Key Takeaways

  • White House AI export restrictions trace a policy spectrum from harmless consumer goods like Coca-Cola to existential threats like selling nuclear weapons technology to Germany or the Soviet Union during World War II and the Cold War.
  • Export policy on advanced computing relies on hard technical thresholds, specifically manufacturing node sizes and interconnect bandwidth metrics, to prevent adversarial military use without killing commercial trade.
  • Sanctions on advanced hardware originated under Donald Trump's first administration, expanded into complex layered controls under Joe Biden, and maintained their core restrictions into Trump's second term.
  • Chinese officials and tech executives have shifted focus over the past two years toward catastrophic risks surrounding artificial general intelligence and superintelligence.
  • Strategic policy planning rests on Collins's Three Struggles of Democratic AI Governance, a method balancing competitive technical leads against state capacity and institutional stability.

The Collins's Three Struggles of Democratic AI Governance

Teddy Collins helped formulate the semiconductor export controls presented to the president and national security advisor during the Biden administration. In his book, The Bitter Struggle, Collins details the high-stakes trade-offs democratic societies face when managing computing power. His method organizes these choices into three sequential challenges:

Struggle 1: Capability Leadership

The struggle for democratic nations to establish and maintain a lead in capabilities.

Struggle 2: Government Adoption and Technical Capacity

The struggle for governments to understand and make use of this technology effectively, because right now there is a big gulf between the private sector and the government.

Struggle 3: Democratic Alignment and Systemic Stability

The struggle then to make sure that the technology doesn't become deeply unstate and doesn't take us away from democratic values once we've mastered it.

When This Works (and When It Doesn't)

This framework guides national security teams when determining where to draw export boundaries between ordinary commercial software and restricted dual-use assets. It works when technology development moves linearly along measurable hardware bottlenecks. Setting limits on silicon node sizes and interconnect bandwidth creates a tangible checkpoint that governments can inspect, trace, and restrict at physical foundries.

The framework fails when capability leaps decouple from physical hardware. If smaller open-source models match frontier model output through synthetic data or algorithmic breakthroughs, export controls on physical compute lose their grip. It also stumbles when the technical gap between Silicon Valley and Washington widens. If civil servants cannot evaluate algorithmic changes or training cluster designs, they end up regulating outdated hardware metrics while missing active proliferation vectors.

What to Do With This

If you run an AI startup touching dual-use capabilities, map your product pipeline against Collins's three struggles this week.

First, check your compute architecture against current trade regulations. If your system depends on distributed clusters or specialized interconnect hardware, verify whether your infrastructure providers face pending export license rules under current Commerce Department revisions. Do not wait for a formal subpoena or an abrupt cloud provider cutoff to learn your network bandwidth crosses an export restriction threshold.

Second, audit your human-in-the-loop workflows. Collins's current venture, the London-based neolab Inherent, rejects the idea of treating research as something to automate in a closed box and hand to autonomous agents. Collins treats scientific discovery as a human-machine teaming problem that requires human judgment and taste. If your product roadmap automates away every human touchpoint, reintroduce structured interfaces where domain experts evaluate intermediate model outputs. Teaming human taste with machine speed prevents drift, satisfies safety audits, and creates defensible enterprise value.