Protecting the U.S. AI Compute Advantage
I’m Ylli Bajraktari, CEO of the Special Competitive Studies Project. In this week’s edition of our newsletter, Konstantin Pilz of SCSP’s Defense team discusses how leading in AI chips and data centers allows America to lead in AI. To maintain this leadership he argues that the United States must boost, retain, and protect AI infrastructure.
Save the Date for the AI+ Expo for National Competitiveness in Washington, D.C. on June 2-4, and be a part of shaping the future. Don’t miss this opportunity to be at the forefront of the technology revolution changing our world as we know it. Visit expo.scsp.ai to learn more today!
The United States currently dominates global AI compute infrastructure, controlling half of data centers worldwide and supplying virtually all advanced AI chips. As AI capabilities become increasingly tied to compute power, dominating this infrastructure allows the United States to lead in AI and provides it with great influence over the trajectory of AI development and deployment globally. Yet, this dominance in AI infrastructure is not guaranteed. In fact, absent a more deliberate public-private strategy that promotes and protects this advantage, the United States runs the risk of ceding its leadership position and foregoing the opportunities it offers. Such a strategy will need to address at least two issues – clearly articulating export controls beyond close allies and pacing adversaries, and implementing a layered defense of America’s data centers that protects them from cyber attacks, espionage, and physical threats.
Why Data Centers Matter
Data centers are the physical infrastructure that enables the development and deployment of AI models. These billion-dollar, industrial-size facilities provide compute at scale, hosting the hardware that runs everything from banking to streaming to high-performance computing tasks like AI training and deployment. As AI advances, it has become increasingly infrastructure-intensive. Training the current generation of AI systems already requires tens of thousands of specialized AI chips installed in these data centers. The next generation of AI models are currently being trained on clusters containing more than 100,000 chips and companies are already making plans for data centers hosting clusters of 2-3 million chips. While paradigm shifts in algorithms or hardware could reduce future infrastructure needs, the current trajectory, as well as empirical scaling laws, point toward continued exponential scaling, at least for the next few years,
The U.S. Compute Advantage
The United States leads in several critical aspects of AI compute:
AI chip supply — Training and deploying modern AI models requires thousands of specialized AI chips. U.S. companies, most prominently NVIDIA, Google, AMD, and Intel control more than 90 percent of the AI chip design market.
Data centers — Hosting AI chips at scale requires industrial-size data centers. About half of all data centers are in the United States.
AI supercomputers — Inside its largest data centers, the United States hosts most of the largest AI supercomputers in the world, such as xAI’s Colossus, comprising 100,000 H100 GPUs. While there is no comprehensive data on the distribution of large AI clusters, plans by Meta, OpenAI and others reveal that U.S. companies are rapidly expanding their largest supercomputers.
Cloud compute — Most AI development and deployment today runs on the compute of a small number of cloud providers. The largest of these, AWS, Microsoft Azure, and Google cloud alone make up 63% of the global cloud market. While there are no figures specific to AI offerings in the cloud, these three likely similarly dominate that space.
Dominating in AI compute provides the United States companies and government with considerable leverage over the direction, speed, and shape of AI development. It enables the United States to:
Deny compute access to adversaries to curtail their threatening AI ambitions. For example, blocking exports of advanced AI chips and scrutinizing the provision of cloud services by third parties to China has allowed the United States to slow down Beijing’s aspirations to dominate AI development globally in the coming decade, particularly for military applications.
Pursue “compute diplomacy.” The current compute dominance gives the United States diplomatic leverage, that if used effectively, could open new markets for U.S. companies, hinder our adversaries, and expand the trade space for broader arrangements.
Set global standards: Controlling the majority of AI compute gives the United States the opportunity to set global norms and standards about AI development, which in turn could help consolidate U.S. competitive advantage and shield American companies from foreign regulations.
