparent low renewable LCOE misleading. Main risk: confusing cheap non-firm energy with datacenter-grade firm supply.
Geothermal & enhanced geothermal · $4–12M/MW · 4–10 yr
Best role: clean-firm 24/7 power where geology is proven. DOE states geothermal capacity factor is generally around 90% and can operate steadily around the clock. Enhanced geothermal can expand the resource base, but drilling and subsurface performance risk remain central. Main risk: subsurface uncertainty and development learning curve.
05
Six deals that reveal how BTM actually plays out in public.
Microsoft–Constellation: Three Mile Island / Crane Clean Energy Center
A nuclear restart-backed clean-firm supply strategy, not a generic SMR deployment and not an ordinary grid PPA. Public federal material identifies the plant at 837 MW with an expected 2028 online date. NRC approval is needed to restore the operating licensing basis; restart oversight resembles the Palisades model. Read-through: this is the best 2020s clean-firm template because it starts from an existing asset. It is still a regulated nuclear restart, not a procurement shortcut.
AWS–Talen: Susquehanna co-location
The doctrine-forming co-located nuclear datacenter case. Public FERC materials identify the PJM Susquehanna co-location proposal and amended ISA; public analysis has described a disputed expansion from 300 MW to 480 MW, with broader market descriptions referencing up to 960 MW of nuclear-backed capacity. FERC rejected the amended ISA. The issues are not just wires, they are open access, backup service, transmission service, network impacts, reserve obligations, and who pays when the generator is unavailable. Read-through: physical co-location helps the engineering case; it does not settle the legal or political case.
Oracle's SMR announcement
A market signal that AI campuses are looking beyond conventional grid supply. Public-company detail sufficient for commitment-grade underwriting was not available in the reviewed record. Any SMR-backed supply would still need vendor-specific NRC licensing, site approval, construction, operating authorization, fuel supply, state/local approvals, and grid/offtake arrangements. Read-through: useful for demand-side market direction; weak as a supply-side proof point until tied to a named reactor, site, license path, financing plan, and schedule.
Meta nuclear RFP
A large clean-firm procurement signal. Deal-specific terms were not publicly sufficient for hard capacity, pricing, or COD claims. The regulatory pathway depends entirely on the winning technology and site, existing nuclear PPA, restart, SMR, advanced reactor, or grid-connected clean-firm portfolio. Read-through: Meta's procurement direction reinforces the move from annual renewable matching to firm clean power. It does not resolve which technologies can scale on the AI timeline.
Google–Kairos
An advanced-nuclear development pathway tied to Kairos's technology maturation. Hermes is a test-reactor precedent, not a commercial datacenter supply license, commercial deployment would require additional design/site licensing, NEPA review, construction, operating authorization, and fuel readiness. Read-through: strategically important; not a 2026–2029 energization solution.
X-energy–Amazon / Energy Northwest
An SMR-backed pathway with a public-power / regional-energy anchor. DOE describes an X-energy four-unit, 320 MWe demonstration with Energy Northwest and Burns & McDonnell as partners. Read-through: one of the more credible SMR pathways because it combines vendor, utility-region sponsor, and federal-program context, but it still needs bridge power for near-term AI load.
06
The BTM era introduces four harder questions for every named deal.
01
Who pays for backup reliability?
If the datacenter self-supplies but still needs the grid during outages, refueling, weather events, or generator failures, other customers will demand cost-causation proof. The acceptance test is not whether the company can buy power. It is whether other customers are protected from backup, transmission, reserve, and stranded-cost exposure.
02
Who bears local emissions and noise?
Onsite gas generation concentrates impacts near communities, even when framed as temporary. The xAI Memphis/Southaven case is the canonical reference: a U.S. Senate Environment and Public Works release describes operation of dozens of unpermitted gas turbines in vulnerable communities, with NAACP and other community groups opposing air, water, and electricity impacts.
03
Who controls scarce infrastructure?
Nuclear restarts, transmission lines, substations, and clean-firm supply can be politically sensitive if they appear dedicated to private AI load while ordinary customers face higher bills. The Pennsylvania PUC's large-load tariff framework and Virginia's SCC ratepayer-protection orders are the public-policy responses to that perception.
04
Who benefits from incentives?
