War with Iran + Pentagon vs Anthropic with Under Secretary of War Emil Michael
In a Nutshell
US and Israel launched Operation Epic Fury against Iran, killing Supreme Leader Khamenei and depleting 90% of its munitions in days, aimed at dismantling nuclear sites, drone/missile production, and terror funding without boots on ground; Trump leverages this alongside Venezuela raid to pressure China on oil supplies for trade/Taiwan deals. Pentagon cancelled $200M Anthropic contract after failed negotiations over AI restrictions on military uses like drone swarms and Golden Dome missile defense, deeming it a supply chain risk due to ideological guardrails—favoring flexible providers like Grok, Google, and OpenAI for lawful DoD needs. Defense shifts to AI-driven drones, relaxed ROE, requirements reform for startups (Anduril/Palantir as new primes), and DARPA innovations to counter China while rebuilding US supply chains.
These notes were generated by AI and may contain inaccuracies.
Emergency podcast, Episode 263 of All-In. Emil Michael, Under Secretary of War for Research and Engineering, working directly for Pete Hegseth. Emil Michael served as Travis Kalanick's right-hand man at Uber, raising $20 billion.
Chamath Palihapitiya and David Friedberg join. Jason Calacanis hosts. David Sacks is out, busy on Capitol Hill.
US and Israel launched joint attack on Iran on Saturday. Today is day six of Operation Epic Fury. Iran Supreme Leader Ali Khamenei killed within hours. 40 senior officials killed. Death toll about 1,000 people. Six US Army Reserve soldiers killed in drone strike on base in Kuwait. US submarine sank Iranian ship off coast of Sri Lanka—first torpedo kill since World War II.
Explanations for war: Rubio said Israel was going to attack, US had no choice but to participate, later walked it back. Trump: not regime change, but to stop terrorism, ICBM development, nuclear bombs (blew up nuclear sites couple weeks ago). Trump: people of Iran should seize moment and take country back. Pete Hegseth: not regime change war, but regime did change, world better off.
Polymarket odds:
- US forces enter Iran (boots on ground) by end of March: 40%; by end of year: 59%.
- Iranian regime falls by June 30: 39%; by end of year: 51%.
Emil Michael: President said weeks not months operation, aimed at disarming regime from supplying Hezbollah, Hamas, Muslim Brotherhood with weapons/money, plus nuclear program. Targeting depots, nuclear sites, prodigious drone production (one-way attack drones hundreds of miles), ballistic missiles aimed at Middle East countries.
No protracted boots on ground like Afghanistan/Iraq 2.0.
Friedberg: Trump prioritizes grand bargain with China, meetings in April. Actions in Iran and Maduro create leverage—90% of Iran's oil goes to China; Maduro's Venezuela has long relationship with China, big economic drivers for China. Disrupting supply chains gives US better footing for deal on American businesses selling into China, trade parity, Taiwan, semiconductors. Could be win-win, crowning achievement pre-midterms.
Chamath: Not regime change or regional conflict. Taking out Maduro and Iranian leadership creates enormous leverage with China. Chinese bureaucracy posted GDP targets of 4.5-5%, lowest in 30 years (pre-WTO). From 8-10% growth to half, with double population/GDP, high youth unemployment—becomes more chaotic. China's interest: invade Taiwan to occupy people/resources, boost production/war machine.
Chart shows preventing war with China: China 20% dependent on imported oil (logistics, transportation, aviation, feedstocks); 1/5 from Iran/Venezuela (now off table). With Russia deal (Whitoff, Kushner, Greenbond), ~40% of China's oil. Redollarize, stop illicit oil funds, hem in China for grand bargain end March/early April, secures global safety.
Emil: China second-order benefit. Not 8 conflicts—inherited Gaza, Russia-Ukraine; Venezuela one raid (few hours) plus drug boats; Houthies (Biden ignored ships being shot at)—freedom of seas. Iran material conflict outside Venezuela.
New approach: no boots on ground. Venezuela went well unexpectedly; Iran so far very well.
Success factors: well-trained military (GWOT generals learned lessons); Chinese military inexperienced (decapitated experienced Vietnam guy). Tech: space, air, land, sea, cyber. Example: 100 guys into most fortified Venezuela compound, took president/wife safely, out with no KIA.
