AnalysisEnergy & Power

The Grid Is the Bottleneck

AI's demand for power is colliding with interconnection queues and transmission limits. Five decisions energy executives should make in the next 18 months.

Arclight Research5 min read

Fig. I · Energy & Power

In brief

  1. IData centers are built on a two-year clock and transmission on a decade-long one. That mismatch is the central planning fact.
  2. IIUnderwrite large-load requests like credit. Some are speculative or duplicated, and tariffs should price that risk explicitly.
  3. IIIPoint AI first at the decisions that drive capex: probabilistic load forecasts, interconnection studies and outage prediction.

For most of the past two decades, electricity demand in the United States was broadly flat. Efficiency gains offset growth, and the hard problems for utilities were plant retirements, renewables integration and rate cases. That planning assumption no longer holds.

Data centers, industrial reshoring and electrification are all adding load, and AI campuses are the most concentrated part of it: single sites asking for hundreds of megawatts on timelines the grid was not built to meet. This analysis looks at both sides of the collision, and at what energy executives should decide while the rules for this new load are still being written.

Two clocks, one constraint

A hyperscale campus can go from land purchase to energization in a couple of years if power is available. A new high-voltage transmission line commonly takes the better part of a decade once routing, permitting and cost allocation are counted. Generation interconnection queues are measured in years, and lead times for large power transformers, high-voltage breakers and new gas turbines have stretched in the same direction.

That mismatch is the central strategic fact. Developers who cannot get grid power on their timeline are turning to on-site generation, co-location at existing plants and temporary bridge supply, each of which raises its own reliability, cost-allocation and regulatory questions. Every serious AI strategy now runs through a substation.

The load is new in kind as well as in size. Reliability bodies have begun studying how large computing loads behave during grid disturbances, including events in which many of them drop off the system at once. A planning model calibrated on decades of residential, commercial and industrial load shapes has nothing in its history that looks like this.

The large-load book is a credit portfolio

Utilities are reporting large-load request pipelines that, on paper, rival their existing peak demand. Some of those requests are the same campus shopped to several utilities at once; others come from developers without a tenant, financing or equipment. Building transmission and generation to the face value of that book would shift stranded-asset risk onto every other customer.

The better posture is to underwrite large loads the way a lender underwrites borrowers. Utilities are already writing large-load tariffs with minimum-take provisions, longer contract terms, exit fees and collateral requirements. The executive decision is how hard to press on each lever, and how to separate committed load from optionality in the planning base case.

Every serious AI strategy now runs through a substation.

This is also a natural first use of AI inside the utility. Document intelligence can extract and compare terms across hundreds of load letters and service requests, and entity resolution can flag the same site or sponsor appearing under different names. Probability-weighted forecasts can then replace the single expected case that planning committees still tend to approve.

Putting AI to work inside the utility

The technology straining the grid can also shorten the work that constrains it. Inside utilities and independent power producers, the value concentrates in four areas, and each fails in a predictable way.

Load forecasting

Probabilistic time-series models that combine weather, metering data, rooftop solar and electrification trends give planners a distribution instead of a point estimate, which is what a capital plan exposed to uncertain large loads needs. Their weakness is regime change. A model trained on history has no basis for a 500-megawatt campus, so large loads must be layered in as explicit scenarios rather than learned from the past.

Grid optimization

Dynamic line ratings use weather and sensor data to rate conductors by actual conditions instead of conservative static assumptions, often releasing capacity from lines already in service. Optimizing topology and voltage can add more. The failure mode is operator trust: if the control room cannot see why a rating changed, operators will revert to the static number, and they will be right to.

Outage and wildfire risk

Models that combine weather forecasts, vegetation data from satellite imagery and LiDAR, and asset age can predict where a storm will cause damage, so crews are staged before it arrives. In wildfire country the same models inform public safety power shutoffs. That is precisely where false confidence is dangerous, and where every model needs a documented human decision-maker.

Interconnection studies

Much of the delay in interconnection is study work: building cases, running power-flow and contingency analyses, and reworking them when projects withdraw. FERC Order No. 2023 has already pushed the industry toward cluster studies with firmer deadlines, and grid operators have begun piloting AI and automation to speed case setup and screening. Machine learning can rank which projects are likely to trigger network upgrades, but it should not substitute for the engineers who stand behind the study result.

An 18-month agenda

Eighteen months is long enough to change a planning process and short enough to be held to account for it. We would put five decisions in front of the executive team.

  1. Set the large-load posture. Decide the tariff protections, flexibility commitments and collateral you will require, and how committed load will be separated from speculative requests in the plan of record.
  2. Rebuild the forecast as a range. Move resource planning and capital budgeting onto probabilistic load scenarios, and ask the board to approve a range with explicit triggers for each step up in investment.
  3. Fund two internal AI applications tied to capex. Interconnection study acceleration and outage prediction are strong candidates because their value appears in avoided cost and earlier revenue.
  4. Fix the data and its regulatory treatment together. Metering, SCADA and outage data need accountable owners, NERC CIP obligations shape what can run in the cloud, and cost recovery for AI software should be settled with regulators early.
  5. Choose a generation position. Independent power producers must weigh long-term contracts with large buyers against co-location and merchant exposure. Utilities must decide how much firm capacity to commit against an uncertain load book.

The fourth decision is easy to defer and expensive to ignore. Under cost-of-service regulation, a utility earns a return on capital but not on operating expense, and most AI arrives as software subscriptions and cloud compute. Unless the commission agrees to different treatment, the accounting quietly favors the substation over the model.

None of these decisions requires certainty about how much AI load will finally materialize. They require a planning process that is honest about the range, and a team that can move faster than the queue.

Questions for your next leadership meeting

  1. How much of our large-load pipeline is backed by signed agreements, financing and collateral, and what would we still build if the rest disappeared?
  2. Which of our capital decisions over the next three years rest on a single-point load forecast?
  3. Where do our engineers spend the most hours in the interconnection process, and what would halving that time be worth?
  4. If a model changed a line rating or a shutoff recommendation tomorrow, who would act on it, and could they explain why?
  5. Has our regulator agreed how AI software and cloud costs will be treated in rates, and if not, what is our filing strategy?

ARCLIGHT · RESEARCH

This analysis is general in nature and does not constitute investment, legal or engineering advice. To discuss how it applies to your organization, request a private briefing.

Begin

Apply this analysis to your own assets.

Tell us about your assets, your board and the decision in front of you. A principal replies within one business day.