Public transcript

Commerce Sec. Lutnick and Palantir CEO Alex Karp participate in a fireside chat — 9/2/2026

By: CNBC Television
Length: 25 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

In this fireside chat, Palantir CEO Alex Karp, joined by Commerce Secretary Lutnick, discusses AI sovereignty and how countries should think about the AI stack — compute, data, models, and the application layer. Karp argues that countries need technical experts, not just money, to regulate and deploy AI securely, and warns that deploying foreign frontier models without an application layer can exfiltrate a nation's secrets. He closes on a conditional note of optimism, urging countries to take AI seriously and repatriate their technical talent.

Key points

Karp says every country except China must decide which parts of the AI stack are powered by American versus local companies.
Palantir positions itself as the application and integration layer, with its ontology in the middle, letting clients train models securely without data exfiltration.
He warns that companies deploying foreign large language models without proper architecture are unintentionally exporting their country's alpha, calling it anti-sovereignty.
Karp urges countries to repatriate their technical experts living in America to help set AI policy and regulation.
He stresses that money matters less than knowledge, and the key question is who receives the spending.
Countries need access to compute and their own data, plus electricity, with France's nuclear commitment cited as a model.
Open models can outperform frontier models on the specific vectors a client cares about.
Karp is optimistic if countries take AI seriously, use classified battlefield data, and find patriots to build around.

Questions it answers

00:00Sovereign AI stacks

How should countries think about who powers their AI technology stack?

Karp says every country except China will have to decide which parts of the stack are powered by American companies and which by indigenous local companies. That expertise is very rare, and countries ultimately have to build sovereign ownership themselves.

Every country must decide which parts of the AI stack are powered by American versus local companies.
China's stack will mostly be powered by Chinese companies.
Deep domain expertise is rare and crucial for sovereign ownership.

01:34Palantir and frontier models

How does Palantir fit into the AI frontier model ecosystem?

Karp describes the ecosystem as compute, frontier and non-frontier models, and an application layer. Palantir provides the application and integration layer with its ontology in the middle, letting clients train models on their precise questions in a closed environment so there is no exfiltration.

Frontier models give estimations that degrade over time and need verification.
An application layer is needed to decompose questions and expose assumptions.
Models must be trained in closed environments to prevent exfiltration of proprietary data.

02:20Supply chain example

What practical problem illustrates why an application layer is needed?

Karp uses the example of a French nuclear company needing to know every supply chain dependency that is not French or European, and how it would rebuild the supply chain at what cost. Frontier models alone give estimations that degrade over time, so you need an application layer to decompose the question and verify every assumption.

A part made by one company differs from another even with the same spec.
Countries need to know every non-domestic dependency in their supply chains.
Frontier model answers are estimations that degrade and must be decomposed and verified.

06:32Critique of regulation

What is Karp's criticism of AI regulation outside America?

He says regulation is being written by people who don't understand the technical material, resulting in companies that deploy large language models and unintentionally exfiltrate a country's alpha to the model provider. He calls this anti-sovereignty and says whoever is advising it should be fired.

Deploying foreign LLMs without proper architecture exports a country's alpha.
Karp calls this 'anti-sovereignty' and says those responsible should be fired.
Regulators must ask whether their technical experts are on someone's payroll.

10:12Model choice and security

Why might a country use weaker or open models for sensitive work?

Karp notes an open model can be better than a frontier model on the specific vectors a client cares about, and that countries like France with internal security operations should not have to negotiate model parameters with an American frontier lab, since exporting their methods would destroy their advantage.

Open models can beat frontier models on specific vectors clients care about.
Europe purges data to protect civil liberties; America keeps it for oversight.
Exporting a security method to a foreign lab destroys its value.

13:01Bring experts home

What should countries do with their technical talent living abroad?

Karp says every country has a titan of industry in America, and they must get those experts to come home and help set policy, regulation, and technology choices. He adds that money doesn't matter as much as who gets it, and countries need knowledge more than spending.

Technical experts are sitting in America, not in their home countries.
Knowledge matters more than money; who receives the money is crucial.
Karp is helping some countries build a plan B so they aren't fully dependent on Palantir.

15:50Compute and electricity

Is access to data centers and compute non-negotiable for countries?

Karp agrees countries need access to compute and their own data, describing the stack as compute, data, data management, model control, model training, and the ability to compete. He also stresses the need for electricity, praising France's move to stick with nuclear, and mentions Palantir is working on unpublished sovereign compute architectures with partners like Nvidia.

Palantir does not build data centers, despite rumors.
The full stack includes compute, data, model control, and training ability.
France's commitment to nuclear power is highlighted as an electricity solution.

18:06Politics and optimism

Is this an optimistic or worrisome time for AI?

Karp is optimistic under the condition that leaders go home and take AI seriously, get access to classified data about what has worked on the battlefield, and find actual patriots to build around. He also acknowledges political pressures from tech-driven inequality, noting the American advantage is producing AI while other countries may have advantages in adopting it.

Wealth inequality makes AI adoption politically harder in America.
The Arab Middle East has an advantage in AI adoption due to social order.
Karp is optimistic if countries treat AI as 'game time' and use classified battlefield data.

Notes

Sovereignty and the AI Stack

Karp argues every country except China must decide which parts of the AI stack are powered by American companies versus indigenous ones.
The stack consists of compute, data, data management, model control, model training, and the ability to compete against adversaries.
Deep domain expertise is rare; countries ultimately need sovereign ownership and should bring home their technical talent currently living in America.

Palantir's Role: The Application Layer

Frontier models give estimations that degrade over time; an application layer is needed to decompose questions and expose every assumption.
Palantir provides the application and integration layer, with its ontology in the middle, allowing clients to train models in closed environments with no exfiltration.
Example: a French nuclear company must know every non-French supply chain dependency and how to rebuild it — a problem every manufacturer has.

Critique of Regulation

Karp says regulation outside America is written by non-experts, resulting in companies that deploy large language models and unintentionally exfiltrate a country's alpha to the model provider.
He calls this "anti-sovereignty" and says the people advising it should be fired.
A key test for experts: are they on the payroll of a company whose product they are wrapping?

Models and Security

Open models can beat frontier models on the specific vectors a client cares about.
Countries like France should not have to negotiate model parameters with an American frontier lab for internal security work.
He contrasts European data-purging civil liberties norms with American data-retention practices.

Compute, Electricity, and Politics

Countries need access to compute and their own data; Karp praises France's commitment to nuclear power for electricity.
He notes political pressure from tech-driven inequality and says the Arab Middle East has an adoption advantage due to its social order.
America's advantage is producing AI; other countries may have advantages in adopting it.

Optimism

Karp is optimistic if leaders treat this as "game time," get access to classified battlefield data on what works, and find patriots to build around.
Money matters less than knowledge and who receives the spending.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.