Summary

Maintaining American leadership in AI will require an infrastructure project at a scale this country has not seen in decades. We must build many gigawatt-scale (GW) clusters,[^1] each requiring the energy equivalent of multiple nuclear power plants. To achieve this, US policymakers must unleash America's industrial capacity. They must radically reduce timelines for environmental permitting and help developers take on the technical risks involved in scaling next-generation energy technologies such as small modular reactors (SMRs) and enhanced geothermal.

However, huge investments in AI infrastructure will count for little if the products of these investments — systems that could both form the basis of the US economy and reshape the global balance of economic and military power — can easily be sabotaged, stolen, or used against us by our adversaries. The AI and computing industry is underinvesting in the level of security required to successfully secure and defend their technologies against nation-state-level actors, if the situation demands it. We must ensure that the future of AI is both built in America, and good for America.

We propose that the federal government establish "Special Compute Zones" — regions of the country where AI clusters at least 5 GW in size can be rapidly built through coordinated federal and private action. Within Special Compute Zones, the government should use federal authorities to accelerate permitting and solve supply chain bottlenecks, and unlock financing for next-generation power plants. In return, the government should require security commitments from AI and computing firms — making nation-state-grade investments in AI security a sensible commercial decision, rather than one which puts a firm at a disadvantage relative to its competitors.

Problem

Ensuring the most advanced AI data centers are built in America will yield two large advantages. Economically, it means American firms can capture the immense value created by cutting-edge AI development and secure priority access to frontier models. Developing frontier models here also means that we have the option of withholding their capabilities from our adversaries. This may become necessary in worlds where the predictions of those at the frontier of AI development turn out to be correct: massive AI-driven workforces achieving cyber-dominance through rapid software development, hacking, and digital reconnaissance; mass-persuasion campaigns and elaborate forms of espionage; and dramatic acceleration of weapons design and military autonomy.

Success requires overcoming two major obstacles. First, our existing energy system is simply not able to compete. American power generation growth has lagged far behind China's over the last 25 years.

To reverse this trend, policymakers must first address the burdens imposed by regulations. More than 70 percent of energy projects in the queue to connect to the grid are withdrawn due to long wait times, which have doubled since 2005. In 2013, around 4,000 new miles of transmission lines were built in America. Today, this figure is close to 500 miles, and it takes 10 years on average to build a new line.

A number of more technology-specific issues also plague any large-scale US energy buildout:

• Existing capacity is not enough. 74 US coal plants totaling 35 GW of generation capacity are due to be retired by the end of 2029. As has already started to happen, these plants could be kept online to power AI data centers, but they are a highly polluting source of energy, and will not be sufficient to power the entirety of the AI data center buildout.
• Thanks to issues with licensing, permitting, and supply chains, building new large-scale clean firm capacity using proven technologies takes a long time: the last US nuclear plant to come online was 7 years late and $17 billion over budget. Hydropower is in a similarly dire state — it takes an average of 7 years to obtain a license for a new project, and a further 5 to 10 years for construction.
• Large-scale generation projects require equipment such as electrical transformers, which are often custom-built for projects. However, transformers have a lead time of one to two years, which has increased by two to four times over the last five years.
• While natural gas plants can be built relatively quickly, supply chains have little capacity. GE Vernova, the world's largest manufacturer of gas turbines, is reportedly sold out beyond 2030. New gas plants also face uncertain long-term economics due to rapidly improving alternatives, regulatory risk, and corporate decarbonization commitments.
• Next-generation technologies like advanced geothermal and small modular reactors show immense promise, but face financing challenges due to cost and timeline uncertainty during the early stages of development.

Given enough time, market forces will solve many of these issues. But to hit aggressive timelines, policymakers must help reduce the burdens on industry imposed by regulation, and help developers take on the financial and technical risks involved in scaling quickly.

The second major obstacle: the benefits of a massive AI data center build out will count for little if our ability to secure breakthrough AI systems lags behind our adversaries' ability to sabotage, steal, and use them against us. At present, the AI and computing industry is underinvesting in the level of security required to adequately secure their technologies against nation-state-level actors. This represents a clear market failure: it is in the American public's interest to ensure that powerful models are not compromised or used against us by our adversaries, but American AI developers and computing firms are locked in a race with each other to build ever more powerful models. If they invest in sufficient security to protect their systems from top Chinese state-backed hacking groups, they risk falling behind.

Solution

We propose an ambitious federal program to accelerate the AI data center buildout within particular geographic regions of the United States — "Special Compute Zones." A place-based policy agenda makes sense: focusing attention on specific areas reduces the number of stakeholders who need to coordinate to build quickly, and allows for targeted public and private investments in shared electricity infrastructure costs. Because AI training clusters can be flexibly located based on power availability, Special Compute Zones can be planned around areas where it is possible to build quickly: including federal lands where local control is limited, areas with existing nuclear capacity or soon-to-be-retired coal sites (where large-scale energy support infrastructure already exists), areas with substantial and consistent sunlight for solar energy production, and areas with high potential for next-generation geothermal production.

