Photo: Just_Super/GETTY IMAGES
“The U.S. possesses all components across the full AI stack, including advanced semiconductor design, frontier algorithms, and, of course, transformational applications. Now the computing power this stack requires is integral to advancing AI technology, and to safeguard America’s advantage, the Trump administration will ensure that the most powerful AI systems are built in the U.S. with American-designed and manufactured chips.”
– Vice President JD Vance, Paris, February 11, 2025
Today, the United States leads the world in generative AI. Its frontier labs set the pace in model development, U.S. firms control more than half of the world’s AI accelerators, and U.S. capital markets are poised to rapidly scale investment in data center infrastructure. Lasting U.S. advantage in AI, however, is not guaranteed.
The global race for compute is intensifying as competitors—adversaries and allies alike—are maneuvering to catch up. Beyond the recent breakthrough with DeepSeek, China is building massive data centers, expanding its power sector, and developing domestic AI chips to reduce Western dependence. France aims to leverage surplus nuclear power to attract data centers and support AI research centers across the country. Japan seeks to overcome space and energy constraints by powering highly efficient data centers with idled nuclear plants. The United Arab Emirates is creating AI-focused economic zones and incentives to attract international companies, with nuclear power as part of its strategy.
To stay ahead in the AI race, the United States should put meaningful distance between itself and competitors across all components of the AI stack—frontier models, data centers, advanced chips, and energy. These constitute the fundamentals of AI competitiveness. While all components are important, by far the most pressing need today is ensuring rapid access to the electricity needed to power large data centers. Simply put, failing to secure energy means surrendering U.S. leadership on AI. At stake in the United States is long-term growth and productivity, market security, and national security.
The central message of this CSIS Economic Security and Technology Department report is that, while the AI revolution is digital in nature, its binding constraint is physical infrastructure. The AI race will be won by whoever scales investment and delivers infrastructure the fastest, most reliably, and in ways that generate maximum positive spillovers for the broader economy.
Our team has developed a range of scenarios to assess the semiconductor, energy, and capital needs for leadership across the AI stack. These scenarios correspond to surges in business investment seen during the Dot-Com Boom, PC revolution, and Second Industrial Revolution. Given the nearly inexhaustible demand for model training and inference (or AI applications) compute, our analysis has focused on the supply-side delivery of compute to the AI sector. A long-term research priority should be to forecast the economy-wide trajectory of compute demand, based on the sectoral adoption of AI, and identify policy options to reduce sectoral barriers to AI uptake, innovation, and growth.
Based on this supply-constrained framework, we estimate that data center expansion will require 80–160 million leading-edge GPUs (measured in H100 equivalents) and 40–90 gigawatts (GW) of new energy demand at a total capex of $2 trillion by 2030. Summarized in the table below, we explore the policy implications of these findings:
To meet the forecasted 2030 target of 80–160 million leading-edge GPUs, global leading-edge semiconductor manufacturing would need to sustain an annual growth rate of approximately 72–96 percent year-on-year from 2025 to 2030. Our high case estimate requires a thirty-fold scaling of chip production beyond what all U.S. firms combined have on hand by the end of 2025.
According to estimates, GPU supply could grow anywhere from 30 percent to 100 percent. To hit the upper bound of this estimate, our model assumes frictionless scaling across every stage of the semiconductor supply chain, from wafer fabrication to advanced packaging. Even minor bottlenecks in foundry expansion or supply chain disruptions could cascade into cost spikes, shortages, and forced recalibrations of AI model deployment strategies. Moreover, given potential geopolitical flashpoints, the United States should cement domestic manufacturing capacity across both chip fabrication and advanced packaging—not as redundancy, but as strategic necessity.
Photo:sdecoret/Adobe Stock
AI data centers run on enormous volumes of electrical power. Our high growth scenario estimates that over 80 GW of new data center demand could be deployed in the United States by 2030, consuming well over 400 TWh of electricity annually. These sites alone will drive 2 percent annual growth in total U.S. electricity demand, a paradigm shift relative to the 0.2 percent growth rate observed since 2007. By 2030, total data center electricity consumption will consume more electricity than the state of California today (around 240 TWh).
The U.S. power sector has essentially zero “spare capacity” to meet new data centers following two decades of near-zero demand growth nationwide. Every new gigawatt of data center demand must be matched by a new gigawatt of effective capacity generation. In this environment, speed-to-power, how fast a potential data center site can access electricity supply, is the critical factor driving investment decisions. Short-term policy should boost speed-to-power for data centers by smoothing the path to deployment for the currently preferred mix of gas, solar, and storage. Long-term policy must seek to establish global electricity supply dominance via support for nuclear power and minimize cost inflation for electricity ratepayers with federal grid investments.
