Material and Immaterial: AI Proliferation and Energy Capacity
By Sam Straus
Productivity Paradox
Over the last few years, we’ve seen a significant rise in discussion surrounding the benefits of artificial intelligence. Billions of dollars have poured into erecting AI data centers throughout the country as Silicon Valley start-ups and legacy tech firms purport global disruption and immense economic benefit. Unfortunately, these sci-fi dreams may be premature as recent reports illuminate the current, nominal productivity benefits–mostly limited to individual tasks like memo writing, spreadsheet manipulation, and schedule planning–to firms and consumers alike from the adoption of AI tools at home and in the workplace.
As The Economist recently reported, real GDP growth and aggregated working hours indicate a productivity growth rate of 1.9% in 2025, below the long-run average of 2% and significantly smaller than the growth rate during the internet boom. In addition, Harvard economist Jason Furnam shared that, excluding direct and indirect investments to data center infrastructure, US 2025 GDP growth may be annualized at around 0.1%. This further suggests that, excluding the direct boost to construction firms, chip producers, and energy providers, AI is having minimal impact on economic growth. Meanwhile, Morgan Stanley’s Counterpoint Global Insights report outlined the low probability for OpenAI to achieve its forecasted five-year sales growth rate of 108% given it falls 9.5 standard deviations right of the mean.
While the productivity benefits of AI remain murky, its immense energy costs are clear. The North American Electric Reliability Corporation (NERC) forecasts US peak winter load demand to rise by 149GW in the next decade, a 21.5% increase between 2024 and 2034. This growth is largely driven by data center expansion and its endogenous computing requirements. For instance, the planned Homer City Energy Campus in Pennsylvania will require 4.5GW which is roughly the same as the city of Philadelphia. And forty miles north of Chicago, T5 Data Centers plans to develop a 1.2GW data center park in Grayslakes, IL with 1.6GW of secured power through ComEd.
To fund these and similar projects, firms are investing heavily. Microsoft, Alphabet, Meta, and Amazon reported $370 billion in 2025 capital expenditures, the majority of which is directed towards AI development and infrastructure buildout. Combining the massive spending, energy demand, and marginal productivity benefits produces a conundrum. If AI is a game-changing technology, where is the evidence? If it is not, why are firms devoting so many resources towards its development? These questions sit at the heart of what’s been labeled the “AI Bubble” and will likely continue to challenge researchers, investors, and technologists in the coming months and years.
Capacity Constraints and a Path Forward
Luckily, one area illuminated by AI proliferation is the stark need to modernize the electric grid. Like other energy consumers, data centers (and their developers) demand cheap, reliable power, albeit on a much larger scale, similar to the size of a town or small city. Furthermore, since data center energy demands share similarities to those of urban areas and existing industrial zones, many of the proposed solutions to AI capacity constraint can be applied outside of the tech sector. Below are three notable examples.
- Bring Your Own Capacity (BYOC): This is where end-users supply their own generation capacity either through on-site resources or off-site Power Purchase Agreements (PPA). For data centers, this means locally installing solar fields, gas turbine engines, or in the case of Microsoft and Constellation, restarting the Three Mile Island nuclear power station in an effort to reduce grid interconnection and transmission delays. Aside from AI, BYOC has broad applications as it enables traditional end-users (individuals, factories, etc.) to procure sufficient capacity within a geographic locality to mitigate energy generation development timelines and constraints. Furthermore, if utilized by towns or municipalities, BYOC may offer a path to energy resilience as locally generated and stored power can lessen demand spikes from severe weather events as well as the risk to long transmission lines from falling branches, traffic accidents, or substation malfunctions.
- Load Factor Flexibility: Load factor is the ratio of average capacity demand to peak demand within a given period. A high load factor is indicative of a system that operates closer to its peak demand for more hours in a given year and the opposite holds for low ratios. It’s a critical calculation when determining energy usage and capacity constraints as systems with low load factors are likely underutilized and therefore potentially better suited to accommodate new data centers or large commercial operations. Additionally, low system utilization may represent underutilization of expensive grid assets: transformers, transmission lines, distribution networks, etc. Given the high cost of infrastructure and the generally low load growth, 1% annually over the past two decades, incorporating load factor flexibility into energy policy offers a means to facilitate data center expansion while new power generation comes online. For end-users, greater load factor flexibility may help reduce rising electricity bills as existing capacity is more efficiently utilized to service new loads, potentially lowering grid operator costs. Moreover, a closer examination of load factor flexibility is a positive step in thinking about the grid, driving producers and consumers to seek tailored solutions in areas constrained by current energy capacity.
- Flexible Grid Connections: Flexible grid connections refer to a data center or industry’s ability to use a mix of power provided from the grid and generated by on-site resources to meet demand. With flexible grid connections, end-users receive stable (grid power) and conditional (local power) energy, where a portion of the load requirements are normally met by the grid and the remaining requirements during high-stress periods are serviced by on-site generation. A recent study by Camus, Encoord, and Princeton’s Zero Lab found that flexible grid connections in conjunction with BYOC shortened data center grid connection wait times by three to five years.
Additionally, flexible grid connections offer a significant mitigation to capacity concerns. Just this past fall, PJM, the organization responsible for ensuring transmission resource adequacy in the Mid-Atlantic region, filed a complaint with the Federal Energy Regulatory Commission concerning the rate at which data centers are coming online and PJM’s subsequent inability to maintain a reliable grid. Pursuing a flexible grid connection policy could reduce these concerns and shift a significant portion of data center energy requirements directly to the data centers. Outside of the AI space, flexible grid connection offers a policy lens to evaluate the responsibilities (or lack thereof) of large energy consumers to the grid as a whole. It also provides a novel solution for regulators and transmission authorities to examine as they overcome current and future energy obstacles.
Conclusion
Though the AI productivity boom may be farther away than anticipated, one clear lesson from the current fixation on data centers and energy expansion is the alarming need for grid modernization and investment in the United States. The rise of AI has performed a valuable, if alarming, service by revealing capacity constraints and low adaptability within the power grid.
These concerns, however, are rapidly giving way to innovation with broad application–providing potential solutions and standards to bolster energy resiliency. Applying BYOC, load factor flexibility, and flexible load connection standards could allow communities and governments to more efficiently meet energy demands. They may also reduce transmission resource costs by locating power generation and storage facilities closer to end-users, providing flexibility to accommodate periods of heavy demand.
Most importantly, without regulation and cooperation between the public and private sectors, data center proliferation may only serve to exacerbate an already strained, aging grid. Whether this leads to a period of painful, inequitable capacity strain and higher energy prices or a large-scale collaborative modernization effort is the central policy and business challenge of the coming years.
Reader question
How should the cost of grid investments, which benefit all but are largely utilized by private industry, be distributed to taxpayers?