The program should include the following measures:

Establish strong and effective leadership

The President should appoint an AI infrastructure czar to coordinate the establishment of Special Compute Zones, with executive branch experience, a deep understanding of energy infrastructure, and the ability to work closely with industry on ambitious security initiatives.

Identify and prioritize Special Compute Zones

The czar should lead a comprehensive interagency review to identify the most promising Special Compute Zones, focused on:

• Compiling an inventory of federal lands suitable for AI infrastructure development.
• Identifying existing energy assets (such as retired coal sites) that could be upgraded or repurposed under the Department of Energy's Loan Programs Authority — Section 1706 of Title XVII of the Energy Policy Act of 2005. Many of these sites reside on or near federal land, with existing infrastructure that can meet the scale for AI training clusters while tapping into existing rights-of-way, reducing permitting hurdles and project timelines.
• Identifying "previously disturbed lands" that qualify for categorical exclusion from environmental review to help lower regulatory uncertainty.
• Identify land available for acquisition under Section 161g of the Atomic Energy Act, and utilize that authority to support nuclear energy infrastructure for AI data centers.

Use federal authorities to accelerate construction

Once Special Compute Zones are identified, the czar should work with the interagency to streamline permitting and unlock financing for AI infrastructure within the zones, including:

• Using DPA Title I authorities to prioritize orders for critical equipment essential for AI infrastructure to the top of suppliers' order books, including gas turbines, high-voltage transformers, and switchgear. Title I can also be used to prioritize not just end products, but also component parts and materials, including silicon steel for transformer cores and specialized cooling systems.
• Using DPA Title III authorities to provide loans for next-generation energy infrastructure, alongside requirements for firms to invest in security measures that protect the AI technologies they are developing from sophisticated attackers and coordinate with the US government on their implementation, following the precedents set by the 2009 Smart Grid Investment Grants program and the CHIPS and Science Act.
• Using DPA authorities to streamline permitting and environmental review, including using NEPA's emergency circumstances provision (40 C.F.R. § 1506.11) to avoid following the conventional, time-intensive process, using legal protections under NEPA for classified and sensitive information to shield projects from legal challenges (40 C.F.R. § 1506.11), invoking Section 7(j) of the Endangered Species Act to obtain national security exemptions, and using the DPA Title III "without regard" clause to waive NEPA requirements for high-priority facilities.
• Establishing new categorical exclusions to NEPA for activities that don't have material impacts, such as design, site characterization, and materials acquisition. This would allow the disbursement of federal loans to accelerate non-disruptive activities, whose expense would normally have to be fronted by developers. Following the 2023 Fiscal Responsibility Act, agencies can adopt categorical exclusions issued by other agencies.

Tie federal assistance to security requirements adequate to protect American AI technology against our adversaries.

The czar should launch new initiatives to radically improve the security of American AI infrastructure, enlisting DOD, DOE, NIST, NSA, CIA, FBI, CISA, and USCYBERCOM. This should include:

• Rapid development of improved security specifications to defend against sophisticated attacks on supply chains, hardware, and networks.
• Creating points of contact in the intelligence community to advise the builders and operators of AI data centers on vulnerabilities to exfiltration, sabotage, and denial operations
• Assisting with the design of secure data centers and inter-data center network infrastructure.
• Establishing a red team to conduct penetration testing of AI infrastructure, and establishing a fast and effective background screening process for roles that involve access to sensitive hardware or data at advanced AI data centers.
• Providing tracking and physical security for shipments to AI data centers, and assistance with screening devices entering data centers to detect and prevent supply chain attacks.
• Launching new research programs to radically advance the state-of-the-art in security for AI hardware devices, including cluster-scale confidential computing and protection from invasive and non-invasive physical attacks.

---

Parts of this agenda have already begun through the Biden administration's AI Infrastructure Executive Order. However, the executive order has three important gaps:

1. While it makes federal lands available for AI data centers, it doesn't contain a comprehensive plan to speed up the permitting issues this creates.[^2] Nor does it provide tools to resolve the looming supply chain shortages, particularly in gas turbines, that energy infrastructure for AI data centers will experience. Both these problems can be addressed through the use of the Defense Production Act.
2. It requires that AI data centers be powered exclusively with clean energy. This neglects the reality that while next-generation clean energy technologies such as enhanced geothermal and small modular reactors offer a longer-term energy solution, natural gas turbines must form a large part of the near-term solution.
3. While it introduces an initial set of security requirements for AI and computing firms, these requirements are not sufficient to adequately protect strategically critical American AI technologies from being stolen by our adversaries.

The Trump administration should issue a new Executive Order to fix these issues, ensuring that increasingly powerful AI systems are both built in America and good for America.

[^1]: A "cluster" is an interconnected group of computers that can be used together to train AI models. Today's largest clusters span multiple data centers.

[^2]: NEPA automatically applies