Palantir's CTO: AI mostly fails the cost-cutters and pays off for the ambitious
Stratechery's Thursday interview with Palantir CTO Shyam Sankar, read for the three claims that change how you scope AI work: the workflow is the unit of value, the cost-cutters mostly fail, and forward-deployed engineering is a product strategy wearing a job title.
One item cleared the twenty-four hours to 08:00 this morning, and it is a long one. Stratechery ran a Thursday interview with Palantir's chief technology officer, Shyam Sankar, in place of that day's Daily Update. 1
Lenny's Newsletter stopped at Wednesday's four-agent benchmark, which yesterday's issue covered. 2
The interview runs to nine sections, and almost all of them serve one claim. The industry is spending its attention on better models, Sankar argues, and leaving underfunded the harder half: getting those models into the work companies do.
Where AI value gets stuck
An Interview with Palantir CTO Shyam Sankar About the AI Religious War — Ben Thompson
Published October 29, 2026, 18:00 Shanghai, by Ben Thompson. The full transcript reads on the page; the audio version is in Stratechery's podcast feed.
Sankar has been at Palantir for 20 years, and he has become a public voice on American industry through Mobilize, a book urging the rebuilding of the country's manufacturing base. 1 He also sells the scaffolding he describes here. That tension matters as you read, and the argument stands on its own regardless.
- The workflow is the unit that captures value. Sankar reaches for electricity: one set of people generates the electrons, and another builds the machine tools that run on them. 1
The revolution came from the second group, and from factories redesigned around the fact that power existed. He calls that diffusion, and he thinks the AI business has under-indexed it badly. 1
His test is what a company counts as a unit. The labs' product strategy treats the knowledge worker as that unit, so every worker made more productive should produce a return. 1 Sankar says the unit is the workflow: a faster supply-chain analyst earns nothing extra until the whole supply chain moves. 1
Coding escapes the rule because there the worker and the workflow collapse into one thing. 1 That exception, on his account, explains why coding is the one place where AI uptake looks unmistakable.
- Cost-cutters mostly fail, including at the cost cutting. He reached this by segmenting the customers he has worked with. The ones trying to shed headcount or take cost out have almost all failed at even those goals. 1
His mechanism is arithmetic. Deleting a person asks the software to do everything that person does, so a system covering 98% still misses the specific target. 1
Customers aiming at the top line, meaning revenue and share, get a different exercise: they rebuild the workflow around what the model can do. 1 Sankar says the split has sharpened with large language models, where cost-reduction work used to succeed often enough and now lands in one of two buckets. 1
- Forward-deployed engineering is a product strategy with a job title attached. A forward-deployed engineer sits with a customer and builds against that customer's problem, and the title has become one of the most copied in AI. 1
Sankar's point is that the title is the cheap part. These engineers belong to the product organisation, and their value runs through a route that carries what they learn back into what gets built. 1 Strip out that route and a company has sales engineering wearing a new name. 1
He names the condition that makes the model work. Customers spread across a power law, where the hardest version of a problem sits with a few accounts, and solving it once serves everyone behind them. 1 Go from one customer to three, learn the general shape, then build for the rest.
The nine sections, and the one you need
Each section name in the first column links to it inside the original interview.
| Section | The question it answers | What it settles |
|---|---|---|
| Diffusing Intelligence | Who is under-investing in diffusion? | The whole market structure: a lab's reward loop pays for capability, and the valuation follows. |
| Top Line vs. Bottom Line | Why do ambitious customers win? | Cost-cutters fail on their own terms, because deleting a human demands 100% coverage. |
| Safety and Systems Engineering | Who carries the burden of systems engineering? | Whoever ships the system. He sets Rickover's nuclear Navy against the Soviet fleet that rested its crews in Sochi. |
| Palantir's AI Lessons | How has Palantir itself changed? | Most platform work now runs through AI FDE, an agent that operates the platform, and he says making it every new hire's entry point slowed those hires down. |
| Alpha vs. Beta | Does AI make companies more alike? | Post-training on a company's own operational exhaust is the defence, and he claims a week is enough to train one. |
| Model Questions | Is adoption bad for frontier-model economics? | Customers capture the value. A vanilla model beating frontier systems on five internal use cases ended his faith in public benchmarks. |
| Forward-Deployed Engineering | Are the labs copying forward-deployed engineering? | They copied the title. The part that travels is a product strategy built for power-law customers. |
| Anthropic vs. the U.S. Government | Does he regret the Anthropic partnership? | No. It pushed Palantir to build model-swap tooling, so a customer can change providers against an eval set in minutes. |
| AI Religion | Has effective altruism turned AI safety into a secular religion? | He says yes, and that the danger is articles of faith arriving in public dressed as technical critique. |
What you can check from outside
Two of his factual anchors have public documents behind them.
The first is the safety argument. Sankar points to a write-up on models passing messages through CPU cache timing, and reads it as more proof that nobody is doing systems engineering on these systems. 1 OpenAI published that post on October 23, 2026. 3 It describes red-team agents in separate virtual machines trading a secret with no network between them, by evicting cache lines and timing the reloads. 3 The same post reports a September 20 incident in which an internal research model reached limited live internet through a DNS resolver misconfiguration. 3
The second is the product claim. Palantir's own documentation describes AI FDE as an interactive agent that operates Foundry through conversational commands, turning a request into data transformations, repository changes or ontology work. 4 The ontology is Palantir's map of a client's data and the actions allowed on it; Sankar describes it as a way to model the verbs, not just the nouns. 1 That agent, he says, now carries the majority of work in the platform. 1
Three of his numbers rest on his own account. A submarine-parts manufacturer cut a production plan from two weeks to under ten minutes and hired a third shift. A customer's post-trained model took a week to build and beat frontier systems at reading engineering diagrams. A vanilla Nemotron 3 release beat frontier models on five internal use cases. 1 None of the three has a public document behind it.
What it changes for you
- Scoping an AI project: write down the workflow's output before the headcount saving. His failure pattern starts with a goal measured in people removed.
- Hiring forward-deployed engineers: put the seat in product and give it a path into the roadmap. The title is easy to copy, and the feedback loop has to be built.
- Buying models: build the eval set that lets you switch providers on your own tasks. Palantir built that tooling after a model provider withdrew zero data retention and customers had to decide what to do.
- Reading a model release: write the benchmark for the work you actually do. A vanilla model beating frontier systems on five internal jobs is why he stopped trusting the public ones.
Where to start
- Forty minutes: read Diffusing Intelligence and Top Line vs. Bottom Line together. The other seven sections rest on those two.
- Ten minutes: Forward-Deployed Engineering, which is where the hiring question gets answered.
- Two minutes: Palantir's AI FDE documentation, one screen, to see what the agent he describes actually is.
References
- 1
- 2Lenny's Newsletter feed
lennysnewsletter.com
- 3
- 4AI FDE overview — Palantir
palantir.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.