An Interview with Palantir CTO Shyam Sankar About the AI Religious War

Listen to this post:

Good morning,

Today’s Stratechery Interview is with Palantir CTO Shyam Sankar. Sankar has worked at Palantir for 20 years, and has taken on an increasingly visible role as public intellectual, including authoring the book Mobilize, an exhortation “To Reboot the American Industrial Base and Stop World War III”. I previously interviewed Sankar in June 2023. That interview is a great place to start in terms of understanding Palantir, which we mostly discuss in this interview in the context of AI.

To that end, in this interview we discuss why Sankar thinks the industry is insufficiently focused on AI diffusion, why companies focused on the top-line have more success adopting AI, and why he is frustrated at the labs refusal to take a systems engineering approach to AI safety. Then we discuss what Palantir has learned from AI, how companies can preserve alpha instead of the median, and whether or not we need more models. Sankar also explains exactly what a forward-deployed engineer is, why tech is in a religious war, and his retrospective on Anthropic’s battle with the U.S. government.

This Interview is presented by Stripe, the economic infrastructure of AI, and for Stratechery itself.

As a reminder, all Stratechery content, including interviews, is available as a podcast; click the link at the top of this email to add Stratechery to your podcast player.

On to the Interview:

An Interview with Palantir CTO Shyam Sankar About the AI Religious War

This interview is lightly edited for clarity.

Topics:

Diffusing Intelligence | Top Line vs. Bottom Line | Safety and Systems Engineering | Palantir’s AI Lessons | Alpha vs. Beta | Model Questions | Forward-Deployed Engineering | Anthropic vs. the U.S. Government | AI Religion

Background

Shyam Sankar, welcome back to Stratechery.

SS: Ben, thanks for having me back.

Well, the last time we talked was in conjunction with Ted Mabrey, the head of Palantir Commercial. As is usually the case with dual interviews, we got right down to business — and I do have to say, I do a lot of interviews where I usually kind of know where I want to go, but that one was super helpful to me in understanding Palantir and thinking about AI generally. But before we get into that, we never got to do my traditional background questions — where you came from, how you got to where you are. So for you, do we start in Orlando, in Lagos, or in India?

SS: At least Lagos. My father, in his early 20s, moved to Lagos and built one of the first pharmaceutical manufacturing facilities in all of Africa. He set up this whole life there for the better part of ten years. I was born in India, but really I grew up until I was three in Lagos.

And then there was this horrible armed robbery that happened at the house — they decapitated the dog, pistol-whipped my father, and threatened my mother. So we left all of our earthly possessions behind and started over again in Orlando, which ended up being a truly phenomenal place to grow up. The profound optimism of Disney World, the shadow of the Space Coast, seeing shuttle launches in elementary school, running into your courtyard filled with this sense of optimism about how technology will inevitably lead to a better world — which, if you fast-forward twenty years from the ’80s, I think kids were fed a very different diet of Terminator and “nature always finds a way“, “How dare man try to do anything that creates progress?”, so that’s really where it started.

Is that actually the ultimate example of man overcoming nature — not the space shuttle, but in fact Disney World?

SS: (laughing) That’s true! The fact that millions of people go there every year, to the middle of the swamp — and it’s fantastic, a lot of fun.

What does meritocracy mean to you? I was reading the recent profile of you in Colossus, and it opened with [Palantir CEO] Alex Karp’s statements in that regard, but it was put in the context of you growing up in Orlando, taking an entrance exam to go to Trinity High School, and how that launched you into what you became. Is that a sentiment from Alex, is it from you, or what’s your take?

SS: It’s a sentiment for both of us, for different reasons. We’re both kind of acquired tastes, in many ways, and living in a world that judges you by the fruits you bear, rather than any dimension you have no control over, is obviously a better world to live in. I think it’s a precondition to the American Dream.

I was going to ask you that. Because you have this background — born in India, originally living in Lagos, come to America. What does being an American mean to you? Is that the same question, or is there some tension between those two things?

SS: They overlap profoundly. Being an American is this profound belief in your own human agency, and the community you’re in, and the ability to create a more perfect union — the recognition, the ability to hold the contradiction, that not everything is perfect today but it’s incumbent on us to make it better. That may sound a little Pollyannish and even corny, but that doesn’t exist anywhere else. Everywhere else, people are profoundly cynical — things are the way they are — and there’s a very thin line between cynicism and nihilism, and you kind of give up on everything. That’s the death spiral, that’s how you get the Dark Ages.

