Grid Access Is Now the Binding Constraint on AI Infrastructure

The Constraint Has Shifted

For two years the conversation about AI infrastructure has centered on GPU availability, memory bandwidth, and networking fabric. Those constraints are easing. The one that is tightening is simpler and slower to fix: electrical power.

US data center power demand is forecast to more than double to 66 GW in 2027 from 31 GW in 2025, driven by an accelerating buildout of AI infrastructure, according to Goldman Sachs Commodities Research. This is not speculative — it rests on facility-level construction data. Year-over-year capacity additions are scheduled to reach 13.6 GW in 2026 and 36.3 GW in 2027, compared with realized additions of 6.4 GW in 2024 and 8.5 GW in 2025.

By 2030 the picture darkens. Data centers are projected to consume 9% to 17% of US electricity by 2030, up from 4% to 5% today, according to EPRI's "Powering Intelligence 2026" report — a range that is 60% higher than the organization's prior scenarios, reflecting the accelerated pace of development. These projections come from the Electric Power Research Institute and Goldman Sachs Commodities Research, tracking consistently with IEA and Lawrence Berkeley National Laboratory modeling.

For operators, the implication is sharp: power procurement is now a supply chain problem, not a utility services problem.

What FERC Actually Did

On June 18, 2026, FERC issued six tailored orders aimed at drastically accelerating grid interconnection for AI data centers and other large-load energy users, acting under Section 206 of the Federal Power Act. The targets are PJM, MISO, Southwest Power Pool, CAISO, ISO New England, and NYISO. The commission issued show-cause orders directing them to explain within 60 days why their existing tariffs remain just and reasonable or propose reforms.

The mechanism matters: rather than launching a Notice of Proposed Rulemaking, which typically takes years to finalize, FERC used Section 206 of the Federal Power Act to issue customized orders to each operator. Speed of process is not speed of power delivery.

Under the orders, each RTO/ISO and its transmission owners have 60 days to either justify why their current tariffs remain just and reasonable without provisions tailored to large loads, or to file tariff changes. Five categories of reform are teed up: efficient transmission service application and study processes; preventing cost shifting and requiring transparency into transmission costs; accommodating co-location agreements and behind-the-meter generation; providing new transmission services for flexible large loads; and developing a process to study generating facilities that serve electrically proximate large loads.

The commission also directed each grid operator and its transmission owners to submit a resource adequacy report within 30 days detailing how they plan to ensure sufficient generation is available to serve existing customers and new large loads. This second requirement may cut deeper. It forces grid operators to publicly state whether they can actually serve projected demand — and where they cannot.

One critical gap: the orders do not apply to the fast-growing data center hub in Texas, which operates a grid outside of federal jurisdiction. ERCOT is not in scope.

The Real Bottleneck Is Physics, Not Paperwork

FERC's orders address administrative friction, not transmission capacity or generation availability—which are the actual constraints.

Only about 50-60% of data center capacity scheduled for the next one to two years is expected to come online on time amid delays and cancellations. The miss rate stems primarily from power, not zoning or permitting. Interconnection queues now stretch years out. Developers trying to build data centers to support AI workloads face the same problem as renewable energy developers: a technically viable site, a willing operator, and no realistic path to getting power at the scale and speed they need.

Grid infrastructure and data centers operate on irreconcilable timelines. A hyperscale campus can be designed and constructed in 18 to 24 months under favorable conditions. Transmission lines take five to ten years from planning to energization, accounting for permitting, right-of-way acquisition, and construction. Nuclear generation takes longer; gas is only marginally faster when pipeline constraints apply.

Regional impacts will diverge sharply, with Mid-Atlantic, Mid-Continent, and Northwest markets facing elevated reliability risks, while the impact in Texas and Georgia may be relatively marginal thanks to plans for additional power generation. This divergence is the operative planning signal. The Mid-Atlantic—Northern Virginia specifically—is already running hot. Virginia, the nation's data center epicenter, sees facilities consuming over 25% of state electricity today; that could rise to 41-59% by 2030.

Within 45 days, RTOs can request 90-day extensions added to the initial 45-day period. That timeline may result in "relatively cursory reports, not detailed reform proposals." Regional stakeholders face having to rapidly diagnose deficiencies and identify reforms for issues that often take multiple years to develop remedies. The 60-day clock may produce paperwork that delays the real reckoning rather than actionable reform.

What Operators Should Do Now

Treat grid interconnection as a first-order site selection variable, not a downstream infrastructure task.

Developers are shifting to locations offering abundant power availability, land, faster permitting, and incentives — moving beyond traditional hubs like Northern Virginia. By 2030, seven additional states — Oregon, Iowa, Nebraska, Nevada, Wyoming, Arizona, and Indiana — could see data centers exceeding a 20% share in the medium scenario. These lack the AI infrastructure credentials of legacy hubs, but they have generation surplus. That surplus is the asset.

"Bring your own generation" trends are emerging, with onsite power — including natural gas, renewables, or storage — helping mitigate grid strain. Co-located generation, once a niche approach, is becoming competitive necessity. Utilities, power producers, and hyperscalers are already pairing data centers with nearby gas, nuclear, renewable, and hybrid generation projects. FERC's focus on electrically proximate generation suggests regulators expect those arrangements to play a larger role as AI infrastructure expands.

Accelerated servers equipped with GPUs, which handle AI training and inference workloads, now account for around 60% of electricity consumption in modern data centers on average, according to IEA analysis. That share will grow. Adding a new GPU cluster without a corresponding power plan is a stranded asset risk.

What to Watch

1. August 2026: The 60-day FERC response deadline. Watch whether RTOs file substantive tariff reform or request extensions. Extensions signal the procedural fast-track is stalling on real constraints.

2. Resource adequacy reports (30-day deadline, mid-July 2026): These will be the first public, operator-level acknowledgment of which grid regions cannot support projected AI demand. Read them as a siting guide.

3. Co-generation deal flow: Track whether major hyperscalers accelerate direct investment in generation assets — gas peakers, nuclear offtake agreements, solar-plus-storage — rather than waiting for utility interconnection queues. Microsoft's West Texas arrangement with Chevron is the model.

4. Regional power price divergence: As constrained markets tighten, watch electricity price spreads between Mid-Atlantic markets and generation-surplus regions like MISO's upper Midwest. Price signal precedes capacity migration.

5. On-time delivery rate: If 2026 and 2027 capacity additions track below 50%, power constraint is definitively the primary AI infrastructure bottleneck — and repricing of grid-adjacent land and generation assets accelerates.