The relentless advance of artificial intelligence is often framed as a software revolution. But beneath the algorithms lies a burgeoning engineering crisis. The demand for compute power, driven by large language models and other AI applications, isn’t simply straining existing infrastructure – it’s exposing deep vulnerabilities across the entire engineering stack, from raw material supply chains to the availability of skilled tradespeople. This isn’t a future problem; it’s happening now, and a new wave of research is focused on understanding, quantifying, and mitigating these constraints.
The Phantom Load: Deconstructing AI’s Power Demand
For years, forecasts of AI-driven electricity demand have treated ‘announced capacity’ – the total power requested by data center developers – as a reliable indicator of future load. However, recent work by N. Milton [1, 2] demonstrates that this is a profoundly flawed assumption. Milton’s analysis, based on data from grid operators like PJM and ERCOT, reveals a significant gap between announced capacity and actual delivered power. “Forecasts treat announced capacity as if it were firm on both sides of the meter,” Milton writes in [1], “This working paper deflates both…using the grid operators’ own resolved data.”
The findings are stark. PJM, which manages the electricity grid for a large swath of the eastern US, has seen its 2030 large-load request cut from 60 GW submitted to just 38 GW accepted, with only 32 GW backed by firm commitments. ERCOT, the grid operator for Texas, shows an even more dramatic discrepancy: metered peak consumption across new data centers is only 49.8% of requested capacity. This “deflation” isn’t random; it’s driven by real-world constraints. Projects are delayed or cancelled due to the cost of necessary network upgrades, lack of interconnection capacity, and, increasingly, the unavailability of essential materials and labor. Milton’s work provides a “two-rail deliverable envelope” – a band of 20-63% – for more realistic planning, urging stakeholders to treat announced headlines as scenarios, not baselines. This shift in perspective is crucial for responsible ratepayer protection and infrastructure investment.
The Constraint Relay
The issue isn’t simply a lack of power generation. Milton’s paper [2] introduces the concept of “the constraint relay,” highlighting how bottlenecks shift across the entire infrastructure stack as demand scales. Initially, the focus was on logic availability (the capacity to run AI algorithms). As that increased, the constraint moved to grid interconnections. Price signals, intended to incentivize investment, were instead met with physical limitations. This led to administrative rationing, pushing developers towards on-site generation, which in turn strained turbine manufacturing and rare-earth material supplies. The analysis details how even seemingly unrelated sectors – like high-voltage electrician apprenticeship pipelines and DRAM wafer allocation – became critical chokepoints. This cascading effect underscores the interconnectedness of modern engineering systems and the need for holistic risk assessment.
Beyond Independence: Modeling Complex Demand Patterns
Traditional online algorithms for matching markets and revenue management often assume that demand is independent and Poisson-distributed. This simplification, while mathematically convenient, fails to capture the complex serial correlations that characterize real-world demand patterns. Ali Aouad and Will Ma [3] propose a nonparametric framework that addresses this limitation. Their research introduces two models – “Indep” and “Correl” – that capture different forms of serial correlation, allowing for more accurate modeling of demand sequences. They demonstrate that relying solely on demand expectations (fluid relaxations) can lead to arbitrarily poor performance, highlighting the importance of incorporating distributional knowledge into algorithmic design. Their new algorithms, coupled with tighter linear programming relaxations, consistently outperform established methods, particularly when applied to real-world data exhibiting high demand variance. This work has significant implications for a range of applications, from ride-sharing platforms to dynamic pricing in e-commerce.
The Human Cost: Displacement and Extractivist Governance
While much of the focus is on technological and logistical constraints, the engineering boom – and the resource extraction it necessitates – has profound social consequences. Ayşe Çağlar’s research [4], though lacking an abstract, points to the rise of “extractivist governance” related to displacement caused by infrastructure projects. This framework highlights how states and capital collaborate to contain and control displaced populations in order to facilitate resource extraction and development. The paper likely explores the ethical and political dimensions of engineering projects that prioritize economic growth over the well-being of affected communities, a critical consideration often overlooked in technical analyses.
Intelligent Control Under Uncertainty: Gaussian Process MPC
Addressing the challenges of constrained systems requires sophisticated control strategies. Jie Wang and Youmin Zhang [5] provide a comprehensive tutorial on Gaussian Process Learning-based Model Predictive Control (GP-MPC). This approach combines probabilistic modeling with receding-horizon control, allowing systems to adapt to uncertainty and optimize performance in complex environments. The authors offer a detailed derivation of multi-step mean and covariance propagation, clarifying the roles of key parameters and providing practical guidance for implementation. Through mobile robot and vehicle platooning examples, they demonstrate how GP-MPC can be used to improve performance and ensure safety in real-world applications. This research is particularly relevant in the context of AI-driven systems, where unpredictable behavior and unforeseen circumstances are common.
The Power of Probabilistic Prediction
GP-MPC isn’t simply about making better predictions; it’s about quantifying uncertainty. By explicitly modeling the probability distribution of future states, the controller can make more informed decisions, avoiding risky maneuvers and optimizing performance under constraints. The tutorial distinguishes between mean-only unconstrained control and uncertainty-aware constrained control, emphasizing the importance of incorporating uncertainty into the decision-making process. This is a crucial step towards building more robust and reliable autonomous systems.
What’s Next? The Bigger Picture
The convergence of these research areas – the quantification of AI’s true power demand, the modeling of complex demand patterns, the recognition of social costs, and the development of intelligent control strategies – points to a fundamental shift in engineering philosophy. The era of simply scaling up existing infrastructure is over. We are entering an age of resilience, where adaptability, resourcefulness, and a holistic understanding of interconnected systems are paramount. Future research will likely focus on:
- Dynamic Infrastructure Management: Developing algorithms and systems that can dynamically allocate resources and respond to changing demand patterns in real-time.
- Supply Chain Diversification: Reducing reliance on single sources of critical materials and fostering more resilient supply chains.
- Ethical Engineering Frameworks: Integrating social and environmental considerations into the design and implementation of engineering projects.
- AI-Driven Optimization: Leveraging AI itself to optimize infrastructure performance, predict failures, and improve resource allocation.
- Closed-Loop Systems: Creating fully integrated systems where data flows seamlessly between sensors, models, and controllers, enabling autonomous adaptation and optimization.
The challenges are significant, but the potential rewards are even greater. By embracing a more nuanced and integrated approach to engineering, we can harness the power of AI while mitigating its risks and building a more sustainable and equitable future.
References
- N Milton (2026). Announced vs. Deliverable AI Power Demand. Zenodo (CERN European Organization for Nuclear Research).
- N Milton (2026). The AI Constraint Relay: How the Compute Build-Out Displaces Physical and Financial Limits. Zenodo (CERN European Organization for Nuclear Research).
- Ali Aouad, Will Ma (2026). A Nonparametric Framework for Online Stochastic Matching with Correlated Arrivals. Management Science.
- Ayşe Çağlar (2026). Extractivist governance of the displaced and their containment: rethinking state, capital, and labor. Ethnic and Racial Studies.
- Jie Wang, Youmin Zhang (2026). A Tutorial on Gaussian Process Learning-based Model Predictive Control. ACM Computing Surveys.