When the world's at stake,
go beyond the headlines.

Will AI eat the world and America’s defense budget? I think of those who toil at the intersection of AI and national security as being divided into three camps: Sprinters hold the most aggressive assumptions and believe profound disruption via artificial general intelligence is imminent; marathoners believe the technology will diffuse selectively, sector-by-sector; and skeptics draw analogies to the dot-com bubble.

America’s near-term AI strategy should align with one of these three approaches. If the sprinter scenario holds, the United States should go all-out to rapidly acquire artificial general intelligence — defined here as human-level intelligence. If the skeptics are right, however, then the United States should do virtually the opposite and avoid overbuilding and overextension. If the marathoners are most correct, then the United States will conduct a complicated, long-term technological competition with a country four times its population.

Adopting the skeptic approach is risky: AI is already a powerful tool. In addition to applying best AI competition practices, policymakers should adopt the marathoner approach for now but maintain flexibility. The marathoner approach will allow Washington to adjust AI efforts as conditions warrant, minimizing the risks of both overreach and underinvestment.

Defining the Camps

Trade-offs between AI and other priorities are already necessary: U.S. private sector investment in 2024 totaled $109 billion, and aggressive estimates hold capital spending could reach $2.35 trillion by 2030. Mapping the three AI camps helps policymakers determine whether Washington’s $3.3 billion Fiscal Year 2025 spend on AI research and development merits a sharp increase or a cautious pause.

Sprinters

This camp believes that AI is on a rapid trajectory toward artificial general intelligence. They foresee world-altering and nearly immediate consequences: initial advantages will unleash enormous and self-reinforcing productivity gains. In this view, the country that first obtains artificial general intelligence will secure enduring — and likely permanent — geopolitical advantages. Similarly, artificial general intelligence’s “inventor” could become the world’s first trillionaire.

Adherents in this camp include U.S.-based technologists like Sam Altman, Dario Amodei, and Elon Musk. In the United States, artificial general intelligence evangelism (and doomerism) carries upside for tech hyper-elites. If the sprint pays off with usable artificial general intelligence, the spoils will be historic. If the marathon approach holds, however, they will still enjoy first mover advantages.

It would be a mistake to dismiss the sprinters solely because of tech actors’ conflicts of interest, however. A slice of the U.S. national security community, like Ben Buchanan, fall into the sprinter camp due to their technical assessments and wariness of the Chinese Communist Party’s secretive nature. Indeed, recent AI advances have consistently outstripped experts’ forecasts. Some analyses of artificial general intelligence arriving before 2030 are exceptionally well-argued. Additionally, as Julian Gewirtz notes, China may be covertly committed to obtaining artificial general intelligence before America. Sprinters’ assessments may be informed by non-public indications of an imminent breakthrough.

Still, few Chinese actors could be categorized as sprinters, based on their public actions — although DeepSeek is an important exception. In China, technologists are disincentivized from promoting artificial general intelligence. Even if privately bullish, they understand that the party views artificial general intelligence as potentially politically destabilizing, especially if it leads to alternative power sources.

There is little public evidence, however, that the Chinese state regards an artificial general intelligence breakthrough as a serious medium-term threat necessitating extreme measures. China is building fewer AI-directed data centers than the United States and many of these facilities are reportedly unused. Chinese policymakers “are far more concerned about near-term diffusion and large-scale adoption,” over artificial general intelligence acquisition, Will Rinehart writes. Chinese AI policy arguably formally endorses the view that the AI competition will be a marathon rather than a sprint.

Marathoners

Marathoners are the dominant camp among the expert community in both the United States and China. Adherents include many Chinese technologists and many U.S. “China hands” like Harry Krejsa, William Hannas, Kyle Chan, Bill Drexel, Elsa B. Kania, and Jordan Schneider.

They see AI not as a singular leap to artificial general intelligence, but as a broad, sector-by-sector process of significant but incremental improvement. Marathoners acknowledge disruption is coming but believe change will be uneven and context-dependent. This camp believes that AI will reshape key industries — perhaps finance, insurance, health, and transportation — but without triggering artificial general intelligence by 2030, or possibly ever.

Marathoners hold that AI inference — that is, the application of AI models by an end user — will prove more important than model training, or creating evermore robust models. Unlike sprinters, who prioritize developing the best AI model via training, marathoners maintain that a “maximum-training-minimum-inference” approach is “primed for change.”

Skeptics

Skeptics believe artificial general intelligence is decades away, if it arrives at all. Figures like Arvind Narayanan, Sayash Kapoor, and Gary Marcus caution that framing AI as a moonshot misallocates capital. They warn against artificial general intelligence comparisons to nuclear weapons, citing the slow pace of real-world deployment, which echoes past tech bubbles.