Tax abatements, federal credits, LPO financing, and utility cost recovery can intensify opposition if permanent jobs are limited. Virginia shows the mature-opposition path; Pennsylvania shows the tariff path; Memphis/Southaven shows the environmental-justice path; Wyoming shows a more constructive path, advanced nuclear can be accepted more readily where it is tied to coal-transition employment and an existing energy community, but schedule slippage can still erode trust.
07
When does BTM actually scale? Three windows.
01
2026–2028: bridge power and tariff sorting
The near-term market will be dominated by gas turbines, reciprocating engines, fuel cells, BESS, grid PPAs, and utility special contracts. The best projects will not claim to be off-grid. They will specify backup service, emissions controls, fuel contracts, telemetry, dispatch limits, interconnection responsibilities, and decommissioning or conversion pathways.
02
2028–2032: nuclear restarts, co-location, and firm clean portfolios
Existing nuclear restarts and co-located nuclear PPAs can matter materially if FERC, ISO/RTO, and state tariff rules stabilize. The Microsoft–Constellation / Crane model is the most credible clean-firm 2020s template. The AWS–Talen / Susquehanna case is the warning that commercial adjacency is not regulatory insulation.
03
2032 and beyond: advanced nuclear and geothermal optionality
SMRs, non-light-water reactors, and enhanced geothermal can become important if early projects demonstrate repeatable cost, licensing duration, fuel supply, construction productivity, and host-community acceptance. The strongest hyperscalers will treat those technologies as portfolio options now, not as substitutes for 2026–2030 energization plans.
08
Five recommendations for hyperscalers, utilities, ISOs/RTOs, regulators, and local governments.
01
Hyperscalers: publish cost-causation principles before announcing BTM deals
The acceptance test is not whether the company can buy power. It is whether other customers are protected from backup, transmission, reserve, and stranded-cost exposure. Stating the principle publicly before the deal disarms the most predictable counter-argument.
02
Utilities: create large-load service products that reward flexibility and direct cost responsibility
Minimum bills, take-or-pay commitments, collateral, interruptible service, onsite-generation telemetry, and non-firm service options can reduce system risk while still allowing growth. The goal is to make hyperscale load an asset to the system rather than a transfer payment from other customers.
03
ISOs and RTOs: standardize co-located-load modeling
The sector needs transparent rules for netting, backup service, generator outages, telemetry, operating limits, and network cost allocation. Without those rules, every co-located deal becomes a bespoke FERC case, and the result is the Susquehanna precedent applied unevenly across regions.
04
Regulators: require realism in public announcements
A nuclear design approval, construction permit, MOU, or RFP is not equivalent to delivered capacity. Public agencies should force a clean distinction between commercial aspiration, permitted construction, and authorized operation, otherwise announcement inflation builds expectations that the engineering and licensing record cannot meet.
05
Local governments: treat onsite generation as infrastructure, not an accessory use
Gas turbines, fuel cells, BESS yards, substations, cooling systems, and gas laterals carry impacts that should be disclosed, conditioned, and monitored. Treating a 500 MW onsite gas plant as a backup-generator accessory to a 'data processing use' is the local-permitting analogue of the AUAR mismatch.
In closing
Behind-the-meter generation is no longer fringe. It is becoming the pressure-release valve for a grid that cannot always deliver large AI loads on datacenter construction schedules. But BTM will scale only where sponsors are honest about its dependencies. Gas and fuel cells buy time. Batteries buy flexibility. Grid PPAs and utility service remain the baseline where capacity is real. Existing nuclear can buy clean-firm credibility. SMRs and geothermal buy long-term optionality.
The projects that work will look less like clever private bypasses and more like integrated infrastructure programs: transparent tariffs, directly assigned costs, credible fuel and permitting plans, enforceable community protections, and finance structures that survive delay. The projects that fail will be the ones that confuse a power announcement with a power plant.
Full brief · 19 pages · PDF
Read the complete brief
Standing note
Independent analysis based solely on publicly available federal, state, and utility regulatory filings, agency releases, court records, and industry publications. For educational and discussion purposes only. Does not constitute investment, legal, or engineering advice and does not represent any of the named hyperscalers, utilities, technology vendors, or regulatory bodies discussed.