Notes — Christopher Kirchhoff on Military Innovation and the Future of War

Notes on Christopher Kirchhoff in conversation with Tyler Cowen — Conversations with Tyler (https://conversationswithtyler.com/episodes/christopher-kirchhoff/), 30 October 2024.

Four questions [Adler frame]

Q1 — What is it about as a whole? Kirchhoff, co-founder of the Pentagon’s Defense Innovation Unit (DIU) and co-author of Unit X, argues that the character of war is changing faster than the American national-security bureaucracy can adapt, and that the bottleneck is institutional, not technological. He ranges across drone warfare’s near-total takeover of the Ukrainian and Middle Eastern battlefields, hypersonic weapons and EMP threats, the AI arms race with China, why defence-primes consolidation after the Cold War produced slow-moving ‘utility companies’ rather than innovative firms, and why DIU itself had to fight bureaucratic and even petty resistance (frozen government credit cards, an eviction threat over a dog policy) to get built at all. The throughline is that Silicon Valley’s iteration speed has become the decisive military variable, and the Pentagon’s acquisition and personnel systems are not built to use it.

Q2 — How is it argued? Almost entirely through first-hand institutional narrative and named case studies rather than abstract argument: DIU’s own founding (Ash Carter’s 2001 ‘Keeping the Edge’ thesis, Lauren Dailey’s discovery of the ‘other transaction authority’ loophole in the National Defense Authorization Act), the Mike Brown whistleblower episode, the Colonel Fear dog-policy standoff, and battlefield evidence from Ukraine, Gaza, and Lebanon. Cowen’s questioning is rapid-fire and concrete — asking for numbers, timeframes, and mechanisms (‘how many people die?’, ‘is it laws or cultural inertia?’) — which keeps Kirchhoff anchored to specifics rather than doctrine.

Q3 — Is it true, in whole or part? Kirchhoff speaks from direct institutional experience (DIU co-founder, National Security Council staff, National Security Commission on Artificial Intelligence, Anthropic policy residency), which lends his organisational claims — DIU’s origins, the Mike Brown case, the $70 billion figure for other-transaction-authority contracts — high credibility as first-person testimony, though none is independently verified within the conversation. [?] His battlefield claims (Ukrainian drone-loss rates on Abrams tanks, the 120 Chinese academic papers on the Ukraine war, Chinese hypersonic test success rates) are stated with more precision than public sourcing allows a listener to check. [?source] His forecast that AI diffusion will not replicate nuclear weapons’ ‘zero-to-one’ strategic discontinuity is a considered but explicitly speculative judgement, not a settled fact.

Q4 — What of it? A first-hand account of the same ‘bureaucratic rigidity cycle’ diagnosed from the civilian-government side in Jennifer Pahlka on Reforming Government — a new reform unit (DIU) gets created but the old defence-procurement infrastructure is never dismantled, so the new unit is added on top rather than substituting for anything. It gives the wiki concrete texture — Lauren Dailey’s loophole, the Mike Brown whistleblower case, the F-35-versus-iPhone contrast — for the abstract claim that commercial technology now outpaces military R&D, and it extends the wiki’s AI-policy material with a national-security-specific angle on dual-use technology and the China AI race.

Glossary

Defense Innovation Unit (DIU) — the Pentagon’s Silicon Valley outpost, founded in 2015 by then-Secretary of Defense Ash Carter and built out by Kirchhoff and his co-author Raj Shah; its mission is to get the US military to adopt commercial (not bespoke defence-industry) technology at commercial speed. [§ On Unit X]

Other transaction authority — a little-used, Apollo-era contracting mechanism that lets DIU sign deals with start-ups in 30–60 days instead of the 18–24 months typical under the Federal Acquisition Regulation; discovered by DIU acquisition specialist Lauren Dailey in a clause of the National Defense Authorization Act and used to acquire $70 billion of technology in the eight years since. [§ On Unit X]

Replicator initiative — a Pentagon push, led by DIU and announced more than a year after Russia’s invasion of Ukraine, to develop autonomous swarming drones for air, sea, and undersea use; Kirchhoff calls it ‘directionally right’ but under a percent of Pentagon procurement spend — too small for the scale of the shift it responds to. [§ opening, on drone warfare]

