Shyam Sankar on Palantir, Reindustrialising America, and AI in Warfare

Palantir’s CTO argues that America lost the capacity to make things — and, with it, the capacity to innovate — and that the only route back runs through ‘heretics’: obsessive founder-figures who deliver against the bureaucracy, in the military and the factory alike.

Key ideas

1. Heretics are founders, and they are the only thing that ever worked. Sankar collects the stories of military mavericks — Hyman Rickover (the nuclear navy), Andrew Higgins (90% of the boats that landed at Normandy), John Boyd (the OODA loop), Billy Mitchell (the Air Force). Each fought the institution, paid an extreme personal price, and was vindicated only later. His claim is total: ‘the things that helped us win all the wars, were the things that the heretics actually did. Nothing that went through the machine delivered anything.’

2. Innovation is downstream of production, so offshoring the factory offshored the innovation. The ‘biggest lie’ of globalisation was ‘we will do the innovation and they will do the production’. Sankar’s counter: the attention-is-all-you-need paper grew out of an incremental 3% tweak to Google Translate — remove the input stimulus of making things and you forfeit the breakthroughs downstream. China ‘went from making a battery to making the whole car’; 50% of clinical-trial drugs now originate there. He is a re-industrialisation maximalist and calls it ‘an undeclared state of emergency’.

3. American manufacturing software doesn’t work — the factory floor runs on Excel. For 15 years Palantir’s software was used to make things, and the one truth no customer would admit before COVID was that their multi-billion-dollar ERP was ill-fitted to the job. On the final assembly line, everyone had hit the eject button and hand-rolled their own spreadsheets — the obvious sign of a system that fails at the point of use.

4. Forward-deployed engineering builds software by back-propagation from the operator. The feedback loop ‘can I sell it?’ is weaker than ‘did it deliver the outcome?’, and only the person on the factory floor or in the foxhole can answer the second. Palantir sends a tech lead obsessed with impact, not a sales engineer chasing the sale, and generalises the specific fix into product. It works only when the problems are big enough to capture real value and when you refuse to fit inside the IT buyer’s architecture diagram.

5. Deterrence was lost and is only now being restored. Crimea, the Spratly Islands, Ukraine, the South China Sea — a decade of conflict America failed to deter. The deeper cause is not acquisition reform but people: the 1993 ‘Last Supper’ consolidated 51 defence primes to five and financialised the industrial base, driving the ‘crazy people’ out to tech. AI is ‘David’s slingshot’ — make the American worker 50 times more productive and you change what can be made on-shore.

Content

Heretics as founders

Sankar’s organising figure is the heretic — the person who ‘gets obsessed with delivering something that it makes no sense’ inside an institution that will end their career for it. His exemplar is Hyman Rickover, who built the first nuclear submarine in seven years while Oppenheimer told him it would fail and the Navy assigned him a women’s restroom for an office to humiliate him into quitting. Rickover designed the reactor to a standard ‘a hundred times safer’ than the assumed minimum — ‘I’m designing this thing so my son could be in it’ — and the engineering culture of naval reactors survives today, underwriting one of America’s last asymmetric advantages against China, built in the 1950s. Andrew Higgins built the Normandy landing craft the Navy repeatedly rejected because it was not their idea; it became 92% of the boats used in the war. John Boyd developed the OODA loop and lived to see it vindicated in Gulf War I. Billy Mitchell was court-martialled and died ‘penniless and depressed’ before his heresy became the Air Force.

The heretic and the founder are the same type — ‘a pathological obsession with winning’ that converts to heroism only through ‘an empirical thing they built’. Sankar is candid that he does not share the personal disagreeableness of a Boyd; his own orneriness surfaces structurally, in refusing to build what the IT buyer asks for when he knows it will not work for the operator, even at the cost of suing the Army to compete.

The worldview: production, innovation, and America

The formative material is personal. Sankar’s father fled violence in Nigeria and ‘almost died’ before settling in the United States in his mid-thirties, started businesses that went bankrupt, and modelled a positive-sum graciousness — ‘every spare moment this man had… he would be on the phone trying to help someone’. Growing up near the Space Coast in 1980s Orlando fixed an optimism that technology makes the world better. From this Sankar draws a claim about national exceptionalism grounded in evidence: Europe has created zero companies worth more than €100bn from scratch in the last 50 years, while America created all its trillion-dollar companies from scratch in the same span. The American advantages he names are a belief in greatness (‘no army that lost its morale ever won the war’) and ‘plasticity of thought’ — a culture that changes its mind when the orthodoxy stops working.

Talent: superpowers, kryptonite, and gamma rays

Sankar credits Alex Karp with modelling Palantir as an artist colony — ‘we have more in common with… a Hollywood talent agency than we do with… a typical software company’. Talented people are ‘highly uneven’. Their superpower is the thing that is effortless — ‘Superman could fly… but that wasn’t some sort of arduous thing for him’ — which young high-achievers misattribute, mistaking the merely-difficult for the exceptional. ‘All of your contributions to the world are going to come from your superpower. Everything else is basically a waste of time.’ The counterpart is kryptonite — weaknesses six standard deviations below average, to be avoided rather than fixed; accepting them makes the artist more valuable.

