a16z Podcast

Emil Michael: Iran, Anthropic and the Future of AI at the Pentagon

Emil Michael (Undersecretary of Defense for Research and Engineering, Acting Director of the Defense Innovation Unit) argues that commercial AI vendors with restrictive usage terms embedded in critical military systems represent an active national security risk, not a policy debate. The urgency is structural: the Department found itself single-threaded on a vendor whose internal "constitution" was, in practice, governing American command and control.

1. Taking Inventory: From 14 Priorities to Six

Michael was sworn in May 2025. He found 14 critical technology priority areas, most unchanged for nearly a decade, written in language he describes as "technobabble" (e.g., "integrated network systems of systems"). He cut the list to six, moved the Chief Digital and AI Office into his group, and designated applied AI as priority one. Within 90 days, 1.2 million of the department's 3 million personnel had used some form of AI. The baseline when he started was 80,000 users.

2. The Anthropic Contract Problem

Reviewing contracts inherited from the prior administration, Michael found AI models baked into Central Command, Indo-Pacific Command, and Southern Command with dozens of usage restrictions: operators could not plan operations or move assets if the action might lead to a kinetic strike. More critically, the contracts were structured so the vendor could, in theory, shut off the software mid-operation if terms were violated. This was a single-vendor situation with no fallback.

The second trigger was a senior executive at the primary vendor asking whether their software was used during the Maduro raid, which Michael calls one of the most successful military operations in recent memory. He frames the question itself as disqualifying: a vendor auditing sovereign military operations against its own internal values document.

3. The Sovereignty Argument

Michael's core argument: AI is trending toward a general substrate that will touch everything, the way the internet or telecommunications did. Telling the military it cannot use that substrate for lawful operations, those passed by Congress and executed by the executive branch, is equivalent to letting a private company's internal document override the U.S. Constitution as the governing authority for command and control. He draws a direct comparison to adversaries: China has stolen American models, removed the guardrails, and may deploy them against U.S. forces. Restricting American warfighters against the same unrestricted model is, in his framing, an Orwellian outcome.

4. Procurement Reform: Fixed Price Over Cost-Plus

The structural fix Michael is pushing mirrors the SpaceX model. Old process: thousand-line RFPs, vendors check every box regardless of physics, cost-plus contracts, endless change orders, multi-billion-dollar overruns. New process: simple outcome requirements (range, payload, environment), industry proposes solutions, firm fixed price contracts, vendors keep margin if they economize. He is also pushing faster yes/no signals to startups, arguing the Pentagon's cultural habit of never saying no wastes founders' time and obscures real demand.

5. What the Department Needs From Startups

The gap Michael identifies is not invention: it is manufacturing scale. Primes hold an advantage in production capacity and quality assurance at scale, not in building new things. He expects startups to cross that chasm in one to two years, and says the signal to watch is not verbal enthusiasm from department contacts but whether they are actually buying or testing.

Key Takeaways

  • Michael cut the Pentagon's critical technology priorities from 14 to 6 and put applied AI first, then scaled active AI users from 80,000 to 1.2 million within 90 days, demonstrating that organizational focus, not budget, was the binding constraint.
  • The Anthropic controversy was not primarily about ethics: it was about vendor lock, where a single supplier with restrictive terms was embedded in the most sensitive military commands with no alternative and the contractual power to go dark mid-operation.
  • Michael's position is that AI is becoming an unavoidable substrate, like the internet, and that a private company's internal values document cannot be the authority governing U.S. military decisions made under laws passed by Congress.
  • The procurement shift from cost-plus to firm fixed price is designed to attract startups by rewarding efficiency and providing clear demand signals, but Michael says startups must build manufacturing and production capacity to compete with primes at scale.
  • The Google Project Maven moment in 2018, when employees protested a defense AI contract, is now viewed inside the department as a cautionary tale that Google itself learned from: Google is now a major government partner, and Michael is signaling that frontier AI companies resistant to defense work will face the same choice.
← All a16z Podcast episodes

Get summaries like this daily.

Subscribe to get personalized insights from a16z Podcast and 20+ other top podcasts.

Get Started — $12/mo