Ben Thompson: Anthropic, the Pentagon, and the Limits of Private Power

What It Covers

Stratechery founder Ben Thompson analyzes the conflict between Anthropic and the Pentagon over AI safeguards, arguing that private companies building transformative AI technology will inevitably face government coercion regardless of legal or ethical positions, drawing parallels to nuclear weapons regulation and Cold War geopolitics.

Key Questions Answered

• •Private Power vs. State Power: AI companies that build sufficiently powerful technology cannot remain neutral toward governments. Thompson argues that when a private executive's decisions over transformative technology become consequential enough, those holding state power will treat non-cooperation as an existential threat — not merely a contract dispute — and act accordingly to eliminate independent power bases.
• •Nuclear Weapons Analogy: Dario Amodei repeatedly compares AI to nuclear weapons, but this framing carries an underexamined implication: governments never allowed private companies to control nuclear arsenals. If AI reaches that power threshold, expecting governments to respect private property rights over AI models is historically inconsistent with how states have handled prior transformative weapons technologies.
• •China-Taiwan Equilibrium Risk: Thompson argues that cutting China off from TSMC fabrication and NVIDIA chips creates a more dangerous equilibrium than controlled access. If the US develops dominant AI while China cannot, China's rational response becomes destroying TSMC — 70 miles off its coast — eliminating the shared dependency that currently deters military action against Taiwan.
• •Intel's Government Sales Model: Bob Noyce's early Intel strategy offers a framework for AI companies: sell to government as a customer, but design products for the mass consumer and business market. Consumer-scale volume funds the hundreds of billions in annual CapEx required for frontier AI — a scale no government contract alone could sustain — producing better results than government-directed development.
• •Democratic Process vs. Executive Discretion: When people argue Congress cannot pass effective AI legislation, they implicitly endorse unelected private executives making consequential societal decisions. Thompson frames this as a binary: either democratic institutions produce laws governing AI capabilities, or individual CEOs like Amodei become de facto unaccountable policymakers — a governance shift with significant long-term implications.