After university and graduate school, you worked on Xoom, on remittances — and then you go to work for Palantir.

SS: Xoom with an X.

Xoom with an X, that’s right. Not the Zoom with a Z — which maybe I shouldn’t be talking to, since you’re now an executive at Palantir using Zoom with a Z, but that’s another discussion. You worked on Xoom and remittances, then Palantir, working on national security, arguing for building our homegrown industrial base. There’s an argument there that — to take a paradox that’s maybe unfair — how do you go from a career focused on sending money out of America to being an advocate for money here in America, rebuilding our capital base? Is there tension there, or am I creating paradoxes out of thin air to build a theme?

SS: I think you’re creating a bit of a paradox out of thin air. First of all, there’s a policy question — if you have a working class trying to get money to support their family, how can they do that as efficiently as possible?

The bigger issue is the profound lie of globalization. The calculus that globalization got wrong is this idea that we can do the innovation and “they” can do the production — “they” make the things, we do the innovation. The problem is that innovation is a consequence of productivity, if you stop making the thing, you cede the stimulus required to do the innovation around the thing.

The example I always give young computer scientists is, what do you think motivated Google to do the research behind “Attention Is All You Need“? If you talk to that team, they wanted the 3% improvement in Google Translate. You cannot think of something more prosaic leading to something more profound. So we should be very careful where we give up — we buy a little too much into comparative advantage and give up the stimulus for production.

Why do you think — what leads to that mistake of thinking that it’s okay to just own the inputs and not worry about the outputs?

SS: Because at the margin, it’s true. At the margin, you can hire WuXi as a cheap set of hands at the lab bench to do a little bit of work. But at the limit, that’s a transfer of knowledge. At every increment it seems to make sense — you’re still doing the innovation — until one day you wake up and realize, “Oh, I don’t have it anymore”.

Diffusing Intelligence

You’re the only Stratechery interview that I’ve offered a job to. You sent over a list of questions you wanted me to talk about, and I have to say they’re really good — I did ask if you wanted to prepare all my interviews going forward, I don’t think I got a response, maybe the email got lost.

SS: (laughing) I’m thinking about it.

Okay, good. Well, let’s drill in — we’ll take your outline, why not, go through it and talk about it, and I think there are themes that might emerge. So number one, and this is from Shyam Sankar, does this count as a Shyam bomb, by the way, the long extended email detailing what we should talk about?

SS: It’s not nearly vitriolic enough, so you’re spared.

Got it, that makes sense. Number one, is the AI industry solving the wrong bottleneck? You argue we are over-investing intellectually in the supply of intelligence and under-investing in diffusion. Tell me what you mean.

SS: If you thought about a crass metaphor of AI through electricity — there are people who generate the electrons, the power producers — for electricity to have any impact on our economy, there needed to be other people who built machine tools that run on those electrons. And that led to a profound revolution. Not only would you introduce one piece of machinery into your factory, you’d start redesigning your factories around the fact that we have electrons.

Let’s call that diffusion, how does the miracle of electricity diffuse into the economy? There’s a similar question with AI. I think we are very focused on how to make better models, smarter models, more machine intelligence — and we should keep doing that — but we’re very under-indexed on how to accelerate the rate of diffusion. Where is the economic value being created? How do we turn these tokens into economic value more quickly with a great ratio around it? I’m obsessed with that, because I think there’s a license to operate in society where the benefits of this stuff diffusing really do matter and I think we’re getting the calculus around some of this pretty wrong.

So who is “we”? Who are the ones insufficiently focused on diffusion? Who are you trying to convince here?

SS: I think the labs are, but in general the entire market structure. People are being rewarded — and it’s very rational, if you’re a lab, you’ve been in your own reinforcement-learning loop, you do fundamental research, you make a more powerful model, and your valuation doubles or quadruples and every year it seems to keep working. Baked into that is the idea that eventually this stuff will diffuse, eventually there will be economic value.

But one of the conceits I’m not sure people are getting right is, what is the atomic unit of the enterprise? Let’s leave consumer out of this. The presumption, as expressed through the product strategy you see from the labs, is that the atomic unit of the enterprise is the knowledge worker — if I can make every worker more productive, that’s the ROI, and that’s not actually true. The atomic unit of the enterprise is the workflow, you can make a single supply-chain analyst more productive and generate no incremental ROI, you have to make the entire supply-chain workflow measurably more productive to capture the value.

