Summary

Shyam Sankar’s defense thesis is that America lost deterrence by turning whole-country mobilization into a specialist industry. In 1989, only 6% of major-weapons spending went to defense-only companies; now it is 86%. Reversing that isolation means reconnecting commercial R&D and manufacturing to national security because “when a country goes to war, it’s the whole country.”
The defense opportunity is not merely adding competitors but restoring founders, heretics, and the leaders willing to protect them. The post-Cold War “Last Supper” reduced 51 prime contractors to five, but Sankar argues the deeper damage was that “consolidation bred conformity,” shifting management toward dividends, buybacks, and cash flow. The counter-model is the Higgins boat: rejected by the Navy, yet ultimately 92% of all boats in World War II.
AI threatens beta SaaS far more than software that expresses a customer’s competitive advantage. Sankar’s rubric separates software that makes every company more alike from alpha-oriented platforms that help each operate differently; AI and vibe coding intensify the latter. Day-two maintenance remains “much harder” and partly unsolved, but he does not think that will protect memetically purchased, standardized software.
His AI-stack thesis places durable value at the chip and AI-infrastructure—or “ontology”—layers as models commoditize. Model companies are moving upward into software “harnesses,” while narrow applications are building downward into infrastructure to support more customers and use cases. That convergence leaves standalone models under pressure while chips and infrastructure may remain defensible, in his theory.
The preferred macro outcome is not labor replacement but AI-enabled reindustrialization that repairs the break between wage growth and GDP growth and reconnects production with innovation. Sankar points to Hadrian making people “50 to 100 times more productive” and calls AI “David’s slingshot” against China’s manufacturing scale. His warning is that separating invention from production was a strategic error: “If you don’t make the thing, you can’t innovate on how you make the thing and what the thing is.”
AI’s immediate organizational advantage belongs to domain experts who can now build instead of petitioning bureaucracies. An intel warrant officer can spend two weeks producing a working application rather than making a PowerPoint and seeking permission; sales teams similarly need an “Iron Man suit,” not replacement for its own sake. The objective, Sankar says, should be to “dominate my industry,” not satisfy an article of faith that people must be replaced.
Sankar ultimately sees America’s largest strategic danger as “suicide, not homicide”—a collapse of national will, agency, and functioning institutions. His answer spans military reserve talent, institutional competence, maximalist reindustrialization, and entertainment that makes heroism attractive again. Hard power and optimistic storytelling share one purpose: mobilizing the next generation early enough to prevent a larger war.

Deep dive

1. Deterrence has a shot clock, and mobilization takes years

Sankar described becoming a public advocate as “equal parts an act of desperation and an act of optimism”: desperation at the Pentagon’s long “frog boil,” but optimism because founders were again willing to build in the national interest.
His evidence of eroding deterrence accumulated across Russia’s 2014 annexation of Crimea, the 2015 militarization of the Spratly Islands, the JCPOA’s failure to stop an Iranian bomb, and October 7. Any event could be dismissed alone; together they said, “We need to act.”
Yet he has seen more departmental change in the past year than in the prior 19. His explanation is leadership—including the election, which he said he did not mean politically—creating a “coalition of the willing, coalition of the capable” inside and outside government.
The World War II lesson is not that industry flipped a switch after Pearl Harbor. FDR used Lend-Lease to mobilize earlier, giving factories roughly 18 months to retool and creating capability and deterrence for the Brits and Soviets so America reached Pearl Harbor at full-rate production: “We will not have the luxury” of waiting for an attack.

