1. Annual electricity is the wrong question

Most AI energy debates begin with terawatt-hours. A project developer faces a narrower constraint: can a specific campus obtain hundreds of megawatts of reliable capacity within 18 months? Is nearby transmission available? When will transformers and switchgear arrive? Can the system remain reliable during extreme weather?

NERC forecasts more than 224GW of additional North American summer peak demand over ten years, 69% above its prior forecast, with AI and digital-economy data centres accounting for most of the increase. The IEA estimates grid constraints could delay around 20% of global data-centre capacity planned through 2030. The world may have enough annual energy while lacking power at the right place and time.

Interconnection is not a ticketing system. A large load can require substations, transmission and generation. The IEA notes that new transmission often takes four to eight years in advanced economies, while lead times for transformers and cables have doubled in three years. Site selection is becoming energy engineering.

2. FERC is rewriting large-load access

In June 2026, FERC directed all six US regional grid operators to justify or change their large-load tariffs within 60 days. The action targets how projects enter queues, who pays, how flexible loads can operate, and whether generation and demand can be studied together.

Projects that disappear should still pay

Cost Recovery Agreements are intended to keep households from paying for infrastructure built for a data centre that never arrives. Escalating readiness requirements also target speculative projects that reserve capacity in several regions and distort forecasts.

Flexible load can connect faster

Non-firm transmission lets data centres limit grid withdrawals under constrained conditions. They exchange operating flexibility for faster access and fewer network upgrades. Schedulable workloads, storage and cross-campus task migration may therefore become interconnection assets.

Bring Your Own New Generation

BYONG is the structural change. A new generator and large load do not need to share a fence. If they are electrically proximate and studied together, they may require fewer upgrades. AI campuses are beginning to arrive with turbines, fuel cells, storage or other supply instead of only requesting utility capacity.

3. Value is migrating toward time-to-power

When a GPU's economic life is shorter than a transmission build, energising one year earlier has substantial value. The winners may be less about who generates the most electricity and more about who shortens deployment, offers credible delivery dates and accepts system responsibility.

Company disclosures show order and backlog signals, not future market share or returns. This is a research watchlist, not a security recommendation.

4. The counter-case matters

• Forecasts may double count demand: one project can apply in several regions; announced gigawatts are not energised capacity.
• Flexibility may beat new generation: curtailment, workload migration and storage can avoid part of the buildout.
• Onsite power is not automatically fast: gas supply, emissions permits, community opposition and equipment lead times can become new bottlenecks.
• Efficiency changes capacity needs: inference optimisation, better chips and higher utilisation can reduce incremental power per unit of AI output.
• Backlog is not revenue: slot reservations can be cancelled and manufacturing expansion can normalise supply.

5. Milestones that matter next

• How six RTOs and ISOs answer FERC, and whether cost recovery, flexible service and BYONG become tariff rules.
• Whether disclosures move from announced GW to energised MW, and whether cancellation rates rise.
• Whether turbine, transformer and switchgear lead times decline as new capacity enters service.
• Whether data centres sign interruptible-load contracts and actually reduce withdrawals during stressed periods.
• The fuel cost, permits, availability and lifetime economics of onsite generation.

Confidence in the direction is medium-high: interconnection has moved from a back-office engineering issue to a core AI deployment constraint, and both regulation and orders are responding. Confidence in specific winners is lower because double counting, technology choices and capacity cycles can still change value allocation.

Sources and evidence tiers

• T1FERC — Large Load Integration Orders, 18 Jun 2026Federal regulator; scope, regions and tariff reform.
• T1FERC — Cost recovery, flexible service and BYONGFirst-party regulatory explanation and intent.
• T1IEA — Energy and AI: AI and energy securityIndependent public institution; delay risk and equipment supply chain.
• T1IEA — Electricity 2026, Executive SummaryQueues, grid investment and flexible connection.
• T1NERC — 2025 Long-Term Reliability AssessmentNorth American peak-demand and data-centre forecasts.
• T2GE Vernova — Q2 2026 resultsCompany-reported orders and capacity.
• T2Eaton — Q1 2026 resultsCompany-reported orders, backlog and demand driver.
• T2Bloom Energy — Oracle partnership, Apr 2026Contracted capacity is evidence; deployment speed remains a vendor case.
• T2Quanta Services — FY2025 operational and financial summaryCompany-reported electric backlog; no inference about data-centre share.