Military-Civil Fusion — Xi Jinping’s doctrine, announced around 2015, requiring any Chinese company producing dual-use technology to notify the People’s Liberation Army and potentially accept a PLA-nominated board observer, so militarily useful commercial technology is captured by the state without a single dedicated acquisition unit like DIU. [§ on AI as an arms race]

Valley of Death — the acquisition-pipeline gap between a promising early-stage defence technology demonstration and an actual procurement contract, where good ideas stall for want of an institutional buyer willing to commit before there is a proven product. [§ on accountability and bureaucracy]

Strategic stability (literature) — the Cold War-era political-science field studying whether the spread of a powerful weapons technology among states increases or decreases the likelihood of great-power conflict; Kirchhoff invokes its consensus finding — that nuclear proliferation among states has, on net, lessened great-power war — against the ‘Stop Killer Robots’ movement’s case for unilaterally restraining autonomous-weapons development. [§ on AI as an arms race]

Key claims by section

Opening — on drone warfare [§ opening, on drone warfare]

The 2024 destruction of a quarter of the M1A1 Abrams tanks the US gave Ukraine by Russian kamikaze drones marks, in Kirchhoff’s reading, the end of man-mechanised land warfare, not just a tactical setback.
Ukrainian drone start-ups — nearly 300 of them — iterate faster than Western defence primes because they can rewrite drone software overnight in garage workshops to counter evolving Russian electronic-warfare tactics.
Ukraine has, in effect, replicated the US Air Force’s entire battle-management system (surveillance drones feeding targeting data to attack drones via open-source fusion software) at a fraction of Western-arsenal cost.
Drones favour whichever side is trying to defeat a conventional military operating conventionally: Hamas’s 7 October border breach, Houthi harassment of shipping, and Hezbollah’s depopulation of northern Israel (85,000 civilians displaced) despite the IDF’s modernity, are all cited as instances.
The US has no adequate early-warning system tuned to detect small drones — the same gap that let the Chinese spy balloon go undetected — and ‘very few if not zero’ US installations are well defended against a surprise drone attack.

On hypersonic weapons [§ On hypersonic weapons]

Hypersonic weapons are strategically distinct from ICBMs on two counts: manoeuvrability (they can change trajectory mid-flight, defeating space-based launch-detection systems tuned to ballistic paths) and compressed decision time, straining the continuity-of-government timeline built around a 20–25 minute ICBM warning window.
There is currently no defence against them; Kirchhoff expects them to shift the military balance of power within five years, independent of their nuclear-payload capability.
China has demonstrated fuller hypersonic flight cycles than other states in public testing; Russia is developing a broader portfolio of exotic weapons, including a proposed nuclear payload on a satellite intended to disable GPS and satellite communications, and supercavitating torpedoes capable of destroying an aircraft-carrier group.
An EMP or neutron-weapon strike would disable the electrical grid and most microelectronics in the blast radius without the mass casualties of a conventional nuclear detonation — Kirchhoff’s flip answer, ‘cooking stoves’, points to the strike’s civilian-infrastructure character rather than a body count.

On AI as an arms race [§ on AI as an arms race]

Kirchhoff respects the ‘Stop Killer Robots’ movement (Human Rights Watch and computer scientists including Stuart Russell) but judges its argument weak on strategic-stability grounds: unlike nuclear or chemical weapons, software-based autonomous-weapons development cannot be verified, so unilateral restraint would not produce mutual restraint — it would simply cede advantage.
The US’s current generative-AI lead over China rests on structural advantages — freer markets, less state suppression of the tech sector — rather than superior state direction; China’s Military-Civil Fusion doctrine nonetheless channels dual-use commercial technology into the PLA without the domestic-censorship constraints that slow China’s population-facing AI deployment.
Kirchhoff argues against nationalising frontier AI labs (a proposal he says has surfaced from former OpenAI-adjacent voices), citing national-labs culture as comparatively moribund next to the commercial tech sector, and favours instead close FBI/intelligence-agency cooperation to protect private labs from nation-state theft — modelled on the 2000s-2010s Enduring Security Framework built for critical-infrastructure cybersecurity.
He is ‘firmly in favour of dual use to the absolute extent we can’, reasoning from his NSC and Pentagon experience that restricting AI’s civilian upside to capture a narrower military-only slice would leave the US weaker in AI overall — the same logic he applies to encryption: net societal benefit outweighs the harm from bad actors’ access.
He does not expect AI diffusion to resemble nuclear weapons’ sudden ‘zero-to-one’ strategic discontinuity; he expects a more gradual, capability-graded diffusion (an ‘iPhone 15 versus iPhone 12’ dynamic) that gives states more time to find new equilibria without triggering sudden destabilisation.