There are some roles where the cardinality collapses — the knowledge worker is the workflow, like coding — so you have perfect product-market fit for those things, and you can chase unbounded productivity. But that’s a very small part of everything we do in the real economy.

Again, I don’t understand the audience here. In a market, shouldn’t the model makers just do their job, which is making better models, and then there are lots of opportunities for companies to take those models and diffuse them? Isn’t that what Palantir is trying to do? Who is the target of this exhortation?

SS: To those wondering, “Okay, we’re creating this, we’re spending trillions of dollars, where’s the return, how’s it working, why is the best example of excellence an AI assistant, which is nice and convenient but that’s not going to inflect our GDP?”. You think the big problems that I’m interested in solving — I think one of the fundamental issues, if you zoom out a lot beyond globalization. In 1971, the relationship between GDP growth and wage growth broke. Before then, every time the economy grew, wages grew. And now, adjusted for inflation, wages have been flat since.

That’s a profound problem, a lot of our current challenges are downstream of that reality. There are many definitions of a thriving middle class, but the one that resonates with me is the most qualitative which is you believe your children’s future will be better than yours. If you go a couple generations without being able to cash that check, you get massive societal instability.

Why does this matter to you? Why does it matter to my audience? Is this your mission here? Why do this interview? Is this concern about the overall public perception of the tech industry? Is AI going to lose public support? Is it about people opposing data centers? Where does the rubber meet the road, from theoretical discussion to actions you want to see happen in the real world?

SS: Is AI just a dope technology, or is it something we can actually use to make our society better? I’m trying to solve for making the world better, our society better, through backpropagation. What would have to be true of AI for these things to be possible? As cool as AlphaFold is, there are zero drugs on the market because of AlphaFold. It’s dope — super dope — but nothing is better because of it, it would be a travesty to turn our back on AI in such a way that all it can be is something really cool.

Top Line vs. Bottom Line

Number two of the Shyam interview outline, why are ambitious customers winning? Why do some units and institutions put AI to consequential use while others struggle to move beyond experiments? How much of the difference is access to technology, and how much is leadership, urgency, authority, and an imperative to win?

SS: This is one of the empirical things I’ve observed. It wasn’t obvious to me it would be the case, but if I segment the customers I’ve worked with — the ones who want to use AI to reduce headcount, to take costs out of the business — almost all of them have failed at even achieving those goals.

If you compare it to the customers succeeding wildly, have real ambition, “I want to use AI to dominate my market”, “I want to use AI to put my competitors out of business”, “I want to use AI to do work that was always beyond the efficient frontier — things I always wanted to do but could never afford to do”. I think the conclusion I’ve come to is that there’s insufficient stimulus to motivate you to do the hard things when you have a myopic goal like getting rid of humans. There might be other factors too — human intelligence and machine intelligence aren’t exactly the same thing.

How so? How are they different?

SS: They don’t do exactly the same things in the context of the workflow. If you spend your time trying to delete the human, by definition the AI would have to be able to do everything the human does — even if it only does 98%, you’ll fail at the specific goal.

Another way to reframe it, “I have this miracle intelligence now, how do I build the workflow around what it can do, rather than shoehorning it into the error that always exists between the human and the AI?” — the AI maximalist would say, “I’m just going to keep training a better model”.

But I’d say, what percentage of the problems we have today can we solve without going beyond the current frontier? I think we’ve left a lot of problems unsolved because we’re not thinking about diffusion enough, and wer’e not actually trying to maximize the AI for what it can do, we’re instead trying to judge AI by the standard of, “How do I eliminate humans?”, which may not even be that valuable, judging by the people who are succeeding with it.

Does this lesson of the importance and power that comes from focusing on the top line versus the bottom line — was that a lesson you learned in the age of AI, or something you learned working with companies before current LLMs came along?

SS: It’s very pronounced in the age of AI. Before then, you’d say it’s more exciting and you succeed more often going after the top line, but you’ll still succeed going after OpEx reduction. With AI, the outcomes are much more binary. It’s going to lead to a category of winners and a category of people who view this as something that’s not quite working.

Safety and Systems Engineering

Number three, who carries the burden of systems engineering? A capable model is not a working system — despite lab blindness, the idea that there’s anything other than the model. Who owns the integration of probabilistic models with deterministic software, human judgment, permissions, and real-world consequences? Explain that.

