A credentialed former White House staffer — someone who briefed actual presidents — sits in a Washington think tank running scenarios where the U.S. military bombs server halls in China to stop an AI that doesn’t exist yet from being built. That’s not a Reddit thread. That’s a CNAS policy report. Jacob Stokes, senior fellow at the Center for a New American Security and former national security staffer under Obama and Biden, published “Superpowers and AGI” on August 27, 2026. When South China Morning Post covered it, the response was the policy equivalent of a grenade rolling across a conference table. The reason it landed that hard: the U.S. lead over China in frontier AI is currently measured in months, not decades, according to Epoch AI data cited in the report. The broader stakes extend to critical infrastructure, as fights over data centers increasingly shape national security calculations.
The Escalation Ladder Nobody Wanted to See
The CNAS report doesn’t open with bombs — it builds methodically toward them.
Stokes isn’t recommending missiles as a first move. Credit where it’s due. The report maps what a “laggard state” — America or China — might consider if it believed its rival was days from artificial general intelligence. The menu escalates deliberately.
What the CNAS Report Actually Proposes:
- The U.S. currently holds a months-long lead over China in frontier AI models, per Epoch AI data cited in the report
- China’s strengths in robotics and “embodied AI” could close that gap through a surprise breakthrough
- Stokes calls on the Department of Defense and NSA to assess what intelligence would justify action
- Cyber sabotage is framed as potentially below the threshold of open warfare; kinetic strikes on data centers are described as the most dangerous, most escalatory option
- The report frames these scenarios as applicable to either superpower — Washington or Beijing
The idea of bombing data centers didn’t originate in D.C. briefing rooms. In 2023, AI alignment researcher Eliezer Yudkowsky wrote in Time that enforcing a global AI moratorium would require countries to “be willing to destroy a rogue data center by airstrike.” That line was fringe when it appeared, reported by DataCenterDynamics as the outer edge of AI-safety maximalism. Three years later, it’s scenario-planning inside a credentialed Washington think tank.
Ideas don’t always have glow-ups in the right direction.
You’ve Seen This Movie Before
The nuclear analogy Stokes reaches for is doing more work than it appears.
Stokes himself invokes nuclear weapons as the template. Policymakers must “work backwards” from AGI’s unique nature, he argues per South China Morning Post’s summary, just as earlier generations grappled with the bomb. That framing is doing enormous work. It positions a hypothetical future capability as a strategic given, then constructs escalatory options — including acts of war — around that assumption. For anyone who followed the pre-Iraq intelligence architecture, the logic will feel uncomfortably familiar.
To his credit, Stokes clarified in a social media thread in September 2026 that the report presents multiple options, framed as scenario analysis for either superpower — not a call to arms directed at Beijing. What South China Morning Post amplified globally, and what policymakers will remember, is that kinetic strikes on data centers now appear on a Washington think tank’s official menu.
AGI doesn’t exist. The U.S. lead is months, not a guarantee. Washington is already debating blowing up server halls over it. The question worth sitting with is what happens when the intelligence on AGI turns out to be wrong.
