AI Power Risk Scoring for Data Center Sites: How Developers Are Quantifying the Primary Constraint
Grid availability is now the binding constraint for AI data center development. Here is how developers are using AI to score it before committing to a site.
Power used to be a box to check in data center site selection. Is there a substation nearby? Has the utility confirmed capacity? Those questions still matter, but in 2026 they are insufficient. Bloom Energy's 2026 Data Center Power Report is direct on this: grid congestion and interconnection delays are now the binding constraint for AI data center growth, not equipment supply or capital availability.
That shift has produced a new category of underwriting work: systematic AI-assisted power risk scoring that treats grid access as a probabilistic variable, not a binary condition. The developers and infrastructure funds doing this well are making better site decisions faster. Those still relying on informal utility calls are taking risks they cannot quantify.
What Power Risk Scoring Replaces
The previous approach to power diligence looked roughly like this:
Check substation proximity on a GIS map
Call the utility to ask about available capacity
Request a preliminary load study
Move forward if the utility seemed positive
This works when data centers demand 5 to 20 MW. It breaks down at 50, 100, or 300 MW, and it breaks down completely when the interconnection queue in your target market runs two to three years deep.
With JLL projecting nearly 100 GW of new data center capacity globally from 2026 to 2030, the pressure on grid infrastructure is structural. Developers who treat power availability as a spreadsheet cell rather than a distribution of outcomes are building models on bad assumptions.
The Five Risk Dimensions AI Models Cover
Effective AI power risk scoring decomposes grid access into five distinct risk factors rather than producing a single composite score. Each factor requires different data sources and different model types.
1. Power availability: will the MW actually be there?
This is the core question. Given the existing interconnection queue, planned utility upgrades, and competing load requests, what is the probability that the site can receive the target MW by the target energization date?
AI models for this dimension ingest interconnection queue data from the relevant ISO or RTO (PJM, MISO, ERCOT, CAISO, NYISO, ISO-NE, SPP), historical completion and withdrawal rates for that utility and voltage level, and planned transmission and substation upgrades from utility integrated resource plans. Large language models are increasingly used to extract the relevant capacity commitments and upgrade scopes from unstructured IRP documents and FERC filings.
The output is a probability distribution over energization dates, not a single "expected" date. A site with a one-in-four chance of a 24-month delay is a fundamentally different underwriting proposition than one with a 90 percent probability of on-time energization, even if both have substations nearby.
2. Grid reliability: what happens after energization?
Once the power is delivered, how often will it be interrupted, and how severely? This matters for both the development underwriting (backup power sizing) and for the tenant negotiation (SLA commitments).
AI reliability models combine historical SAIDI and SAIFI data from utility reliability filings, outage cause codes, asset age, vegetation management records, and climate hazard layers. Swiss Re's 2026 analysis of AI data center risks notes that single large campuses can exceed $20 billion in total exposed value, and that climate-driven perils, particularly hail, tornadoes, and heat waves that simultaneously raise cooling loads and grid stress, require explicit probabilistic modeling.
The output feeds backup power sizing decisions (generator capacity, fuel storage, UPS runtime) and informs the lease terms around facility SLA commitments.
3. Interconnection queue risk: how exposed is the queue position?
Not all queue positions are equal. A data center requesting 80 MW at a substation with 400 MW of projects ahead of it in various stages of study is in a different position from one requesting the same capacity at an undersized queue with mostly smaller projects likely to withdraw.
AI systems trained on historical queue data by ISO can calculate:
Expected time-to-interconnection given queue position, project mix, and historical attrition rates for that market
Probability that queued capacity ahead of the developer will actually be built versus withdrawn (solar and wind projects in particular have high withdrawal rates in congested markets)
Estimated upgrade cost exposure based on queued projects at the same substation, which affects both timing and cost allocation
This scoring dimension directly informs whether a site is worth pursuing formal utility engagement or whether it should be filtered out earlier.
4. Power price and regulatory risk: what will power actually cost, and can the rules change?
Even if power is available and reliable, unexpected cost changes can materially affect facility economics on 15-year leases. Two risks matter:
Tariff and cost-allocation risk: Several states, including Oklahoma, Virginia, and North Carolina, have moved toward requiring data center developers to fund their own grid upgrade costs rather than socializing them across ratepayers. Tracking active PUC dockets and legislative activity on data center cost allocation is now a standard part of site diligence. LLMs can parse these filings and flag changes that would affect a specific site's power costs.
Market price and curtailment risk: For developers pursuing co-located renewables or PPAs with market price exposure, AI models that forecast nodal price volatility and curtailment risk at specific interconnection points are relevant. Historical LMP data combined with capacity expansion scenario modeling produces a distribution of future power costs rather than a point estimate.
5. Community and policy risk: will the approvals hold?
Power availability can be undermined by local opposition, moratoriums, or regulatory changes that block or condition large loads. The New York statewide data center moratorium passed in June 2026 and the active moratoriums in 13 other states illustrate that political risk is now a material underwriting variable.
AI systems can monitor zoning dockets, PUC proceedings, local government minutes, and media sentiment to flag sites where community or regulatory opposition is escalating. This is not prediction, but it is far better signal than a developer's own local knowledge.
How the Scoring Feeds Development Decisions
AI power risk scores serve three distinct decision contexts:
Site selection. Early-stage screening against a portfolio of candidate sites. A well-structured scoring model can eliminate sites with high queue congestion or reliability risk before a team spends on formal utility engagement or option payments. The developer is essentially buying information before buying land.
Investment committee underwriting. Once a site is in active diligence, the power risk model inputs directly to the financial model. Probability-weighted energization timelines change the IRR distribution. Reliability scores change the backup power budget. Tariff risk changes the revenue assumption on MG+E leases.
Lease negotiation. Power risk scores inform which provisions the developer insists on. A site with high queue risk but strong underlying demand is a different negotiating situation than one with power certainty. Understanding the probability distribution of outcomes allows the developer to price the risk into the deal structure rather than accepting the hyperscaler's standard terms.
What AI Cannot Do
The scoring frameworks described here are analytical tools, not substitutes for engineering or utility relationships.
A probabilistic queue risk model tells you the historical attrition rate for similar queue positions in that ISO. It does not tell you whether the three projects ahead of you will actually withdraw, or whether the utility's study team will identify a different upgrade scope than the model predicts. Engineering sign-off and direct utility engagement remain required.
The policy risk models can surface active moratorium proceedings. They cannot predict how a city council votes.
The value of AI power risk scoring is not in eliminating uncertainty. It is in quantifying uncertainty early enough that developers are making decisions based on a realistic probability distribution rather than an optimistic point estimate.
Implementation Notes
Most leading data center developers are building these scoring frameworks as custom integrations, not buying off-the-shelf products. The components are available: GIS platforms, ISO queue data via public FERC filings, LLM services for IRP and regulatory document parsing, and ML libraries for predictive modeling. The integration work is non-trivial but achievable with a small data engineering team.
For developers without that capability, the practical starting point is more structured use of publicly available data: ISO queue reports, utility IRP summaries, and PUC docket tracking. These do not require AI tooling to yield better site decisions than informal utility calls alone.
The developers building systematic power risk capability now are accumulating a proprietary data advantage that compounds over time. Every site screened and every utility engagement adds to the training data for the next model iteration.
Power is the constraint. Scoring it is the work.