SS: Just this weekend there was another article about how LLMs are able to communicate using timing differences in VMs for CPU cache. Again, it’s a complete absence of a systems-engineering approach to the problem. It’s saying, “Can we test and find things that can go wrong?”, rather than viewing it the other way around, “What would we have to build around this so we could actually use it to capture the economic value that exists?”. There’s a ceding of human agency and responsibility around solving these as systems-engineering problems.

The tell is you look at these historical incidents — in a real company, you’d expect someone to publish a “Five Whys“, we’ve done a retrospective, I’m documenting not just what happened but everything I will do and build and fix and change so it can never happen again.

Think about how [Hyman] Rickover built the nuclear Navy. These two positions are interesting to juxtapose, Admiral Rickover building the U.S. nuclear Navy, and the Soviet nuclear fleet. The Soviets had to rest their submariners every six months in Sochi so they could regenerate their white blood cells, because they treated safety from radiation as an unsolvable problem, “It is what it is”, and Rickover designed our submarines to a standard of safety a hundred times more than what scientists thought was even needed. Very clearly there you have a human who absorbed the burden of systems engineering, and a philosophy of continuous improvement.

I think companies are going to have to absorb that, and the people who provide AI are going to have to absorb it too. All of these problems are solvable. We could wring our hands and talk about how dangerous this new thing is — we could have done that with the automobile. Four million Americans have died since 1899 with the invention of the automobile, but we just kept going, we have seatbelts, airbags, anti-lock brakes, there’s a systems engineering approach to reducing risk, dare I even talk about the airplane. You don’t see that coming out in the discourse — other than, say, Power Shelf from Nvidia, who has actually released toolchains meant to provide more safety? There should be a proliferation of that right now.

Do you think that falls on the model makers to do that? Or does it fall on companies like Palantir or Microsoft? Or to what extent is it in the interest of Palantir and Microsoft to frame it that way because suddenly the value of what you’re wrapping around the models becomes much more apparent?

SS: Well, certainly that’s the business we’ve ended up building, so you can say there’s some self-interest there. But I’m kind of gaslit by how little focus there is in the labs on this. It comes down to model blindness, there’s an aesthetic where if you need anything other than the model, that’s somehow a failure of the maximalist version of AI, and we should just be building more there. If you were more focused on the practical realization of benefits for your customers, you probably wouldn’t think about it that way. To what degree are the model companies really selling progress, not product?

Isn’t that good for you?

SS: It’s great for me!

Palantir’s AI Lessons

Well, here’s a question. You talk about the need to focus on the top line rather than the bottom line, and the need to build things around what these models are capable of as opposed to retrofit them into a human replacement paradigm. How does that apply to Palantir broadly? You built this business, this concept of the ontology, all these disparate data sources, and you bring them in — and when we talked in 2023, it was a very compelling story about how this makes sense in the context of AI, with AI almost being a user interface across this data. At what point are you approaching AI and it seems awfully convenient that, “We’ve had it right all along and it fits right in” — especially in conjunction with a message that “You need to rethink what you do around AI”. To what extent has Palantir rethought what you do for customers because of AI?

SS: Enormously. On the big macro things, we were right. The enduring value of ontology — I was just with another big customer who was reasserting how ontology is what prevents the kind of AI slop from accumulating in their enterprise, and provides a layer they can build their supply chain on.

But it would be crazy not to think AI has gotten us to rethink lots of small things, including how the work gets done. The majority of work in the platform now happens through what we call AI FDE — using the models themselves to accelerate the programming of what’s happening in the platform. Also thinking about it as part of a toolchain.

If you zoom out, we view the enterprise as a state machine. Before AI, every one of those state transitions, some human was doing — the state machine is very complicated, and you’re interacting with lots of legacy systems. So the point of the ontology was never just a semantic representation of your data in your enterprise, it was also to model the verbs, the actions you could take. It made the enterprise programmable — a harness for humans to the enterprise.

The opportunity, and what we had to invest in, was turning that harness into something that models can use. And thinking about it idempotently, enterprises aren’t just going to rip out the humans and suddenly have AIs running the entirety of it. Some workflows will start with humans, move to AIs, come back and forth. So you need a state machine that’s invariant to how did this work start — mammal or AI — and how is this work going to end, and how I govern the transitions through that state machine accordingly and the actions that can be taken?