2. Defense became an island optimized for finance, not invention

In 1989, only 6% of major-weapons spending went to defense specialists; today, Sankar said, it is 86%. Chrysler once built both Minuteman missiles and minivans, meaning ordinary purchases helped subsidize national security while defense could access commercial price-performance curves.
The “Last Supper” compressed 51 primes into five, but Sankar rejected lost industrial competition as the main diagnosis. The government was a monopsony, and much innovation arose from competition among services; the real result was financialization—“dividends, buybacks, cash flow”—and the expulsion of heretics.
His analogy is the Galápagos giant tortoise: exquisite under the monopsony’s special conditions, but unable to survive among mainland wolves. When Palantir began, outsiders had no Defense Department front door; only In-Q-Tel offered one through intelligence. The opening that began around 2015 created a decade-long bench of founders ready for 2024’s “18 theses.”

3. Heretics produce breakthroughs only when powerful leaders protect them

The old industrial base was not an abstraction but Glenn Martin, Jack Northrop, and Leroy Grumman—founders building beyond quarterly performance. Andrew Higgins embodied the pattern: the Navy resisted and tried to take his landing-boat design, yet Higgins boats became 92% of all boats in World War II.
Sankar’s contemporary exemplar is Colonel Drew Cukor, whose experience with a failed attempt to evacuate Yazidi refugees exposed the cost of poor intelligence interpretation: a Marine’s mistaken RPG sighting would have made it unsafe to land and exfiltrate, after which thousands were tortured, enslaved, or raped. Cukor later built Project Maven from a Pentagon basement with almost no resources.
Opponents even alleged Cukor was housing Iranians in his basement. Investigators found a devout Mormon with four daughters living in a 1,400-square-foot house that had no basement—an absurdity illustrating how bureaucracies attack people who put mission above career.
John Boyd’s test was “to be or to do”: become somebody or accomplish something, but not both. His Air Force hated him while the Marines absorbed his OODA loop and high-low mix; equally important was the unnamed three-star who protected him. Bernard Schriever likewise fired the difficult Edward Hall, then rehired him because Minuteman required him.

4. Industry veterans are a dormant military capability

Israel showed Sankar what reserve expertise can unlock. Roughly 360,000 reservists mobilized on October 8, many returning with 20 years of commercial experience; they modernized the IDF more in four months than Sankar had observed during the previous ten years.
America is “drowning with that talent,” yet makes voluntary civil-military fusion difficult while China makes it compulsory. World War II direct-commissioned 100,000 specialists resembling today’s Detachment 201. Sankar’s own contrast: “I’m not sure I really had a lot to give the Army at 24, but I think at 44” he does.
He joined alongside Bob McGrew, Meta CTO “Boz,” and Andrew Wheel, a former chief product officer at OpenAI who is now head of science, advising Army leaders while working directly on projects. His remit includes long-term force structure and treating software as a “malleable weapon system” that commanders wield through operational data teams.
The surprise was the quality of self-taught “green suiters.” A domain expert who once could only make a PowerPoint can now build an application in two weeks and force an empirical discussion. The remaining institutional trap is rushing from zero to N before proving one—something startups understand better than Fortune 100 companies.

5. AI separates alpha software from beta SaaS

Katherine posed both SaaS-apocalypse views: AI may erase code, data, and UI moats, but nobody can casually vibe-code products with deep integrations and distribution. Sankar’s answer was, “Both things are true,” but the decisive split is alpha versus beta.
Beta software makes customers resemble everyone else and is built around the feedback loop “Can I sell it?” Alpha software expresses how a company is different. AI can make bespoke software easier, raising the premium on toolkits that embody strategy while pressuring standardized products bought because everyone else uses them.
Sankar said COVID should have been a “Sputnik moment” for the software industry: no one talked about ERP implementations that saved their supply chains because they fell over “like paper tigers in two weeks.” CEOs instead talked about Zoom and Teams enabling remote work. Palantir benefited because customers could adapt at disruption speed. Day two remains harder and partly unsolved, but Sankar does not think operational difficulty rescues low-value beta software.