Dan Wang, a China-focused analyst, offers a related but unique view: Beijing may prioritize political control over innovation. While China has a poor safety record with many technologies, it may pursue “party first” AI development rather than unconstrained innovation.

The skeptics will likely be proved right in some domains, but AI tools have rapidly made profound improvements, are already valuable, and will very likely become even more powerful.

Camp-Specific Strategies

Each approach brings different strategies, trade-offs, and risks. The skeptics’ camp is the most risk-accepting. While their approach limits overbuilding risks, there are few hedging instruments if China’s AI capabilities begin to outstrip America’s. Sprinters can diffuse risks, to a degree, but the marathoner approach maximizes flexibility and hedging.

Sprinters

If the sprinter scenario holds, the United States should go all-out to achieve artificial general intelligence before China. Training AI models will be the most important AI workload. Accordingly, the United States should rapidly site data centers and take an all-of-the-above approach to resourcing their electricity needs as quickly as possible and prioritize “speed-to-power”; hoard advanced chips while restricting China’s access; take active measures to stunt Chinese artificial general intelligence development; and protect key personnel and infrastructure — such as the Spruce Pine facility that produces ultra-pure quartz for semiconductors — from Chinese offensive actions. If the sprinter dynamic holds, policymakers should expect Beijing to act aggressively — even violently — because whichever country first secures artificial general intelligence will very likely become the permanent superpower.

This approach is high-risk, high-reward. Given artificial general intelligence’s overwhelming importance, sprinters require the United States to deprioritize every other long-term national security investment. Otherwise, the United States could lose the most strategic technology in history. Tangibly, a sprinter approach could mean dialing back support for, or even scrapping, the Next Generation Air Dominance fighter or the Next-Generation Attack Submarine programs. Accordingly, if the sprinters’ “all-in” bet on AI and artificial general intelligence doesn’t materialize, at the cost of hundreds of billions of dollars, China could cement its industrial dominance and establish technological leadership.

To hedge against an artificial general intelligence bubble while adopting the sprinter approach, the United States could encourage cost-sharing by allies and partners, loosening chip controls on the Middle East, and encouraging data center construction abroad. Cost-sharing diffuses financial risks but could also enable Chinese companies to access high-end chips needed to train AI models.

Marathoners

If the marathoner approach holds, AI adoption — rather than artificial general intelligence acquisition — will prove decisive. Accordingly, while speed is still important, sectoral adoption and controlling costs should be prioritized. Instead of training AI models, the United States should concentrate on optimizing AI workloads for inference — that is, applying already-built models.

Distributed inference workloads will necessitate diffusion. Since inference workloads are relatively energy-weighted compared to AI model training, AI developers will be more selective and methodical about data center deployment than in the “sprint” scenario. Rather than constructing massive data centers, the United States will need to site smaller data centers closer to inference demand.

Electricity generation sources will need to align with distributed inference demand. Advanced nuclear reactors capable of providing baseload power hold outstanding promise for inference applications (and military microgrids) but have yet to be deployed at scale or cost efficiently. Given that substantial deployment of advanced reactors will likely not occur until the 2030s, other near-term solutions are needed. Solar holds unique appeal: it is low cost, can be deployed on virtually any rooftop, output crests during the summer, and its diurnal electricity generation profile aligns well with working hours — when inference demand will likely peak. Incorporating batteries and expanding the grid will substantially enhance the reliability of solar energy solutions for inference workloads. Still, sunlight intermittency will constrain its reliability. To resource AI electricity demand, the United States will need to adopt a patchwork approach, leaning on solar, batteries, and natural gas to meet short-term incremental demand; resourcing next-generation technologies (such as advanced reactors and geothermal) over the medium-term; and relying on its existing fleet of natural gas, nuclear, and, when necessary, coal plants to prevail in the AI competition with China.

Advanced batteries will also prove crucial in the marathoner scenario, given the need for on-device or on-board AI inference for autonomous vehicles and platforms. Working with allies and partners like South Korea, Japan, and Taiwan will prove critical in developing U.S. advantages in advanced batteries’ dual-use applications.

America’s most valuable technological partner in the marathoner scenario may well prove to be India — at least in some respects. The world’s largest democracy will be disproportionately critical for supplying much of the brain power and human capital used in the AI competition with China, although the United States will need to identify and recruit leading AI researchers from everywhere. Cooperation with other countries — especially traditional allies — will also prove crucial.

While marathoners strike a middle ground between the other two camps, this approach brings trade-offs and carries risks. Marathoners can hedge against a potential AI overbuild via other technologies, such as quantum. Alternatively, they could pursue pre-sprinter activities, including laying the groundwork for a massive, sovereign-scale, artificial general intelligence-directed data center campus — something China may be considering for its vast “Beijing Military City” complex. Fortunately, the marathoner approach is a natural hedge and allows for scaling AI efforts up or down based on emerging realities.