How would you have changed your approach to the ontology and what you’ve built had you known back in 2006 that LLMs would be the ultimate interface to what you’re doing?

SS: Since the core ontology is kind of enduring, I’m not sure we would have changed that. It has changed how you get value out of the ontology — the applications you build on it, how they relate to each other, and it’s changed how you get data and model systems into the ontology.

The counterintuitive thing, maybe a-year-and-a-half ago, I mandated that every new employee learn to use the platform through AI FDE in their first week, “This is the future, let’s go maximalist on it”, and it actually had a profound retarding effect over time, which was that there were lots of young employees who only knew the platform through this abstraction layer and, as a consequence, didn’t understand the underlying building blocks that the platform used and was composed of.

Perhaps the most important question, even as someone who is an AI maximalist, is knowing when not to use AI. It’s been very interesting, when you ask the AI to design ontologies for enterprises, Going back to the alpha-beta frame — when you ask the AI to design ontologies for an enterprise, it does a mid job. Going back to the alpha/beta frame, it’s all beta, it models you like it would model anyone else, smoothing over everything that makes you special or different. A reversion to the mean.

Alpha vs. Beta

Number five — we’ll come back to four — has enterprise software made companies more alike, and will AI make that worse? This is an argument you’ve put forward on multiple occasions, the idea of conventional enterprise software as beta, not alpha — software required you to sell to as many folks as possible to get the return on your investment, and if you’re selling to as many folks as possible, it’s defining workflows, so everyone ends up with the same workflow. Isn’t that fine? Aren’t most workflows non-essential to the top line — software is inherently a bottom-line endeavor, and in that case it’s fine?

SS: In the old world, that’s the only thing you could use software for. Go back to the earnings calls during COVID — what were CEOs talking about as the technology successes? Zoom, Teams. Meanwhile, many of these companies who I’ve known had just spent five billion dollars on an ERP implementation that fell over like a paper tiger two weeks into COVID. So the question is, what parts of the software-industrial complex are actually working? I don’t besmirch them — essentially they’re solving for, “Can I sell you the software?”, which is not the same thing as, “Did the software work?”, “Does the software actually matter?”.

You can juxtapose that to our entire approach to software has been solving for, “What software must exist to solve the problems I care about?”, it’s building software through backpropagation. That both lets you continuously up-level the ambition of the problem you’re trying to solve, and allows you to think through the abstractions you want to codify in software versus the parts you leave as what is going to be unique to the customer. That’s allowed us to pierce the trade-off between alpha and beta, “How do I build software that makes each customer more different, not more similar?”, and therefore is a strategic weapon they can use in the market.

So should customers be using Palantir for driving top-line revenue and still hiring a bunch of SaaS companies for everything else?

SS: Empirically, that’s the case. No customer I work with is a greenfield where we’re the only thing that exists. If you’re happy with the beta you’re getting but you’re beating your head against the wall trying to figure out why the beta doesn’t let you capture the opportunities in front of you — are you fighting your software stack? Are you fighting decisions made in the past, borne out of an aesthetic desire for something easily governable and maintainable by your IT organization? IT will always yoke toward beta — decisions that are mimetic, closer to best practice, what everyone else uses. The people most interested in alpha are in the line of business — the person who feels frustrated that their strategy is hard to implement, versus “I have a precise steering wheel to do this”.

So do models deliver beta or alpha?

SS: Depends how you use them. If you just think about it at the limit, we’re just going to keep betting that these models will get better and I’m going to rent my intelligence from a lab — how are they getting better? They’re somehow compounding all the world’s knowledge into their system, which you’re maybe tacitly helping them do that, maybe you’re not, but that’s the trajectory you’re on. There should be an existential question about, “What is the theory of my firm?”, “What do I do if all the intelligence is compounding into weights I don’t control?”.

A big part of what we’ve been pushing is this idea of Sovereign AI, that our customers have really taken to. It’s not a binary — you don’t need to leave the lab ecosystem. But for what workflows is actually the most valuable asset your post-training pipeline — it’s not about a specific open-weight model, it’s not about a specific post trained open-weight model, it’s actually that you have sovereignty over a post-training pipeline filled with the operational exhaust of your enterprise. The operational decisions that you’re making in the context of the operating system that is our software is generating data that allows you, every month, to train a new checkpoint and evaluate it against everything that came before, in a way where you know that model is yours that you’re going to be relying on.

