Loading page…
In a new Stratechery interview, Sankar argues that better models alone will not deliver economic returns. Interviewer Ben Thompson challenges who should be responsible for turning them into useful business systems.
Listen to this story
Sankar’s argument shifts the test of enterprise AI from whether it makes an employee faster to whether it improves the workflow where a company captures value. He says model research should continue, but adoption work deserves more attention; otherwise, gains in machine capability may not translate into business returns or wider economic benefits. Stratechery’s Ben Thompson challenged whether labs need to take on that work themselves or whether it is an opportunity for companies building applications around their models.
Sankar says coding is an exception because an individual worker’s output can more directly represent the workflow, helping explain coding tools’ strong product-market fit.
He compares AI diffusion to electricity: economic gains required not just power generation but machines and factory redesign built around it.
Sankar describes labs’ focus on stronger models as rational under incentives that reward research and valuation growth, while adoption returns are deferred.
Making workers faster with AI does not necessarily make a business more productive, Palantir CTO Shyam Sankar argues. In an October 29, 2026 interview with Stratechery’s Ben Thompson, he challenged the industry’s focus on smarter models, saying the harder task is turning their capabilities into measurable gains across an entire business process.
Sankar is not calling for model research to stop. He explicitly says the industry should keep building better models. His criticism is about the balance of attention: too much goes toward increasing the supply of machine intelligence, he says, and too little toward getting that intelligence into productive use throughout the economy.
His central example is a supply-chain analyst. An AI tool could make that person more productive without generating any additional return for the company, he argues. To capture value, the business must make the whole supply-chain workflow measurably more productive. The distinction is between improving one person’s work and improving the process that person helps run.
That challenges what Sankar describes as an assumption embedded in the labs’ product strategies: make every knowledge worker more productive, and the return on investment follows. He says the relevant unit is usually the workflow instead. His argument asks businesses to judge AI at the level where they capture value, rather than treating individual productivity as sufficient proof.
Coding is an important exception in his account. In some roles, the worker and the workflow can effectively be the same unit, allowing productivity gains to translate more directly into value. Sankar says that helps explain strong product-market fit for coding tools, but cautions that those roles represent only a small part of the real economy.
The atomic unit of the enterprise is the workflow
Shyam Sankar, Palantir CTO, speaking to Stratechery
To explain what he calls AI diffusion, Sankar draws an analogy with electricity. Generating power was only one part of its economic impact. Other people had to build machines that used it, and factories could then be redesigned around those machines. He wants a comparable focus on how businesses put AI to work, not just how models improve.
He attributes the imbalance partly to market incentives. In his telling, labs conduct research, produce more powerful models and receive sharply higher valuations. The expectation that widespread adoption will eventually generate economic value is built into that cycle. Sankar calls the behavior rational for the labs; his objection is that the payoff from adoption is being left for later.
Thompson challenges the assignment of responsibility. Why shouldn’t model makers concentrate on better models, he asks, while other companies turn those models into useful applications? He points to Palantir itself as a company pursuing that opportunity. The exchange raises a division-of-labor question: is the gap evidence that labs should change their priorities, or an opening for companies building around them?
Sankar answers by pointing to people asking where the returns on AI spending will come from. He questions whether convenient AI assistants alone can meaningfully change economic growth. His concern also extends beyond business returns: he says the spread of tangible benefits matters to AI’s continued acceptance in society. For him, useful deployment is part of the technology’s justification, not merely a downstream commercial opportunity.
Loading discussion...
Join the conversation
Explain where model makers’ responsibility should end.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.