The Sign-Off Layer: Where AI Is Landing in Energy Infrastructure, 2026

01The pressure

The IEA's 2026 update puts the growth of data-centre electricity demand at 17% in 2025, with AI-focused data centres growing 50%. It projects consumption roughly doubling from 485 TWh in 2025 to about 950 TWh in 2030, around 3% of global demand. The largest technology companies' capital expenditure exceeded USD 400 billion in 2025 and is expected to rise a further 75% in 2026 ^[1].

Grid operators see that demand as interconnection requests. PJM expects demand to grow by up to 70 GW by 2038, driven largely by data centres ^[2]. In Texas, large-load interconnection requests reached about 474 GW in August, “more than five times” ERCOT's record peak and about 90% data centres; the governor ordered an audit of every data centre in the queue and ERCOT delayed its first large-load study, while cautioning that its own forecast may be inflated. New York had already paused new data-centre approvals for up to a year ^[3]. In October, NERC's first reliability standards for computational loads passed their final ballot, covering sites of 50 MW or more connected at 100 kV and above, after a July 2024 event in which roughly 1,500 MW of data-centre load dropped off the grid at once ^[4].

Each of these responses asks for evidence: verify the load, verify the site, verify the model of the equipment. The new supply that has to meet that load faces the same requirement over a longer timeline.

02The bottleneck is paper and steel

Berkeley Lab's 2026 count found 2,061 GW of generation and storage actively seeking interconnection in the United States at the end of 2025, down 10% from the year before and from a peak of nearly 2,600 GW at the end of 2023. Gas in the queues rose 86% to 253 GW, while solar (773 GW), storage (749 GW) and wind (220 GW) all fell ^[5]. Of capacity that requested interconnection from 2000 to 2020, just 13% had come online by the end of 2025 and 75% had withdrawn. More than 40% withdraws even after signing an interconnection agreement. The median request took well over three years to reach an agreement and more than five to reach commercial operation ^[5].

Delays continue after the agreement. In the first quarter of 2026, more than 6.4 GW of US clean power expected to come online was delayed, adding to a backlog of 53 GW of delayed projects; developers attribute the delays to lengthy permitting, backlogged queues and fluctuating equipment prices ^[6]. Plants that are built start behind: across 373 GW of US solar, the average loss at commissioning was 4.46%, loss rose 55% between the second and third years of operation as EPC warranties lapsed, and the average technician now covers 70% more megawatts than five years ago ^[7].

Almost every step on that path is a document: an application, a site-control agreement, a title commitment, a study, a permit, a schedule, a commissioning record, a work order. AI started to land on those documents.

03Where AI landed in 2026

Interconnection intake

PJM's first reformed interconnection cycle drew 811 applications; 715 projects totalling 201.5 GW qualified, after a process that now requires up-front financial commitments and proof of site control ^[2]. Tapestry's HyperQ handled the site-control review. It turned 4,581 raw site-control documents into 2,328 agreement bundles and ran 9,312 compliance checks covering contract term, chain-of-title continuity, exclusivity and minimum acreage, with a median of 6 minutes 15 seconds per check, and returned page-level citations to the PJM engineers who made the final determinations ^[8]. Integration took ten months and 234 historical applications.

Two caveats come with it. The claimed 20-fold speed-up uses what Tapestry calls a conservative, anecdotal baseline of four hours per application ^[8], and PJM says it will evaluate the tool's impact throughout the cycle ^[2]. Neither caveat changes the division of work: HyperQ reads and cites, and PJM engineers decide.

Grid studies

Grid Agents, published in November, runs power-system studies in its own engineering environment, relies on established solvers for every number it reports, and leaves decisions to engineers; its authors verify each answer across 519 tasks and the work behind it, by replaying the agent's tool calls ^[9]. The US Department of Energy's $11.5 million GridFM 2.0 project, under the Genesis Mission, aims to let utilities evaluate a billion grid scenarios in 24 hours, with two utility deployments ^[10].

The counterweight arrived the same autumn. Across 15 grids of up to 70,000 buses, a sufficiently parallel CPU solver matched or beat the throughput of the learned power-flow surrogates tested while staying orders of magnitude more accurate; the authors found “no evidence” that these surrogates are yet competitive for power flow ^[11]. A grid foundation model distilled to a 32.5× smaller student for edge inference became more accurate than its teacher on some quantities, but its physics-residual term was consistently harmful ^[12].

Permitting

PNNL and OpenAI's DraftNEPABench tests AI agents on 102 real drafting tasks taken from environmental impact statements across 19 agencies. The drafts are grounded in authoritative sources, “with pointers to the original documentation for expert review”, and the researchers state plainly that the systems “still require human oversight” ^[13].

Operations and maintenance

The IEA, drawing on a survey of 25 network operators, finds forecasting, maintenance, inspection and planning scaling first, because they improve decisions “without directly controlling the power system” ^[14]. In solar inspection, a vision-language model reached 93% accuracy on 1,927 real modules across eight defect categories while writing diagnostic rationales; in a blind concordance study, a certified inspector's assessments aligned closely with them ^[15]; a ground robot now ties each defect under the array to a specific asset and its 3D position ^[16]. Raptor Maps' data associates more frequent inspection with lower loss: sites inspected quarterly had 36% lower power loss than sites inspected once a year ^[7].