6. Models commoditize while chips and operational infrastructure hold value

Asked whether applications or infrastructure will capture AI economics, Sankar described model companies moving upward into software around their models—sometimes diminutively called a “harness”—while narrow vertical applications earn their way downward into infrastructure as their use cases expand.
Palantir’s theory is that value accrues at two potentially defensible layers: chips and AI infrastructure, which it calls ontology. Models themselves face continuing commoditization and margin pressure.
Sankar resisted forecasts phrased as “AI is going to do X.” Humans will use AI to produce an outcome, creating both agency and obligation. The Economist’s possibilities—from vertical AGI to economic collapse to 2% growth—are not fate: “It’s up to us to pick which one we want.”
Most technological revolutions are tool revolutions, he argued, invoking the telescope, microscope, power loom, and personal computer. Galileo did not invent the telescope; he wielded it. Model inventors may possess their own “jagged intelligence,” so technical brilliance does not guarantee correct predictions about social consequences.

7. Reindustrialization reunites production with innovation

Sankar sees a chance to repair the 1970s break between wage growth and GDP growth by giving workers AI “superpowers.” Hadrian’s technology can make people “50 to 100 times more productive,” offering an asymmetric “David’s slingshot” rather than an attempt to replicate Chinese production methods.
Globalization’s “great lie” was that innovation could remain here while production moved elsewhere. SpaceX co-locates R&D engineers with production to shorten feedback cycles; in the negative case, Sankar said, we once viewed WuXi as a cheap set of pipetting arms for contract pharmaceutical research, while 50% of all clinical trials are now being done in China.

8. The winning AI organization augments people and restores engineering authority

Sankar contrasted Europe—zero companies created from scratch in 50 years worth over €100 billion—with America’s homegrown trillion-dollar companies. His explanation was founders, followed by a warning that America’s mega-caps had begun “Europeanizing” through professionalized conformity.
Intel chose its CFO over Pat Gelsinger as CTO, a decision that appeared successful for ten years before “it fell off a cliff”; Sankar called that financial engineering rather than engineering. He thought Boeing’s last engineer-CEO was in 2004 and recalled Andy Grove telling salespeople: engineers create the value; sales moves it around.
When a host suggested sales may be least automatable, Sankar rejected replacement as the objective. A “mixed mammal-AI team” can give salespeople an Iron Man suit and spread top performers’ methods. CEOs ask him how to destroy competitors and grow, not how to fire everyone.
His sharper critique is that some labs have filled Pascal’s “God-shaped hole” with AGI, confusing potentially true articles of faith with empirical facts. China may have an advantage in its pragmatic focus on productive forces, though Sankar would still bet on America “100 times out of 100.”

9. National will, institutional competence, and culture form one strategy

Despite being a China hawk, Sankar said America’s greatest danger is “suicide, not homicide.” He distinguished free trade in soybeans from smuggling agricultural fungus, but argued the decisive contest is whether Americans retain agency, common purpose, and institutions that work—from government services to planes whose doors stay attached.
Physical AI therefore requires maximalist reindustrialization, not a little friend-shoring. America invented mass production and nuclear power; the belief that its people cannot do so again “beggars belief.” He cited “Maduro” and “Midnight Hammer” as reminders of U.S. capability and will, while arguing Russian and Chinese systems failed to protect their partners.
Sankar said his assimilation began through The Hunt for Red October, Red Dawn, and the Rambo films: as a five-year-old, he “knew what it felt like to be an American before I knew civics.” He said he had heard from the maker of 300 that Navy SEAL recruitment went through the roof afterward, an example of entertainment inspiring people without becoming “Pravda.”
His proposed Rickover film concentrates the thesis in one founder. Rickover endured a women’s-restroom office and opposition from both the Navy and Oppenheimer, yet he said he built the first one, Nautilus, in roughly five or six years. Rickover was a four-star admiral for 30 years, demanded reactors “safe enough for my son,” and set a standard 100 times safer than required. Soft power and hard power meet in the same mission: inspire the next generation and prevent World War III.

Terence Tao – How the world’s top mathematician uses AI Skill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI