Palantir's Shyam Sankar on enterprise AI autonomy, defense reformation, and why the forward deployed engineer can't be cargo-culted

Summary

Shyam Sankar, Palantir's CTO and its first forward deployed engineer, argues that enterprise autonomy is the defining AI conversation right now — not tariffs, not Nvidia export controls. The model layer is largely solved. The real work is implementation.

Sankar's position is that we are well past the threshold where models are powerful enough to automate consequential things. Incremental model improvements still matter — they change how many agents you need to decompose a problem — but the value creation has shifted to tooling. His analogy: electricity's value didn't accrete to turbine makers, it went to the companies building machines that ran on electricity. Palantir is betting it is the latter.

The commercial results he cites are specific. Sepsis deaths at Tampa General have fallen by half after automating monitoring. AIG's insurance underwriting, which previously reached only 10% of submissions over three weeks, now processes 100% in under an hour. Sankar frames these not as cost-savings stories but as competitive-moat stories — winner-take-most dynamics that mirror the OODA loop logic in defense, where speed of decision is everything.

Model agnosticism

Palantir's approach to foundation models has been "K LLM" from the start — run multiple models rather than bet on one. Sankar points to two practical reasons. First, open models have converged with and in some cases surpassed closed models on capability benchmarks while inference costs have dropped sharply, confirming commoditization. Second, models get end-of-lifed: the original GPT-4 is gone. Any enterprise that built assuming a specific model would persist indefinitely is already exposed. The toolchain Palantir has built is designed to evaluate models in parallel and migrate safely as the landscape shifts.

Robotics and physical autonomy

Sankar pushes back on the idea that physical autonomy is a qualitative leap. Palantir already has more than 300,000 blue-collar workers operating inside its software daily, from Chrysler factory floors to every Airbus airframe to HD Hyundai shipbuilding. At Rio Tinto, the railroad cars hauling iron ore from mine to port are fully autonomous, as are the three-story dump trucks moving ore on site. His read is that it is a difference of degree, not kind — the system just keeps getting smarter.

On digital twins, he sees the real value in CI-check-style scenario modeling: before making a change to a complex system, you simulate the downstream consequences. His example is the procurement manager who buys raw material at 30% off list price while the production manager absorbs a 40% yield hit — a value chain problem that only becomes visible when you model it end-to-end.

Defense reformation

The executive order on defense procurement is, in Sankar's view, the single most important step toward avoiding a peer-competitor crisis. His diagnosis: after the Cold War, the DoD became a monopsony — one buyer dictating what gets built, at what price, under what terms. He describes it bluntly as the last institution still running five-year plans in a world where China and Russia abandoned that model decades ago. Deterrence, he argues, effectively ended around 2014 with the militarization of the Spratly Islands and the annexation of Crimea.

The EO enforces a 1994 law — the Federal Acquisition Streamlining Act — that already mandates buying commercial items when they exist, modifying requirements to fit commercial products when possible, and only then pursuing custom development. Sankar calls it the most violated law in America. His read is that enforcing it is the only credible path to fielding meaningful deterrence before 2027, because there is almost nothing developmental you can stand up in that window.

The historical throughline he draws is consistent: the Higgins boat, rejected by the Navy until a young marine forced a competition; the Predator drone, developed entirely on private capital by General Atomics and despised by the Air Force until 9/11. At Intel, Bob Noyce capped government revenue at 4% of R&D funding to preserve engineering control over the roadmap — even in 1969 when 96% of Intel's actual revenue came from Apollo and DoD contracts. The defense primes consolidated from 51 to 5 after the 1993 Pentagon dinner known as the Last Supper, cutting per-dollar defense spend by two-thirds overnight. Sankar's argument is that the real damage wasn't reduced competition — it was that consolidation bred conformity and expelled the founder-type engineers who actually made things work.

His reason for optimism now: more than $100 billion has been deployed in the national interest by founders like Palmer Luckey at Anduril, the Sang brothers at Shield AI, and Dino Mauro and Chris Brose at Serco. Founders are back in defense, and Sankar thinks that changes the trajectory.

Cyber warfare

Sankar agrees cyber conflict is dangerously underweighted in public consciousness — partly because it doesn't photograph well. His concern is that normalization breeds nihilism, a fatalism that nothing can be done, which is precisely the wrong precondition for raising standards. He expects that changes when physical systems get compromised visibly at scale — 100,000 humanoids stopping simultaneously in a single city will photograph very well.

AIP and the platform

Palantir's ontology — what Sankar calls a declarative backend — lets companies describe the shape of their data and the logic of their enterprise without having to imperatively wire it all together. For third-party startups building on AIP, the practical advantage is access to 20 years of integrated defense data through an SDK, letting them compete on product quality rather than on navigating procurement bureaucracy. During Hurricane Helene, Army soldiers in the 101st Airborne built their own common operating picture on the platform without Palantir involvement — what Sankar calls the code version of the OODA loop.

Culture and career

Sankar institutionalizes rebellion explicitly. Every new employee's first AMA with him ends with the entire room telling him to f*** off in unison. The intent is to signal that Palantir is an artist colony, not a factory — career paths are undefined, but access to motivating problems and exceptional colleagues is guaranteed.

His career advice tracks the same logic. Growth is not progressive overload; it is a near-fatal dose of gamma rays with a 50% chance of killing you. Linear career progression with a clearly mapped path is, in his words, lead shielding preventing the real exposure. The question to ask is where you will work with the most compelling people and have the greatest surface area on hard problems — whether that is a seed-stage company or a scaled one depends entirely on specifics.

On forward deployed engineering, Sankar is clear that cargo-culting the model kills it. Former Palantir employees describe what an "echo" is and get told it sounds like customer success. It isn't. The methodology works through total ownership of implementation — that is the source of the feedback loop, the quality, and the improvement. Strip that out and you have a veneer.