Skeptics  

If the skeptics are right, AI’s primary risk is overbuilding and overextension. Overhyping AI could lead to stranded infrastructure, inflated energy demand, or even financial contagion from an AI investment bubble. Skeptics believe in committing scarce resources only where adoption holds indisputable benefits and oppose funding “moonshots” like artificial general intelligence.

Skeptics hold that America should prioritize non-AI national security capabilities. Accordingly, given the likelihood of continued Chinese technological improvements, America’s diminishing qualitative edge will necessitate investments in quantities of traditional national security capabilities, such as shipbuilding and standoff weapons.

The skeptic camp advocates a low-cost approach — but at a huge risk of miscalculation. If skeptics are right, then the avoided costs will be substantial. If skeptics are wrong, however, then the United States will fall behind in a potentially decisive technology where benefits are uniquely self-reinforcing. Accordingly, the United States would need to either accept Chinese technological domination or scramble to catch up by investing even more than it would have committed under the sprinter or marathoner approaches.

Skeptics advocate for the riskiest approach: this camp has few hedges beyond monitoring AI developments. However, AI has inherent recognition and observability lags, and its development could continue to be exponential: Inference costs for an AI performing at the level of GPT-3.5 fell 280-fold between November 2022 and October 2024. If the United States applies a skeptic approach incorrectly, the consequences will be devastating.

Listen to this episode of The Insider featuring author Joseph Webster by becoming a War on the Rocks member here

Divergences

The three camps hold different views on the appropriate policy mix for the Sino-American AI competition. While the United States should apply universal best practices for its AI competition with China, it will also need to make consequential choices across four key policy areas: resource allocation, talent, infrastructure and energy, as well as alliances. Which policy playbook Washington follows may decide who dominates AI.

Sprinters direct all available resources to the existential artificial general intelligence competition, even if that means delaying next-generation military technologies. Marathoners hold that AI is a highly important technology that should be prioritized on a case-by-case basis. Skeptics, meanwhile, apply a “guilty until proven useful” standard to AI.

On AI talent, sprinters prioritize immediate visas for elite AI talent; marathoners scale domestic science and technology education while deepening talent pipelines with like-minded partners, perhaps especially India; and skeptics hold that AI’s risks — especially surrounding misinformation — require broad digital-literacy upskilling.

Sprinters hold AI’s energy and infrastructure requirements entail constructing massive data campuses to train AI models, building all types of new energy as quickly as possible, and even constricting competing energy demands — such as cryptocurrency. Marathoners also take an “all-of-the-above approach” to energy but emphasize inference-directed energy generation, especially solar and batteries. They also emphasize long-term grid buildout, especially advanced reactors and transmission. Skeptics, meanwhile, worry about overbuilding inefficient generation.

The camps also adopt different alliance strategies. Sprinters push for strict chip embargoes but think artificial general intelligence development could entrench U.S. leadership, granting Washington enduring leverage. Marathoners, meanwhile, focus on building a long-term ecosystem of like-minded AI partners, prioritizing traditional allies while cultivating partnerships with India and other potential “swing powers” — and AI talent sources — across the Middle East, Africa, and beyond. Finally, skeptics view allied chip controls as potentially counterproductive, given their hope for Chinese AI overinvestment.

Expect a Long Run, Be Ready For a Sprint  

Adopting the marathoner approach best serves U.S. interests, for now. AI is already a powerful tool and will very likely become more capable, suggesting limitations to the skeptic camp’s approach. Still, artificial general intelligence acquisition appears unlikely in the near-term, diminishing the appeal of sprinters. Crucially, the marathoner approach allows the United States to scale AI efforts up or down in accordance with real world developments.

Regardless of which camp one falls into, however, U.S. policymakers should adopt general, scenario-agnostic recommendations. American AI companies and U.S. security services should deepen full-spectrum cooperation. Additionally, as labor is typically the most important cost driver for training AI models, the United States should maintain access to high-skilled AI-relevant labor, both domestically and from abroad, and resource its leading universities. While reducing exposure to Chinese supply chains is critical, the United States should be very careful about applying tariffs on AI-relevant goods from trusted partners, as these measures raise costs for American companies and increase China’s ability to obtain advanced technologies via trade diversion. The United States should adopt a pragmatic, all-of-the-above approach to energy, which could be a critical bottleneck for American AI efforts. Finally, policymakers should recalibrate AI approaches based on the latest developments.

AI will be a critical and perhaps defining element in the Sino-American competition. The United States should do everything it can to ensure the right side prevails by applying best practices and adjusting policy for whatever camp — sprinters, marathoners, or skeptics — proves most accurate.

Joseph Webster is a senior fellow at the Atlantic Council’s Global Energy Center and Indo-Pacific Security Initiative, and editor of the independent China-Russia Report. This article reflects his own personal opinion.

Image: Charlie fong via Wikimedia Commons