War games planned years ahead (Midnight Hammer in Iran). Refresh tactics, techniques, procedures to minimize life loss, maximize success.
Discombobulator real but can't discuss.
Could not pull off Venezuela mission 5/10 years ago—not just tech maturation, but relaxed rules of engagement (ROE). Past ROE insane (e.g., Afghanistan: match gun size). Objective: unfair fight. Past ROE required legal understanding every moment; now: clear objective, red lines, use judgment (Colin Powell approach: overwhelming force).
Secretary Hicks highlighted ROE punishing in Afghanistan/Iraq.
More drones deployed this week than US military history.
Evolution: Predator (big, 10-15 years ago); Obama balcony strikes; Trump Soleimani drone kill; Russia-Ukraine: 70% casualties from drones, drone-on-drone/robot-on-robot future.
Drone Dominance program: build arsenal. Small drones cheaper than $35k; LUCAS one-way attack drone (500-700 miles, airplane speed, big warhead): $50-80k. Uses: drug drones at border, protect bases, 2026 World Cup, 2028 Olympics, surveillance.
AI: sophisticated drone war = drone swarms controlled by AI. Uses: decoy discrimination, automatic target recognition, identify persons safely. Ukraine/Russia kill box: lose comms (jammed), AI searches box for target. China has drone swarms, heterogeneous autonomy (interoperable drones, e.g., DJI forced).
AI can hallucinate/mistake. Ukraine/Russia hot war less thoughtful.
Not at FSD moment (magnitude better than human). Gradation by risk: low-risk (drone vs laser—miss okay); high-risk (dense population—less risk). Develop procedures/tactics per scenario.
Debate with Anthropic: need AI for Golden Dome (Chinese hypersonic missile: 90 seconds to separate/decoys; hit from space—human can't react/discriminate).
Next 10 years: AI applications where safe/low-risk.
American military fought with one hand tied (past ROE frustration for soldiers).
There was broader frustration about purpose and secondary frustration about inability to do the job. Regarding full autonomy in military action, systems are not close to readiness for real civilian risk scenarios. Focus is on basic autonomy in drones, underwater unmanned vehicles, and collaborative aircraft that fly with jetcraft for more firepower while tethered. Golden Dome exemplifies unopposable defense: a laser autonomously taking down drones threatening a military base with sleeping soldiers.
Debate is scenario-specific, not broad due to Skynet being unrealistic. Told Anthropic: models are getting stolen by Chinese who will unguardrail them for use against us; weakening our models is counterproductive. Benchmark: 10-15 years ago Wikileaks' Collateral Murder showed Apache tragically taking out journalists; current technology likely would have avoided it. Similar to Kuwaiti aircraft hitting American due to identification failure. Self-driving analogy: scary without human but safer, more precise in many scenarios.
Israeli government desires regime change in Iran. Tucker Carlson and MAGA base claim US captured by Israel. Debate in Republican party/MAGA: no wars, Israel driving actions per Rubio's quote that US follows because Israel goes anyway.
President not captured by Israel; decides US best interest, incorporates allies like capable Mossad. Operation Epic Fury: four days in, Iran 90% depleted of munitions, no more missiles from Iran, fleets of drones/planes waiting knowing Iranian positions. Rely on allies when deciding action. Contrast: UK, Spain pontificating, Europe taking weekend off before statement.
Iron Dome first generation, building Iron Beam; early but works. Israel technologically sophisticated, small, invests due to necessity, smart people. Primarily works on rockets. Logical evolution from 80s/90s space-based lasers for ICBMs to avoid nuclear war, taking out every nuclear warhead. Space-based harder but more valuable for space threats; need ground layer for cruise missiles, drones—multi-layered defense.
Directed energy lasers improving: more powerful for bigger weapons farther away, more capable against variety at standoff distance (ideal over enemy land, not over Tel Aviv). Laser interceptors demonstrated in field. Collaborations with Israel; they have own, US has own. Trump said it's US technology. Evolving science: cohering beams for distance, high-powered microwaves to drop drones.
Pro-regime change thoughtfully, isolating dictators like Putin, Kim Jong-un via diplomacy. Free Iran after 50 years of subjugation. Trust Trump; has more info. Independent, pro-Trump judgment. Learnings from Iraq, Iran-Iraq war, 1953 US/British deposing Mossadegh for Shah (later booted). Avoid 20-year forever war; tweet: not every war needs three decades/trillions (to friends in Virginia/Maryland/DC). One-and-done possible.