Based on current trends, our high growth scenario estimates that data center capex (inclusive of the cost of leading-edge chips) could total $2.3 trillion by 2030. With investment approaching more than $500 billion per year at the end of this decade, data centers will be a source of investment far exceeding annual investment in oil and gas exploration and production in North America (around $200 billion in 2024) or the power sector ($179 billion in 2024).
Both the data center industry and power sector have no problem attracting private capital to fund the boom in capacity requirements. Federal strategic capital needs to be deployed to mitigate risks the private capital markets are not allocating for. On energy, strategic investment in nuclear power and energy transmission is predicated on the ultra-long (80 years or more) value proposition of that infrastructure. In the chips sector, federal strategic capital needs to drive domestic chip fabrication investment to reduce U.S. exposure to Taiwan supply chain risks.
The following recommendations are designed for consideration by the offices of the assistant to the president for science and technology, the White House AI and crypto czar, and the national security advisor, who are tasked with preparing an AI action plan by a recent Trump administration executive order. Key elements that require action by Congress could be considered in the 2025 legislative agenda.
Objective: Improve speed-to-power for data center projects in power supply-constrained regions.
The newly established National Energy Dominance Council (NEDC) should leverage emergency authorities to accelerate permitting for energy infrastructure projects, with a particular focus on the Northern Virginia computing cluster where data center expansion faces severe constraints. This strategy should encompass fast-tracked approvals for the gas midstream, electric transmission, and generation projects, including offshore wind development along the Virginia coast.
The NEDC should also consider emergency federal siting authority for high-voltage transmission projects, especially those approved by PJM to integrate the northern Virginia cluster with surrounding power sources, as well as similar projects serving emerging industrial and data center clusters in the Midwest, Southeast, and Southwest regions.
The NEDC could recommend that agencies employ emergency authorities to retain coal capacity on a case-by-case basis. Most plants are unlikely to need support in the near term due to improved economic prospects, where such authorities are used, they should be supplemented with support for new long-term generation capacity (such as nuclear generation).
Objective: Bring forward timelines for new nuclear construction.
Congress should appropriate funds and grant new authorities to the Department of Energy (DOE) to apply the “anchor tenant” currently authorized for transmission projects to offtake contracts for new nuclear power. The DOE, possibly in concert with federal power marketing administrations or the Tennessee Valley Authority, would be directed to support a minimum of 10 GW of nuclear projects by 2030 across multiple sites and reactor technologies. As projects near completion, the DOE would sell offtake capacity to private firms (e.g., hyperscalers) or transfer to utilities to enable ratepayer access to nuclear power. Financial risk sharing should be authorized insofar as it is shared across private parties; this authority should not be implemented as a form of cost overrun insurance.
Objective: Resolve coordination problems between states, utilities, nuclear developers, and hyperscalers that constrain nuclear power deployment.
The White House should instruct the National Energy Dominance Council to designate “nuclear-powered computation hubs,” which will empower states to rapidly develop nuclear power that serves long-term data center demand scaling, improves grid reliability, and reduces system-wide costs for existing ratepayers.
Nuclear-powered computation hubs take a state-led and federally supported approach, which would encourage partnerships between state energy offices, data center operators, and power producers, targeting sites capable of hosting 2 GW data centers and 2 GW or more nuclear capacity. Sites should have access to the high-voltage transmission network and can include additional generation resources (gas, solar, geothermal, and solar) on or off-site to support data center load prior to the nuclear commercial operation date. Selected hubs would receive expedited federal permitting under emergency authorities, federal loan guarantees, site-development grants, grid expansion funding, and DOE nuclear offtake support while participating states must streamline their own permitting processes and establish nuclear, data center, and engineering workforce development programs.
Objective: Accelerate grid investments to improve data center speed-to-power without burdening existing ratepayers with increased costs.
Grid investments that enable AI leadership deliver strategic value to the nation but risk burdening existing ratepayers with increased costs. The DOE’s Grid Resilience and Innovation Partnerships (GRIP) fund, with its remaining $2.4 billion, should be refocused to prioritize high-voltage grid investments that facilitate speed-to-power for data centers. Grid investment is crucial as both hyperscale computing clusters and large generation facilities require access to high-voltage transmission, as evidenced by Meta’s 2 GW Richland Parish data center’s location near the Southeast’s 500 kV backbone and the attraction of data centers to AEP utilities’ 765 kV grid system.