Promote U.S. prosperity and national security AI systems are rapidly becoming and will continue to become increasingly capable, providing opportunities to further strengthen U.S. national security, as well as enhance the wellbeing of Americans. Leading in compute could ensure leadership in the next, crucial milestone of AI development — attainment of artificial general intelligence (AGI) — and provide the United States with a unique window of opportunity to pursue the enormous benefits that AGI could unlock, and ample time to prepare for potential downsides.
Threats to U.S. leadership in AI infrastructure
Through decades of leadership and capital investments in digital technologies and platforms, the United States has been able to field and expand its data center infrastructure. However, three dynamics are converging that could threaten U.S. leadership:
Foreign pursuit of AI infrastructure. China is the nation with the second-largest investments in infrastructure, and the most serious contender to displace the United States from its leadership position. However, other nations are increasingly investing into AI and attempting to increase their share of global infrastructure. For instance, Saudi Arabia is planning a $40 billion AI investment fund and the UAE has already launched a $100 billion fund this year, entirely focussed on AI. Other nations, despite their more limited resources, will likely follow suit.
The growth in supply of energy in the United States may not keep pace with planned growth in the footprint of data centers. As the United States rapidly expands its compute footprint to support training larger models and deploying them to a growing user base, data centers and power providers are struggling to keep up with rapid demand. In particular, permitting requirements for power generation and transmission are causing delays in the data center growth required for maintaining U.S. compute dominance. This may lead companies to build future data centers outside of the United States, which could compromise the U.S. lead and thus diminish its position in AI.
Adversaries will seek to hold U.S. data centers at risk. As AI becomes a general purpose technology that permeates our society, economy, health sector, education, and national security, it will undoubtedly become a target for adversarial attacks, both cyber and physical. While these attacks will most certainly be directed against the entire AI stack, data centers will likely become high value targets — given their importance as single points of failure (or success) for AI development and deployment.
Perpetuating the U.S. Compute Advantage
The central role of compute in AI development means the United States must consider a more deliberate approach to perpetuating its dominance. Some core elements of such an approach would include:
Establishing a coherent strategy and promulgating clear guidance on AI chip exports to countries that are not U.S. allies. The Bureau of Industry and Security (BIS) at the Commerce Department recently updated its AI chip export controls, restricting access to additional production equipment for advanced AI chips that China and other adversaries could use for military modernization. However, these restrictions still lack clarity or safeguards for exports to countries that fall between clear allies and adversaries. This is particularly troublesome at a time when American companies are exporting an increasing volume of chips, including to destinations where their end use is unclear or not policed. In some instances, foreign recipients may have close ties to China and may not respect end-user concerns of the United States. To prevent such leakage and further erosion of export controls, the U.S. government should:
Track all AI chip exports, including re-exports by foreign countries. Such tracking would not only help strengthen export control, but it would also have the benefit of generating a better understanding for the United States on global distribution of compute power and advantage.
Ensure technology exports align with U.S. interests. As with some of the other technologies or dual-use products, the U.S. government should establish a predictable process and designate offices that regularly re-evaluate if exporting a powerful dual-use technology to non allied nations is in the U.S. interest. As AI chips become more closely linked to military power, this process could also involve relevant Congressional oversight committees, similar to exports of defense articles.
Requiring foreign companies to monitor access of end-use and reports on customers and activities on AI chips. To prevent unauthorized re-exportation or third-party use of AI chips, or provision of AI cloud compute to China or Russia, countries other than close U.S. allies should implement know-your-customer procedures, agree to monitoring access, and provide regular reports to U.S. authorities. This could help ensure that U.S. technology is not used against U.S. interests and would allow the U.S. government to quickly react to unauthorized use.
Ensure adequate resources for export control enforcement. As production of advanced AI chips has increased to millions of chips this year, reports of chip smuggling have become widespread. While smuggling is always a phenomenon, part of its explanation lies in the fact that BIS is understaffed and underresourced for its growing national security mission.