The development method is deliberate over-exposure: ‘How did Bruce Banner become the Incredible Hulk? … a near-fatal dose of gamma rays.’ Palantir gives a person a project they are unqualified for, then builds enough transparency into the environment to intervene before it goes off the rails. The alternative — the structured career ladder — ‘is perfectly designed to make you feel comfortable that you are growing’, but ‘it’s all fake’. The maximal rate of learning is ‘coincident with your maximum ability to tolerate pain’. The bet on a person: ‘If I give this person an inch, can they turn it into a mile?‘

Forward-deployed engineering and the ontology

Sankar’s critique of the ‘conventional software industrial complex’ is that its feedback loop stops at the sale. Forward-deployed engineering closes the loop at the outcome, on the factory floor or in the foxhole, and builds the product ‘through back propagation’. Two engineer archetypes must be held in tension: those who build the right thing (hackers, ‘more like MacGyver’) and those who build it the right way (artists chasing architectural elegance). The seam between them ‘you actually kind of need a dictator-like figure to manage’ — which Sankar names as his own narrow superpower. The model pays off only when customers sit on a power law, where a handful ‘living in the future’ reveal a valuable problem no one else has solved: ‘We’ll be 5 years ahead of everyone else if we lean into this problem.’

Asked what Palantir is — ‘the hardest question in the world for me’ — Sankar describes an enterprise operating system built on the thesis that ‘decisions, not data’ create value. Its ontology is a model not only of the data but of the actions (‘the kinetics’), an API layer that ‘makes the business programmable’. At Airbus, ramping the A350 in 2015, Palantir sat with ‘wrench turners’ on the Toulouse assembly line to distinguish a one-off non-conformity from a recurring design defect, then compounded quality into production planning and in-service uptime — the natural evolution along a customer’s ‘decision chain’.

Deterrence, China, and the industrial base

America has ‘the best military in the world’ but has failed to deter a decade of conflict. Sankar traces the cause past acquisition reform to people and the industrial base. At the dawn of World War II the country direct-commissioned Bill Knudsen — GM’s number two — as a three-star general to run war mobilisation, and Lend-Lease let it reach full-rate production before Pearl Harbor. Then only 6% of major weapons spending went to defence specialists; 94% went to ‘dual-purpose’ companies — Chrysler built missiles and minivans, General Mills built torpedoes, Kodak supplied the first spy satellite — so that civilian R&D subsidised national security. Today that 6% is 86%. The 1993 ‘Last Supper’ dinner told 15 of 51 primes they would not be saved; consolidation to five primes ‘financialised’ the base and drove ‘the crazy people’ to tech.

On China, Sankar’s diagnosis is that ‘it’s not enough for China to be prosperous, America must fail’. Their conception of war is deception and ‘system destruction warfare’, waged below the threshold of open conflict. Their asymmetric advantage is long-term planning — systematically mapping US dependencies since Gulf War I; America’s is that ‘we’re crazy’ and unpredictable, pivoting the whole economy onto AI on a dime. The remedy runs through incentives: cost-plus contracting is ‘anti-heresy’ because it caps upside and removes skin in the game; companies like Anduril, absorbing product risk on their own specs, point the other way. On-shoring can be catalysed cheaply — extend patent length for drug-makers who build worldwide branded supply in the US, and the generic-manufacturing capability follows.

AI, self-disruption, and where value accrues

Palantir must disrupt itself: ‘the only way you can do that is by disrupting yourself.’ AI ‘plays to the beginner’s mind’ — the earliest and best adopters were the youngest hires, whose ‘overwhelming productivity’ humiliated or inspired the rest, so Sankar spent no energy on the cynics and instead built momentum behind the insurgents. Delivery time for customer value has compressed from eight weeks two years ago to about a week. On the AI stack, he expects value to accrue ‘at the chips layer and at… the ontology layer’: model companies must run up the stack because the model commoditises, and pure-play AI applications run down, reinventing the ontology from first principles as they scale. Closing on Boyd, Sankar frames the choice every builder faces — ‘you can either be somebody or you can do something… you can’t really have both’ — and places Palantir, and himself, on the side of doing.

See also

• Nabeel Qureshi on Palantir and Forward Deployed Engineering — the same company and the same forward-deployed model, from an engineer who lived inside it.
• Dylan Patel on the Token Economy, AI Supply-Demand, and the Permanent Underclass — the AI supply chain and the chips layer where Sankar expects value to accrue.
• Gavin Baker on AI Infrastructure, Power Constraints, and Semiconductor Investing — the sibling Invest Like the Best conversation on the compute stack underpinning AI’s economic weight.
• 7 Powers — the competitive-advantage lens on Palantir’s ‘monopoly on the capability’: solving a power-law problem years ahead of rivals is a structural barrier, not a feature.
• Shyam Sankar — the guest’s speaker page.