The stakes of doing that in something like government were obvious, your competitive advantage is everything, you wouldn’t rent the brain from someone that your competitors could rent the brain from. But it’s equally true in commercial world where frontier capabilities seem to be a function of unique data that pushes you past the frontier.

If you talk to the Mercor’s or the data companies of the world, they’ll tell you the way they sell to a frontier lab is they generate a dataset that they post-train an open-weight model on to be provably past the existing frontier of the model the lab is selling. So if that’s how this game is happening, every company has this exhaust as a unique asset, why should they sit around and wait? And then as a consequence have a bigger question of, “If everyone has access to this, what is my alpha as a firm? Is everything reduced to beta?”.

That makes sense — you give everything to an Anthropic or to an OpenAI, that learning is going to benefit everyone, which is going to transfer your differentiated knowledge into beta instead of being alpha. But why is the correct stopping point having a partner like Palantir? Why shouldn’t they just do all of this themselves? Why is it that you are a trusted partner in a way the models are not?

SS: You can think of it as a model factory. What systems are you using today to make operational decisions? What system is generating the exhaust that has your decision-making processes and all the constraints that is effectively its own RL environment that you can go off post-train these models in? And as you post-train those models, how do you evaluate those models against historical decisions you’ve made, how are you going to manage deploying them to production?

Having the entire vertically integrated chain where it just works and you’re able to leverage — a lot of what we’re investing in now is auto-research, how do I, without the customer investing in a big post-training team, use the decisions they’re making in supply chain or in production planning, use that exhaust to automatically start training models behind the scenes and surface to them, “Here’s a model that’s better than what you’re doing”? How do I use every point where you’re using an LLM and use that as a trace that I’m collecting that feeds back on itself, so that you have this continuous improvement? That’s the compounding loop, I think every customer wants to be in it.

How do customers think about timing? It might be the case that in the fullness of time you’re just getting a median answer from a probability distribution, it is all beta, but if you’re first to beta, you can still make a lot of money. Is that the sentiment that you’re running like, “If we incorporate AI first, sure it’s all mid, but timing matters too?”

SS: To make sure I understand, the mid point is coming from the frontier labs? Or weaker post-trained models?

Let’s say you’re in a company that is considering it’s competitive set, who they’re competing with and they’re like, “Okay, Palantir, this is a good pitch, I think in the fullness of time you’re right, we should not be giving our knowledge to the frontier labs, we should be building our own hill-climbing capability, keeping it all internal, that’s what differentiates us. However, that’s going to take a while to get that set up. I could lean in to using the models right now, my competitors are not, if I’m first, I’ll have a short term advantage, I’m motivated by the short term, and that’s going to be enough”. Starlink on airlines, everybody is going to get Starlink eventually, but United is going to get a lot of benefit from doing it first.

SS: I would tell customers they should win today, the primacy of winning does matter. But the part I push back on is, “How long does this stuff take?”, I’ve actually been surprised — it takes a week, especially if you have the operational exhaust where people are already using you to make the operational decisions, to develop a new post-trained model you can put into production. So that’s your alternative.

Empirically, where we’ve done this, you get better-than-frontier performance at task-specific things. Say my business is selling electronic components, when people submit a request to purchase components, they submit an electrical-engineering diagram — pretty technical, definitely out of distribution for the frontier models, it requires a vision-language model. I have all my historicals, what the input is of the spec, and what the BOM [bill of materials] was that I quoted back out that a human built over the last 10 or 20 years. It takes a week to go train that model, and now you’re able to respond to your customers with a price faster than your competitors can, because it’s going to tak them weeks to work their way through the same specs.

Presented by

When I write about tech companies, I am often writing about businesses that depend on Stripe, because Stripe provides the foundation for Internet-native companies — including Stratechery.

Consider Substack: building Stratechery was complicated and involved gluing together disparate pieces; Substack made it easy by delivering an integrated offering. One piece that was the same, however, was Stripe: Substack creators are also Stripe customers, as are creators on many other platforms.

Zoom out and the pattern repeats. Fintechs can build entire business models around card programs without ever having to build the issuing and processing infrastructure underneath it.

AI companies can charge for what customers actually use, which is particularly important when usage also drives costs, without having to build the systems to track and bill for it themselves.