In the next five years, utilities will invest billions in grid infrastructure to support data center expansion, potentially increasing electricity prices or slowing deployment as costs of that buildout are allocated. Congress should consider replenishing and refocusing the GRIP funds to offset a portion of this investment surge and support proposed nuclear computation hubs.
Objective: Expand energy siting options for states, data center developers, and power developers by improving access to federal land and utilizing emergency permitting authorities.
Identify and prepare federal sites through the DOE, the Department of Defense, and the Department of the Interior for data center development, leveraging existing infrastructure and streamlined permitting authorities. Sites should be selected based primarily on their ability to deliver speed-to-power across an all-of-the-above portfolio of generation technologies at the speed and scale necessary for 5 GW data center cluster development. Site selections should be available for partnership and participation in state-led nuclear computation hubs to improve opportunities for states with large amounts of federal land.
Objective: Secure the supply, including domestic supply, of leading-edge semiconductors to power AI data centers.
The Chips Program Office within the Department of Commerce should ensure the timely implementation of committed funding agreements under the CHIPS program for domestic fabrication of high-end AI accelerators (GPUs, AI application-specific integrated circuits) and advanced packaging (2.5D and 3D integration).
The International Trade Administration in the Department of Commerce should focus the upcoming SelectUSA Investment Summit on foreign direct investment attraction to finance the AI stack to include domestic chip-making capabilities and nuclear power generation. The department should monitor committed investments in leading-edge fabs in the United States.
Consistent with the goals of the CHIPS and Science Act, the secretary of commerce should encourage hyperscalers to diversify the sourcing of high-end chips across U.S.-based suppliers including through a voluntary advanced market commitment (a U.S. version of the First Movers Coalition).
The Department of Commerce, working with the National Center for the Advancement of Semiconductor Technology, should establish a common framework to assess in real-time the development of a chips workforce to include a long-term plan for developing homegrown talent with a strategic immigration solution to meet short-term needs.
Objective: Protect U.S. federal investments in advanced chipmaking by ensuring that high-end chips do not fall into the hands of mercantile and malign actors.
Appropriately resource the Bureau of Standards and Industry in the Department of Commerce with expert staffing and data capabilities to better enforce export controls and identify violations with respect to semiconductors.
The White House should instruct relevant departments and agencies to share information, with the Department of Commerce as a clearing house, to map connections between suspicious entities in Chinese shadow networks.
Objective: Maintain and grow U.S. leadership in advanced chip design.
Congress should extend tax credits for domestic design and manufacturing of leading-edge AI chips for break-ground dates between 2026–2031.
Objective: Facilitate the most efficient use of data centers and their existing capacities.
The National Telecommunication and Information Administration should establish a data center connectivity fund using its Bipartisan Infrastructure Law authorities, specifically for ultra-high bandwidth fiber infrastructure connecting major data center clusters. This fund should prioritize projects that deliver a minimum capacity with scalability between major data center regions, reduce latency for distributed AI workloads, and support multi-region redundancies. The program should include technical assistance for state broadband offices, data center operators, telecommunications providers, and AI research institutions to identify critical connectivity gaps between major computation clusters.
Objective: Accelerate innovation of AI applications for commercial use across sectors.
The Economic Development Administration in the Department of Commerce, using its authorities under the CHIPS and Science Act, should create a specialized designation program within the tech hubs framework that identifies and connects regions with complementary AI strengths. These “AI innovation zones” would receive priority funding for developing shared infrastructure including high-capacity data centers, specialized cooling facilities, and renewable energy resources specifically scaled for AI workloads. Require designated zones to implement standardized permitting processes for AI infrastructure development with expedited timelines for projects meeting predefined sustainability metrics. The “AI innovation hubs” should also include financial incentives for collaborative projects that distribute compute resources across multiple hubs to create resilient, load-balanced AI research networks. They should also support AI applications across sectors with a view to driving leadership on inference.
Congress should fully fund the tech hubs program to include “AI computation hubs” with a focus on prioritizing AI applications in a wide array of sectors such as biotech, clean tech, finance, and other areas to drive leadership on inference.
Navin Girishankar is the president of the Economic Security and Technology Department at the Center for Strategic and International Studies (CSIS) in Washington, D.C. Joseph Majkut is the director of the Energy Security and Technology Program at CSIS. Cy McGeady is a fellow with the Energy Security and Technology Program at CSIS. Barath Harithas is a senior fellow with the Economics Program and Scholl Chair in International Business at CSIS. Karl Smith is an economic consultant providing economic analysis for the Economic Security and Technology Department at CSIS.
The research informing this commentary was made possible by generous support from OpenAI and general support to the Economic Security and Technology Department at CSIS.