Venezuela/Iran freedom: 5% world under autocracy flipping to democracy, more for spread than any president in decades/lifetime. Noble, just. Give Trump Nobels, all prizes if frees people.
Oil at $84/barrel Wednesday, standstill. Strait traffic slowing, ships taken out. Unless opens, 3.3 million barrels daily production lost early next week. Insurance cancelled war risk coverage for Gulf vessels effective March 5th. Supertanker traffic dropped 94% first 48 hours. Trump: US provides political risk insurance for Gulf maritime trade, especially energy.
Modern insurance from 17th century Lloyd's of London coffee shop for maritime risks (safest routes vs pirates/weather). Lloyd's has 78 syndicate members underwriting big risks like Hormuz oil tankers. IRGC shut down Strait, risk of mines/ships attacked. Premium spiked 5x from 0.25% to 1.25% of ship value, then markets shut down.
Massive risk to global energy prices, inflation, US economic security. US International Development Finance Corporation (from 2019 USAID output) provides insurance. Shipping companies levered (debt for ships), debtors require insurance. Fills market gap. Could reshore/onshore maritime insurance to US, new industry historically European.
Pentagon notified Anthropic as supply chain risk (first for American company; done to Russian/Chinese like Huawei). Cancelled $200M contract Friday. Disputes: Anthropic concerns on fully autonomous weapons (Dario: tech not reliable, wants assurances); mass surveillance of Americans (uniquely powerful beyond webcams/7-Eleven cameras). Pentagon wanted all lawful use.
Biden AI executive order limited compute, grandfathered winners like Anthropic for control. Anthropic sold to combatant commands (CENTCOM for Iran, INDOPACOM for China), intelligence agencies; forward deployed Palantir-style, sticky in workflows.
Emil took AI portfolio August, reviewed contracts: 20-page restrictions—no kinetic strike planning, satellite movement, war games involving harm. Irrational for Department of War mission. Negotiated direct relationships, rational terms. Gave scenarios (Chinese hypersonic, drone swarm); they added exceptions, but unpredictable for 20 years. Wanted all lawful use; follow internal policies/Congress, not hurt unnecessarily.
Turned PR game: inflammatory issues like robot weapons, mass surveillance (not FBI/Homeland). Anthropic didn't want bulk collection of public info; example: querying Chimath on LinkedIn violates terms. Months of this.
Trigger: post-Maduro raid, Anthropic exec called Palantir (prime contractor) asking if software used (classified); implied violation. Raised alarms: potential shutdown, guardrail refusal, poisoning, hallucination, insider threats risking people.
Culminated Tuesday meeting: Hegseth, Emil, Dario; Friday deadline blown. Model in AWS GovCloud, Palantir serves, Anthropic controls plane/weights—can change/interfere.
Insight: Anthropic's political perspective/terms reflect philosophy, changeable. Government reasonable: can't rely on unreliable provider disallowing reasonable uses. Example: pro-life state healthcare—no abortions; Anthropic pro-choice engineers change model? Private right, but governments/states/cities/companies can't rely solely on one due to moral tirades unrelated to law.
Everything is about subjectivity. This weekend's events tore this open. Failing to be multimodel and agnostic across models takes on enormous business risk after Friday because you can't tolerate providers doing that. It's too critical a technology. This is deplatforming all over again, but times a thousand—not about posting on social media, but using fundamental technology to advantage or disadvantage your business.
Leaders of these companies say they're going to cause 50% white-collar unemployment. This is as powerful as a nuclear bomb—50,000 geniuses in a data center. A small country could coerce the world. It's a general substrate of intelligence applicable to many things, not like Workday HR software with competitors. It will be part of everyday life, with control over moral conscience. Anthropic has its own constitution and soul, not the US Constitution, subject to whims and changes. Scary for Americans, becoming a bigger deal.
Grok: All in for all use cases across classified and unclassified networks, as expected from Elon's truth-seeking. Department of Defense doesn't want ideology messing with operational decisions—nothing fake or tilted.
Google: Surging Google for all lawful use cases on classified networks, moving to classified networks. They need to build infrastructure; it's complicated. In compliance as a partner.
