Hidden AI workload data is weakening power grid forecasting

The expansion of large language models (LLMs) is creating a new problem for power grids: electricity demand from AI data centers is rising rapidly, but some of the information needed to forecast that load remains locked inside the companies running the models, researchers report in a new study.

The study, titled "Bridging the Information Gap: A Mechanism Design Approach to Forecasting AI's Power Grid Load" and published in Energies, proposes a two-stage forecasting framework that uses incentive mechanisms to obtain verifiable demand-related information from large language model providers before feeding those parameters into machine learning load forecasting models.

AI electricity demand is rising faster than grid planning cycles

Data center electricity consumption more than doubled between 2019 and 2022, reaching 460 terawatt-hours, according to figures cited in the paper. The International Energy Agency (IEA) projects global data center demand could exceed 945 terawatt-hours by 2030, more than the current electricity use of Japan. In the United States, data centers are projected to account for 9% of electricity consumption by 2030, up from 4% in 2022.

This growth is being driven by generative AI, cloud expansion, high capital investment and the strategic race to deploy AI infrastructure. However, the power system has not been built for such rapid, concentrated growth. Large AI data centers can create dense local loads that stress transmission and distribution networks, especially in regions where grid upgrades move more slowly than new server campuses.

The researchers point to a 2025 Deloitte survey in which 72% of power company and data center executives identified grid capacity constraints as a major barrier to infrastructure build-out. The study also cites warnings from the North American Electric Reliability Corporation that parts of the US grid face reliability risks, underscoring the tension between fast AI deployment and slower power-sector planning.

The challenge is not only that AI data centers use more electricity. It is that their load behaves differently from many traditional commercial loads. Large language model services respond to real-time inference requests, which means demand depends on the volume of user requests, prompt length, response length, model architecture, hardware type, batching policy, attention implementation and serving configuration.

Grid operators may see historical load and some aggregate usage patterns, but they often do not know the private technical parameters that determine how an AI provider's model turns tokens, prompts and responses into electricity demand. A model update, hardware migration, new batching strategy or change in inference stack can alter power demand in ways that historical forecasting tools may not detect quickly enough.

Traditional forecasting methods rely on past electricity demand and observable variables. Such methods can work under stable conditions, but AI workloads may introduce structural breaks when providers change their systems. The authors argue that forecasting AI-induced electricity demand requires not only better algorithms but also a way to obtain demand-relevant information that providers may consider commercially sensitive.

Incentive mechanism seeks private AI load data without full architecture disclosure

The framework proposed in the study has two stages. In the first stage, a grid operator uses a mechanism design approach to elicit reduced-form demand parameters from large language model providers. In the second stage, those elicited parameters are incorporated into forecasting models alongside observable variables such as demand history, calendar information and usage indicators.

Providers are not asked to disclose their full proprietary architecture. Instead, the mechanism targets verifiable reduced-form parameters that summarize how a provider's AI service translates usage into power demand. These parameters can reflect architecture and deployment factors while avoiding direct exposure of sensitive model details.

The authors model the problem as one of information asymmetry. AI providers know more about their own systems than grid operators do. But gathering and reporting accurate internal demand information also carries costs for providers, including monitoring, measurement, audit preparation and possible confidentiality concerns. The mechanism therefore uses payments to encourage providers to report at a chosen level of precision.

The study's theoretical model links the requested reporting precision to each provider's private cost of supplying that precision. Providers with lower reporting costs can be asked for higher precision, while those with higher reporting costs are assigned lower precision. The mechanism is designed to be incentive-compatible, meaning it aims to align the provider's reporting behavior with the grid operator's need for useful information.

Verification is crucial to the proposal. The researchers stress that incentives alone cannot guarantee truthful reporting unless the reports can be audited. They suggest that reduced-form demand parameters could be checked against ex-post measurements such as metered electricity consumption, aggregate token throughput, request volume, prompt and response length statistics, GPU utilization logs and third-party audit records.

The framework also allows for penalties if reported parameters fail an accuracy test. In practical terms, this means an AI company could protect its proprietary model details while still providing auditable demand descriptors that help grid operators forecast load more accurately.

The demand model used in the paper is intentionally reduced-form. It separates usage variables, such as request arrival rates and response lengths, from provider-side coefficients that capture how a given model and deployment configuration translate those usage patterns into electricity demand. The authors note that real LLM serving is more complex than the model, because energy consumption can be affected by KV-cache reuse, FlashAttention, grouped-query attention, sparse attention, mixture-of-experts routing, batching policies, memory hierarchy and GPU scheduling.

The proposed framework is not a full operational model of how every AI system consumes power. It is a proof-of-concept showing how privately held provider-side information can improve forecasting when made available in a verified, reduced-form way.

Forecasting errors fall when provider-side parameters are included

To test the approach, the researchers built calibrated synthetic scenarios using public data center energy reports, open LLM inference energy benchmarks and secondary public estimates. The study uses 16 representative providers over 90 days of hourly synthetic operation, with forecasts made 24 hours ahead. The authors emphasize that the results should not be treated as validation on actual proprietary AI data center load measurements, but as a controlled demonstration of the framework's logic.

The models tested included persistence, seasonal naive, Ridge regression, gradient boosting, an architecture-agnostic ResNet, a ResNet with elicited parameters and an oracle ResNet that uses true demand parameters without reporting noise. The study also tested gradient boosting with and without elicited parameters to see whether the gains depended on a specific neural network architecture.

The results showed that adding elicited reduced-form parameters substantially improved forecasting performance. The architecture-agnostic ResNet produced an average mean squared error of 2.02 × 10⁹, while the ResNet with elicited parameters reduced that figure to 6.72 × 10⁸. That represents a 65.1% reduction in mean squared error relative to the baseline ResNet.

The oracle version, which used the true reduced-form parameters without reporting noise, reduced error further to 2.93 × 10⁸. This served as an upper-bound benchmark showing how much forecasting could improve if provider-side demand information were perfectly available.

The same pattern appeared in the gradient-boosting comparison. Adding elicited parameters reduced mean squared error from 4.80 × 10⁸ to 1.90 × 10⁸. That finding is important because it shows the main value comes from obtaining otherwise unavailable provider-side demand features, not from the ResNet architecture itself.

The study also found that higher incentive budgets can improve forecasting by encouraging more precise reporting. When the budget available for eliciting provider reports rises, the variance in reported parameters falls, which lowers forecasting error. Conversely, when payments become more costly for the grid operator, the mechanism requests lower precision and total transfers decline.

The authors argue that this makes the framework useful as an information-procurement layer for power systems. It is not meant to replace unit commitment, economic dispatch, reserve procurement, optimal power flow or network-security analysis. Instead, it could improve the demand forecasts that feed into those downstream grid-planning and operational tools.

It is important to note that the empirical work relies on calibrated synthetic data rather than operational data center load traces paired with proprietary serving logs. The inference-energy model is simplified and does not fully represent all technical features of modern LLM deployment. The operator objective does not directly model full power-system constraints such as voltage stability, transmission congestion or reserve requirements. The audit mechanism also does not fully eliminate risks such as collusion, coordinated manipulation, strategic non-participation or data tampering.

The authors suggest that future work should test the framework with real AI data center measurements, richer serving models, privacy-preserving audits and integration into full power-system planning models.