Actively promote favorable conditions for data center construction in the United States. The United States must ensure that data centers continue to be rapidly built within its borders. However, as this will require large amounts of new energy, a significant streamlining of permitting requirements will need to occur to avoid crippling delays in crucial energy projects. SCSP recently hosted an AI + Energy summit to generate ideas on how to avoid energy becoming the limiting or pacing factor of AI development.
Data Centers for Defense
Preserving and perpetuating a U.S. advantage in AI infrastructure is also vital to America’s national defense.
The U.S. national security enterprise lacks anything comparable to private companies’ capacity in AI infrastructure. The most advanced AI systems require tens of thousands of AI chips for both training and deploying them on a large scale. Yet, according to unclassified data, DoD’s AI compute infrastructure is miniscule compared to large AI labs. The Air Force Research Laboratory’s Raider that came online last year hosts 128 NVIDIA A100 GPUs. Compare this to xAI’s Colossus, a supercomputer containing 100,000 of NVIDIA’s next generation H100 GPUs. Google, Meta, and Microsoft likely all have similar resources. This scale of supercomputers outpaces even the Department of Energy’s largest machines, such as Argonne National Lab’s Aurora supercomputer.
DoD relies on private companies for much of its compute power. Lacking its own large-scale AI compute infrastructure, DoD has had to rely on private companies to leverage AI advances for its mission. In 2022, the Pentagon made a $9 billion cloud deal, splitting services across Google, AWS, Microsoft, and Oracle. DoD has also recently begun partnering with major AI companies such as OpenAI and Anthropic.
Commercial data centers need protection against attacks. Leopold Aschenbrenner, a former researcher at OpenAI, recently warned that a daring nation state operation could evade current physical access controls of data centers, or could gain remote access to sensitive AI systems. This means U.S. adversaries could potentially sabotage, infiltrate, and steal information from commercial U.S. data centers. This risk would be dramatically more acute if the data centers were located outside of the United States.
Increased importance of AI makes threats more severe. As data centers host critical military or intelligence services, and especially if the United States is the first country to develop AGI, U.S. data centers will likely become high value targets for America’s adversaries, including for deliberate targeting in war. In fact, data centers have already been a target of domestic terrorism before. New technologies like consumer drones now pose an additional threat, as shown in both Ukraine and Israel. Additional risks include supply chain attacks, especially given the complexity of the semiconductor supply chain, and even hypersonic missile attacks in the event of a high intensity conflict.
Given the capital intensive requirements of the current AI infrastructure build-out, it remains cost-prohibitive for the U.S. government to build and maintain its own compute infrastructure and develop cutting edge AI models. That could change as hardware prices fall and algorithms get more efficient. Until such time, however, while DoD continues to rely on private sectors’ compute prowess, it should at least put in place plans for enhanced data center security, such as:
Establish a public-private partnership on advanced data centers security. Commercial companies cannot provide layered defense for data centers running national security applications on their own. DoD and the U.S. intelligence community have the expertise and resources to augment infrastructure security against a range of threats and should explore partnerships with AI developers to enhance their protection. For instance, the U.S. government could work with companies to harden their data centers, and ensure their resilience to both physical and cyber attacks.
Ensure U.S. security agencies monitor potential threats against data centers, including insider threats, sabotage and supply chain risks. Given their dual-use nature, the FBI and other intelligence agencies must put similar emphasis on protecting them as they are on protecting production sites by major defense contractors.
Conclusion
AI chips and the data centers that host them are essential for AI development and deployment, and thus critical to U.S. leadership in AI. As AI models get increasingly capable and integrated into our economy, society, and national security, the U.S. government must better manage and protect this infrastructure. This could include limiting and overseeing exports of advanced AI chips to countries that fall between allies and adversaries, facilitating the expansion of energy supply for AI data centers, and improving the physical and cyber security of AI data centers. These measures will protect the U.S. compute advantage and ensure AI development aids U.S. national security and democratic values.