The control room

A September 2026 study put a language-model agent on a digital twin of the Greek transmission network, behind a governance layer it could not bypass: only whitelisted tools run, no run exceeds its step budget, nothing with side effects executes without operator approval, and every number in an answer is rendered from backend results ^[17]. Across 590 runs the agent chose the right tool 96.5% of the time and completed 93.7% of tasks, and all four rules held. Removed from the layer, the same model executed all 45 approval-requiring runs without authorisation, and only 39.2% of its answers carried numbers backed by data. Enforcement cost 12 to 16 milliseconds per request.

General models are not yet reliable on this kind of evidence without such structure. On PowerBench, 300 questions that require retrieving and reasoning across telemetry and 24,939 operational documents, the best frontier model reached 74.2% joint accuracy ^[18].

04The pattern: cited evidence, human sign-off

The IEA states the pattern directly. AI adoption in grids is advancing unevenly, “and the reason is not model performance”. Real-time control is “constrained by trust and auditability rather than accuracy”: models must also be explainable, auditable, secure against manipulation, bounded by deterministic fallback rules and embedded in clear chains of accountability ^[14].

Every deployment above that scaled shares three properties. Its output is evidence: page-level citations ^[8], pointers to sources ^[13], numbers from established solvers ^[9]. A named person decides ^[8,9,13]. And the work can be replayed: tool calls re-run ^[9], an audit trail kept for every run ^[17].

Regulation is converging on the same shape. The EU's Digital Omnibus on AI entered into force on July 27, 2026, moving the deadline for high-risk systems under Annex III, which covers critical infrastructure, to December 2, 2027, when conformity assessments, technical documentation and registration fall due ^[19]. In the UK, Lucy Yu's review of AI in electricity networks recommends a framework defining autonomy in the grid by the first half of 2027, a use case for moving AI up the “delegation levels”, and probabilistic, risk-based operation by 2035 ^[20]. NERC's new standards require computational-load sites to re-verify their models within 180 days of changes to hardware, software, firmware or settings ^[4].

05What still holds adoption back

The IEA's survey of energy companies names the lack of digital skills as the single largest barrier to AI adoption, alongside fragmented data and concerns over data protection and cybersecurity; only 10% of global electricity consumption is covered by open electricity-data policies ^[1]. In grids, it finds that scaling depends “less on the AI model than on whether the network is observable, data are reliable and outputs can be integrated” ^[14].

Industrial operators report the same. In Takepoint Research's 2026 survey, 87.7% of industrial organisations were using or planning AI for OT security, but only 7.9% had deployed it across multiple functions. Data quality (45.4%) and legacy integration (42.4%) were the main obstacles, only 15.6% had OT-specific AI governance policies, and some compliance mandates prohibit OT systems from connecting to the cloud at all ^[21].

06What this means for owners, developers and operators

1. 01Start where a person already signs off. Interconnection intake, site-control and title review, permitting drafts, inspection triage and schedule reviews each have an accountable reviewer and a document trail. In 2026, AI scaled at stages with both.
2. 02Require citations. Ask for the page and region behind every claim, and for each claim to be checked against it. Page-level citations let PJM's engineers verify HyperQ's work at the source ^[8].
3. 03Measure against an honest baseline. Tapestry's speed-up rests on an anecdotal baseline by its own account ^[8], and a parallel CPU solver erased the speed advantage of learned power-flow surrogates ^[11].
4. 04Keep the data where it is allowed to be. Some OT rules forbid cloud connections ^[21]. Models that run on premises or at the edge, such as a grid model distilled to 32.5× smaller ^[12], fit that constraint.
5. 05Log the decision and the outcome. The EU's high-risk obligations for critical infrastructure arrive on December 2, 2027 ^[19]. A record of what the system proposed, who decided and what happened next is both the compliance file and the training data.

07Where Lumerian is building

We build for the stretches of the lifecycle where the evidence is heaviest and a person signs off. Prism reads title, survey and site-control evidence of the kind PJM now gates its queue on ^[2], with engagements on US solar projects in Virginia, Ohio and Texas; code finds every claim's quote on the cited page. Helm, open to its first design partner, targets the milestones behind a 53 GW backlog of delayed projects ^[6]: it records every date change with its reason code, and slip prediction switches on only after one real portfolio has run two quarters and beaten two baselines. Aperture addresses operating loss, which across 373 GW of US solar averages 4.46% at commissioning and rises once warranties lapse ^[7]; it runs on six utility-scale solar parks in Hungary, with 82,368 radiometric thermal frames from seven campaigns between 2023 and 2025. The rule across all three is the one this report found in every deployment that scaled: every answer carries its evidence, and a person makes the decision. Luminol, the model layer, ships each model only past its gate, and Lumerian Core runs hosted or on premises, so the evidence stays where operators need it.

The supply that will meet this load is still paperwork today: 2,061 GW of generation and storage was seeking interconnection at the end of 2025 ^[5].