AI Companies' Capacity for Rapid Infrastructure Building Exposes Their Failure to Coordinate on Existential Safety Measures
Source: Gideon Lewis-Kraus. "Our A.I. Problem | The New Yorker." September 20, 2026. www.newyorker.com
The Gist
AI companies can build giant data centers almost overnight, proving they're capable of moving mountains when they want to. So why can't these same companies, whose leaders openly warn that AI could cause human extinction, simply agree among themselves on basic safety rules? The author argues this failure isn't because it's impossible—it's because competition and self-interest are getting in the way of the will to act.
Conclusion
The AI industry's demonstrated ability to rapidly mobilize massive capital and build unprecedented infrastructure proves that its failure to coordinate on basic safety measures against existential AI risk is a matter of choice and misaligned incentives, not genuine impossibility.
Premises
- AI companies have shown an extraordinary capacity to 'just do things'—building data centers and marshaling trillion-dollar capital investments faster than any government infrastructure project in decades
- AI lab executives themselves genuinely believe and publicly state that superintelligent AI poses serious extinction-level risks within roughly three years
- Despite this acknowledged danger, executives have failed to reach agreement on even voluntary safety frameworks like Amodei's proposed transparency and accountability standards
- The obstacles cited for this failure—commercial competition and geopolitical rivalry, framed as a prisoner's dilemma—are the same kinds of coordination problems these companies routinely overcome when building infrastructure
- Public anxiety and the data-center backlash reflect legitimate concerns about loss of control, compounded by broader political resentment that government and institutions can't 'do things' anymore
- Those who create technological dangers bear special responsibility to prevent the harms they generate, yet the industry suffers a credibility deficit in fulfilling that responsibility
Assumptions
- The capacity for unilateral corporate action (building infrastructure) is comparable to the capacity for multilateral coordination among competitors (agreeing on safety standards)
- Executives' stated beliefs about AI risk are sincere rather than strategic or exaggerated for other purposes
- Voluntary industry coordination, if achieved, would meaningfully reduce existential risk
- The primary barrier to safety coordination is willpower/incentive alignment rather than genuine technical or structural impossibility
- Government regulation is currently insufficient or absent, making voluntary industry action the relevant lever