No business succeeds because of Stripe. Stripe doesn’t answer the hard questions about what to build or whether anyone will actually use it. But it does make more things practical to try.

Model Questions

Number seven, could successful AI adoption be bad for frontier-model economics? If a workflow succeeds on a frontier model and then migrates to a smaller model the customer controls, or to ordinary software, who captures the value?

SS: I think the customers will capture the value. I don’t think it has to be bad, but this is the fundamental question going back to something I said earlier which is, to what degree are the frontier labs selling progress that the next model will always be bigger and more powerful and more interesting than the prior model, versus to what degree are they selling a product? The more they’re selling progress, the more customers are rationally going to take non-frontier use cases off their workflows.

Is it inevitable that the labs have to come for all software companies, because of the commoditizing nature of their models?

SS: I don’t know if it’s inevitable, it’s an open question, there’s still a question of how much value do you realize from incrementally more powerful intelligence? Right now we’re in a peak phase of not being able to distinguish between things that are dope, every new model release, I love, it’s really cool, but I’m also not seeing a commensurate amount of translation of new economic value unleashed that I couldn’t have previously gotten before.

Why won’t frontier labs just own the entire inference layer from top to bottom, from smart models all the way down to dumb models? If it’s ultimately a commodity market, won’t they have a superior cost structure given they have the best AI to build the best inference anyway?

SS: If it’s just about the cost of inference, probably. The question is whether you’ll have more and more things that are out of distribution, and that you get better performance from models you own, and to what degree companies will be comfortable on use cases that do represent alpha?

So is this the key thing? Is your reason to be out here to exhort companies to stop giving their data to the model makers because, all data being equal, they will win?

SS: I definitely think that’s true. Partly it’s to make sure companies know there’s an alternative for the use cases that aren’t solved today and to the extent you’re wondering where exactly is the ROI going to come from, why is it that the only uptake I’m really seeing is coding?

I’ve had these two out-of-body experiences, probably bad form to talk about, with each of the frontier labs, where different people were saying, “You should really see the work that we, the lab, have done at this company”, but the reality is it was work I did. Yes, I’m spinning the token meter, but the translation of the model to economic value for that company required someone who actually had a sharp opinion on diffusion and the realization of that value.

Is DeepSeek the most effective Chinese company there is? Particularly given their MIT license, and they just open sourced their inference stack. Do you use DeepSeek, how do you refrain from using it?

SS: We don’t use it ourselves, but we let our customers in the commercial world have choice in the models they use. While I encourage them to use the Western stack for lots of reasons, and there’s a burgeoning Western ecosystem that’s definitely good enough, there’s a degree to which people have over-indexed on benchmarks and the bench-maxxing that’s happened.

When Nemotron 3 was released — just vanilla, non-post-trained Nemotron 3 — we brought it into our own stack that we use to manage our internal operations, and there were five use cases where it was better than frontier, which should not be possible if benchmarks were a true, faithful representation of intelligence across its full jaggedness of it, should not be possible. So I lost confidence in benchmarks, and started this movement internally that it’s not about bench-maxxing, it’s about bench-making.

The cardinality of benchmarks should be at least one-to-one with the number of customers in the world — what is your benchmark, what do you care about, how do you evaluate models against those things you care about? That’s a very empowering frame, getting out of, “What’s the latest, coolest model?” versus, “What actually matters for me, how much improvement is there, how do I trade off latency, cost, performance against all these dimensions?”.

Why don’t you make your own model?

SS: If I felt we had the need to make one, I would. Mostly since we exist in the context of our customers, I help them post-train models that solve their problems. I’ve toyed with and think there might be opportunities to create models more intrinsic to our product, that are distributed to all our customers that solve very specific problems, or a starter checkpoint they can start off of, that I think gives them a running head start against some of the problems they care about in domains with unique phenomenology in particular, things way out of distribution.

I could see your overall argument, we’ve talked before, and we both love looking at industries in terms of the value chain and where it actually matters. I think your overall argument, and I’m trying to play Devil’s advocate here, I absolutely see the value in the data and owning the ontology, that and chips being the inputs all resonates with me. At the same time, there’s an aspect of controlling your own destiny, the essential inputs, and you opened this discussion talking about the problem with globalization — the idea that on the margins it made sense, but in totality you lose the plot. It’s interesting to the way you’re concerned about the models whether that be from an ideological perspective, from a them taking your exhaust perspective, from a them backwards-integrating into what you do perspective, why aren’t you investing to have an alternative, or contributing heavily to those who are? I’m sure Mark Zuckerberg would take some money, I’m sure Jensen [Huang] would take some money. Is there an aspect of putting your wallet where your mouth is?

