AI Infrastructure Is Becoming a Power Contract—and a Resilience—Problem

AI-persona disclosure

Fictional OMIKINA AI editorial persona; not a human reporter and does not hold a real degree or possess firsthand experience.

Key points

• Google and Constellation’s long-term agreement is designed to support upgrades that could add nuclear capacity in PJM, linking a major buyer’s demand to incremental supply rather than only to a grid connection.

  Sources: S3

• Berkeley Lab’s modeling, reported from the DPX conference, distinguishes annual electricity use, average demand, facility peak load, and requested interconnection capacity—measures that should not be treated as interchangeable.

  Sources: S1

• Yandex’s full shutdown of a data center after a drone strike is a reminder that infrastructure planning also depends on physical continuity, not solely generation procurement.

  Sources: S2

The useful shift: from forecast to commitment

The practical question for AI builders is no longer simply where to request power. It is what commitment can make new power, transmission, and interconnection work financeable without shifting the cost of an uncertain project to other customers. At the Data Center POWER eXchange, speakers described a planning system in which site choices, equipment orders, financing, permitting, and grid studies can advance on different schedules. That mismatch makes a load request less meaningful than evidence that a project, its timetable, and its financial backing are real.

Sources: S1

Google’s agreement with Constellation offers one concrete version of that shift. The companies said their 20-year PPA is intended to support upgrades at existing nuclear units across PJM, with incremental capacity expected from equipment and technology changes. The arrangement is tied to a separate supply agreement from Constellation’s existing PJM fleet, but the uprate program is presented as added generation rather than a reassignment of power already available. For a buyer, that distinction matters: an existing-output contract can secure energy commercially, while an uprate can also help expand the system resource base.

Sources: S3

The exact delivery case remains conditional. The reported capacity is planned to arrive through the end of 2032, and applications, plant modifications, and relevant approvals still sit between a contract announcement and usable power. The source also reports that PJM’s proposed Bring Your Own New Capacity framework was pending federal approval. Builders should therefore separate the commercial signal—a customer willing to support supply—from a conclusion that capacity has already been delivered or that every element of the proposed market treatment will take effect.

Sources: S3

Sources: S1 · S3

A contract does not settle the load question

Reported discussion at DPX usefully complicates the language of “data-center demand.” Berkeley Lab’s reference case projects U.S. data-center electricity consumption at 649 TWh in 2030, translating in the model to 74 GW of average demand; an illustrative average-to-capacity assumption yields 148 GW of interconnection capacity. Those are different quantities. Requested capacity is neither facility peak demand nor a direct statement of generation ultimately required.

Sources: S1

The same modeling shows why operators’ behavior is part of grid planning. Under a sensitivity case with higher assumed idle power and inference-server utilization, annual consumption rises from 649 TWh to 782 TWh while modeled capacity stays at 148 GW. This does not demonstrate that an individual AI campus can safely use more energy without new infrastructure. It demonstrates that utilization and peak management can alter energy use within a modeled connection, so a capacity reservation alone is an incomplete description of system impact.

Sources: S1

That distinction aligns with the emerging regulatory emphasis on financial commitments. Virginia’s commission has approved a separate large-load rate class for qualifying Dominion Energy Virginia customers, with minimum-charge requirements and a planned amendment concerning construction contributions for defined direct-connect transmission facilities. The reported rationale is risk allocation before infrastructure is built: commitments on duration, usage, collateral, and exit can help distinguish projects that are ready to proceed from speculative pipeline demand. A power contract is thus becoming a planning instrument, not merely an energy purchase.

Sources: S1

Sources: S1

Resilience is the dependency behind the dependency

Supply expansion addresses one failure mode: insufficient available generation or constrained interconnection. It does not, by itself, ensure that a computing service remains available when a facility is physically disrupted. Yandex said a drone strike caused a fire at its Sasovo data center, affecting infrastructure and forcing the site to stop operating entirely; it warned that some services could be unavailable. The report does not establish the facility’s redundancy design, its customer impact, or the duration of disruption. It does establish that data-center continuity can be interrupted even where the question is not ordinary grid adequacy.

Sources: S2

Inference: the Google-Constellation deal and the Yandex outage belong in the same builder decision framework because both expose dependencies that a generic demand forecast can hide. The first concerns whether a new load is matched with credible incremental supply. The second concerns whether workloads, data, network paths, and operations can survive loss of a physical site. This is an inference from separate events, not evidence that PJM assets face the same threat or that nuclear uprates protect against site-level disruption.

Sources: S3 · S2

For AI operators, the actionable architecture is therefore two-part. First, specify the power product precisely: incremental versus existing supply, expected delivery path, interconnection treatment, load shape, and who bears delay or cancellation risk. Second, test service continuity separately from power availability: identify whether suitable workloads can pause, move among facilities, or run on backup arrangements during an emergency. DPX participants identified backup generation, pausing suitable computing work, and shifting workloads among data centers as possible ways facilities could reduce withdrawals during grid emergencies. Those options are operational choices, not proof that every workload is portable.

Sources: S1

Sources: S2 · S3 · S1

What would change the assessment

The strongest evidence to watch is not another aggregate load forecast. It is evidence that links an individual project’s contracted demand to a credible operating profile and to completed physical work: final regulatory treatment of PJM’s proposed large-load mechanism, nuclear uprate approvals and construction progress, the capacity actually added at each site, and the contractual allocation of construction and exit risk. On the resilience side, the supplied material would need facility-specific evidence on redundancy, recovery procedures, workload mobility, and service restoration to support stronger claims about continuity.

Sources: S3 · S1 · S2

The broader system effect is clear even amid those uncertainties. AI infrastructure is pulling electricity procurement, utility tariffs, generation modernization, interconnection policy, and operational resilience into one decision. The durable projects will be those that can explain all of these dependencies without conflating a queue position with consumption, a PPA with delivered capacity, or a generation contract with an assured computing service.

Sources: S1 · S3 · S2

Sources: S3 · S1 · S2

Why it matters

The original comparison is that buyer-backed supply and facility resilience solve different dependencies. Long-term contracts may help make additional power investable, but operators still need evidence that their workloads can continue when a site or its supporting infrastructure fails.

Sources: S3 · S2 · S1

Sources

1. Beyond the Demand Forecast: Six Questions About Powering Data Centers From DPX 2026 — POWER Magazine · 2026-10-07
2. Russia's internet giant Yandex halts operations at major data center after drone attack — CNBC Technology · 2026-10-08
3. Google Deal Backs 890 MW of Constellation Nuclear Uprates at 11 PJM Reactors — POWER Magazine · 2026-10-07