OpenAI: Sam Altman playing both sides. Called him needing multiple solutions if things go sideways; he'd like OpenAI as one. Sam offered to help the country, tried protecting Anthropic by negotiating terms, avoiding calling them a supply chain risk. They compete for the same thousand top researchers like baseball players in Moneyball. Lose 20% and they launch Claude Code before Codex, numbers change dramatically. Sam acted patriotically, helping while Anthropic trashed him and recruited from OpenAI. Wants all providers on same terms for redundancy to see if they diverge or converge—maybe need only two over time, but too early.
Anthropic deeply embedded, so must unentangle. Others not as heavy on enterprise sales, forward-deployed engineers, government business—they must catch up on serving government, not just model capability. Bible way ahead. No view on models uniquely advantaged. Claude Code innovative and ahead, but in 12 months Codex won't be close. Asymptoting happening—confidence intervals shrinking, models becoming similar with enough compute, generating similar results. Even more important to have a complexion of models.
Anthropic posted revenue ramp. Speaker's small software company 8090: opex tripled since November 2025 from inference costs to AWS (ginormous), Cursor, Anthropic—spending millions, more per unit and in aggregate. Costs up 3x every three months, revenues not. Token use addicting. Infatuated with Ralph Wiggum loops—send thing off to figure out, but it never does, huge bills from Cursor wrapping Claude Code. Deprecated Cursor. Tool usage fun, feel like genius, but ROI important. Will become issue as use cases sophisticated, more consumptive.
Not punitive. If their model has policy bias from constitution, culture, people—don't want Lockheed Martin using it to design weapons, or componentry providers. Risk of poisoning enters defense enterprise. Fine for Boeing commercial jets, not fighter jets—don't trust outputs wedded to policy preferences.
Why not handle quietly? Two to tango. Anthropic has sophisticated press operation, painted DoD as mass surveillance over commercial database for recruiting soldiers (DoD runs schools, hospitals, not just wars). Worried about them shutting off system in moment of need or messing with it.
If selling batteries or laptops for lawful use, should be enough unless peaceniks not wanting involvement in weapons—Google's position for years to recruit aligned talent. If don't want stuff used for Department of Defense, don't sell to them—it's in the name.
In room of 20 people, gave Golden Dome scenarios; Anthropic rep said "just call me if you need another exception." But if balloon goes up for decisive action, not rational to call. Holy cow moment on how they think.
Anthropic lost all Republicans. Fleeting with Democrats—progressives hate Silicon Valley and technology, won't let uncontrolled god king. Pissed off every constituent. Progressives want more control, Anthropic pushes back, same situation.
Don't sell bullets if don't want to be in war. What was senior management and board doing? Breathless on revenue curve—added $6 billion ARR in a month, absurd growth. Great product, Open Claude drove it. On board, close eyes, shut up if working. Company not worth $350 billion anymore.
Pick between OpenAI, Anthropic, or Google. Google worth $3 trillion already—doubling adds $3 trillion, market value creator bet. Anthropic multiple bet—$1.5 trillion market cap unless blows up (5x vs. 3x or 2x). Anthropic worth more than $350 billion, undervalued vs. ChatGPT. Anecdotally, everyone on Claude/Copilot (Co-work), fulfilling AI promise. Google will flip virtual version integrated with G Suite in 90 days, sweep market.
Elon expects Grok to exceed coding models in May, for sure June. If all models asymptote by October, build governance layer, indifferent—compete on price. Love indifference: one main and redundant, or two mains. Anthropic out if posture continues, so three. Want two or three; all in San Francisco vibe except Elon's. Google long-term advantage: own cloud, no margin pass-on like Anthropic.
Google Workspace now integrated for agents, 40 agent skills.
Dario: Real reason Pentagon/Trump admin dislike Anthropic—no donation to Trump, while OpenAI/Grok did. List of nine companies (Microsoft, Apple/Tim Apple, Nvidia, Amazon) participated in inauguration, White House CEO dinner, Melania documentary—all except Anthropic. Singled out for not genuflecting, not paying cover charge; pay-for-play accusation.
Dumbest thing heard. In Department of Defense to win wars—if you help without transition waste, thrilled. Not Trump dictating. Spent three months negotiating simple standard—why if just kicking out? Internal psychosis.