On Unit X [§ On Unit X]

DIU traces to Ash Carter’s 2001 article ‘Keeping the Edge’, which observed that federal R&D spending was flatlining exactly as the commercial technology economy took off — by the time Carter became Secretary of Defense in 2015, Apple, Microsoft, Google, and Amazon individually exceeded the market capitalisation of the entire US defence industry.
Kirchhoff ran the working group that designed DIU under Carter, then co-led it with Raj Shah from 2016; its central innovation was procedural — using ‘other transaction authority’ to contract with start-ups in 30–60 days rather than 18–24 months, discovered by 29-year-old acquisition specialist Lauren Dailey in a clause of the National Defense Authorization Act.
$70 billion of Pentagon technology acquisition has used Dailey’s mechanism in the eight years since.
DIU deliberately staffs on rotation, DARPA-style, to avoid the institutional ossification Cowen presses him on; Kirchhoff frames this as one round of a recurring ‘Hegelian dialectic’ in defence-agency history (DARPA itself spent years in the wilderness after its founding before becoming dynamic).
China has no single DIU-equivalent unit; instead, its Military-Civil Fusion doctrine mandates dual-use technology disclosure to the PLA across the whole commercial sector — a structurally different, state-directed route to the same end.

On accountability and bureaucracy [§ on accountability and bureaucracy]

Executive authority inside government has eroded to the point that even a lunch talk at a think tank required Kirchhoff, as an NSC aide, to file an ethics waiver.
The inspector-general system can be ‘weaponised’ against effective public servants: a whistleblower complaint against DIU’s second director, Mike Brown, over paying Silicon Valley engineers near-market rates and a minor office-snack-fund accounting irregularity, was investigated and cleared internally (an 82-page report), yet resurfaced in the media on the eve of his Senate confirmation as Pentagon acquisition chief and forced his withdrawal despite no substantiated wrongdoing.
Post-Cold War defence-industry consolidation, driven by Bill Perry’s 1993 ‘Last Supper’, produced five large primes that Kirchhoff likens to utility companies rather than private firms — burdened by requirements and audit regimes that produce slow innovation and cost overruns.
Newer product-oriented defence start-ups (Anduril, Shield AI, Palantir, SpaceX) are winning contracts once assumed to go to the primes — Anduril and General Atomics, not a traditional prime, are the two finalists for the Air Force’s 10,000-aircraft Collaborative Combat Aircraft programme.
DIU itself faced small but real institutional resistance when founded, including government credit cards being shut off out of spite before a first Washington trip, and a colonel (‘Colonel Fear’) threatening eviction over DIU’s dog-friendly office policy — resolved by reclassifying the pets as HIPAA-protected service animals.

On government commissions [§ On government commissions]

Kirchhoff’s Cambridge political-science dissertation on government commissions, following early-career work on the Space Shuttle Columbia accident investigation and the Iraq reconstruction history, taught him that expertise and ‘settled notions of how business should be done’ are often wrong — especially when they get people killed.
He divides commissions into three types: disaster-management commissions (created to deflect political blame, rarely effective), agenda commissions (like the National Security Commission on Artificial Intelligence — created to build expertise and bipartisan consensus on a complex, ongoing issue, not in response to a single crisis), and crisis commissions (like the 9/11 Commission — created after an acknowledged rupture, when openness to fresh diagnosis is highest and adoption rates for recommendations are correspondingly high).
He remains a strong advocate for commissions as an instrument of government, noting the British government uses them more freely than the US does.