Theory: Massive Claude instance internally for strategy—Claudebot saying zig for press, tell DoD to pound sand. Stupidest decision.
Advice to Dario: Genius, best tool—not battle to win, unpatriotic. Tim Cook at Melania premiere—support president, show up for America. Don't have to donate, be patriot. Biden excluded Elon, he ankled.
Better advice: Call Emil back, apologize, press release: support customers' lawful use of models; terms of service in stone, solidity/reliability. Misstep.
Emil response: Always wanted reliable partner for autonomous systems, not wigging out early stages.
Dario called Trump wannabe dictator in 5,000-word essay—apologize to more than me, rewrite position. Kumbaya.
Early days reducing China dependency. Critical minerals action. Office of Strategic Capital: $200 billion lending authority, Treasuries +100 bps loans to domesticate solid rocket motors, batteries, fiberglass—20 critical things outsourced to China. Critical minerals done this term from US/allied sources. Batteries next—outsourced technologically and lithium. Takes years to build plants. Hope next administration continues; free market but outsourced crippling assembly.
No munitions risk now, but plus up needed—Europeans slow on Ukraine contribution. Russia consumed munitions worldwide, plus US conflicts. Fighting with 1980s Cold War weapons largely, not modern. Regenerate: nuclear missiles 50 years old, planes 40. Need next generation.
More defense tech venture.
Defense tech venture capital is 3x more than last year and growing. The department needs companies like Seronic to win big contracts quickly, enabling more money to flow in, attracting more entrepreneurs, and allowing more purchases. Warfare is shifting from big carrier ships costing $20 billion and a decade and a half to build to mass-produced, low-cost, attritable things that new entrants can provide. Those companies need to succeed to create a flywheel with venture capital, entrepreneurs, and capabilities.
The market is moving from old primes to new primes, with a small set of big winners like Anduril and Palantir, plus three or four others, where most value will accrue, along with seconds and subs. Anduril and Palantir want that dominance. At least a second layer of innovative companies is needed to disrupt the first layer. Example: A mom-and-pop wholly owned company makes Rams missiles sold and sent to Ukraine; 30 people produce 1,000 a year because they've designed and manufactured efficiently.
Primes are few partly because they mastered contracting with the government and navigating bureaucracy, which became a competitive advantage. The goal is to take that advantage away.
Part of the solution is requirements reform. Previously, services like Army, Navy, Air Force specified detailed requirements for a fighter jet (stealthy, hold a missile, hold four humans), creating unbuildable things. Contractors didn't care under cost-plus contracts, delivering years late and billions over budget without proper engineering.
New approach: State the common operational problem, e.g., need missiles that go 500 miles or more with a certain blast effect. Companies provide solutions with as few requirements as possible. On contracts, move to commercial-style contracts. Startups excel here with fixed-cost pricing: pay less if late, more if early. This is super disruptive to the existing system.
Example application: Announce Strait of Hormuz is super important, need devices to keep it open, and invite creative entrepreneurial ideas instead of goosing profits.
DARPA is the father of the modern internet and has created many critical technologies. It's still an honored profession with the best talent; great engineers go there instead of just private companies like SpaceX.
Public projects:
- Using biology to synthesize critical minerals, pulling them from the ground without messy, dirty refining. This would transform access to critical minerals, getting them faster and leapfrogging China in tech.
- Deep work in cyber, as cyber attacks are the next huge threat with AI. Example: Creating agents to attack systems, as happened to Anthropic.
Much of DARPA's work is classified.
On aliens: Fought for the alien portfolio but didn't get it. Team wanted it badly, but requires 100% access to everything.
On China: They have had the greatest military buildup in world history in the last 15 years. US was asleep at the wheel, focused on global war on terror, so China advanced without threat perception. US still has sophisticated operational expertise, subs, space layer, and the best stuff in the world, but the gap cannot narrow. Cannot be complacent.
President Trump is allocating money decisively. Recent events in Dubai, Doha, and Tel Aviv—attacks on residential homes—remind how fortunate the US is with layers of protection.
Defense is critical for freedom, even for libertarians favoring small government. It's an amazing, valuable institution.
Thanks to Emil for candor, thoughtfulness, and insight.
Keep All-In Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