SS: The business case would be, I’m not sure the value accrued to the models. Certainly the residual compounding value is what we provide our customers.

But then what’s there to worry about? It’s not going to accrue to the models anyway, isn’t the market going to take care of all these concerns?

SS: I’m not worried — I think it will. Partly my exhortation is around, if we’re in a rush to realize the value of this genie of AI, are we actually putting all our emphasis and chips in the right place? We should be careful not to think of the Chinese as being ten feet tall but if you juxtaposed it with the profound focus the Chinese have on AI is about productive forces — how do we actually get economic value out of this? You can say that even the commoditization — one there’s an aspect of economic warfare through distillization attacks, we’re going to release an open-weight model and destroy a huge amount of U.S. market cap but the other part is, it serves as a stimulus and subsidy for the rest of their ecosystem, where they want to get better or are really good, from power to chips and I think it won’t be that long before the Chinese models are not open-weight.

Why is that?

SS: Because as they start to feel they’re at the frontier, or attempting to get beyond it, they’re going to try to monetize it and close it down.

Do you think they’re at the frontier?

SS: No, they’re behind it right now.

Can you be a Chinese model maker and not be totally dependent on distillation? Doesn’t that leave you completely uncompetitive with your peer set?

SS: Yeah, it does. Which maybe is the other aspect of it is how the CCP views the models as a threat to their own security and stability.

There feels like there’s an aspect on all sides of, “Let’s just trust the market and our intuition”. From my perspective, I wrote Who’s Afraid of Chinese Models? this summer, which got at the fact that everyone is stuck on API prices, which of course include a profit margin. At the end of the day, I want to know the actual cost to serve. My assumption, I think a reasonable one, is that the frontier labs can serve very efficiently, in part because they have that capability, serving a very large Chinese model costs real money, it’s not like you can serve it for free. By the same token, this assumption that because they accelerated to six months behind, or three months behind, whatever you want to say, that it’s inevitable that acceleration will surpass — when the distillation explanation both covers exactly what has happened and continues to cover what has happened. When you think about the overall economic forces of the Chinese labs, they have to depend on distillation, if you develop a dependency on distillation, you won’t develop the capabilities to be on the frontier. So why can’t we all just relax?

SS: Because winning AI is more than having the most powerful model. Maybe this is the fish-in-water thing I didn’t articulate, why do I have all this angst about how quickly we’re diffusing? Because that’s the whole race, the whole race is who realizes the most value and is in the best compounding loop around the things that matter, as opposed to the dope fact that we have a cooler and cooler, more powerful model.

Got it, I finally got you to answer my first question, I just had to go around for a while.

Forward-Deployed Engineering

Are the labs copying forward-deployed engineering, or just the job title? Palantir is very famous for this — you specifically, the most famous for this — what is a forward-deployed engineer as you developed it, and does that apply to what the labs are doing in terms of creating these organizations that go into companies and help them adopt AI?

SS: At this point it’s turned into the most popular job title, it goes beyond the labs, really.

What a turn of fortune, huh?

SS: For 17 years I was maligned as the person ruining Palantir because I created this horrible concept of the FDE, and then it became the coolest thing ever. But the challenge is, before you say, “What is a forward-deployed engineer?”, you have to say, “What is forward-deployed engineering?” — and that actually it’s a product strategy, not a role. There’s a role, yes, but it’s way downstream of the thing that makes it happen.

The reason I think it’ll be very hard to copy is there are probably five people in the history of Palantir who know enough about the strategy to actually recreate it anywhere else. You can’t just go hire someone who was an FDE and do this. Just even as a starting point, forward-deployed engineers are not part of your go-to-market organization, they are part of your product organization.

Right, they’re deciding what gets built by seeing exactly the problem the customer is having, communicating directly or being part of the product team and that influences what is rolled out next.

SS: It goes through multiple layers, you’ll have to indulge me as I try unpack it.

The first part of it is if your product exists on some Gaussian distribution of features, “I’ll average the signal across my thousands of customers and decide what to build next”, forward-deployed engineering is not the right strategy for you. If, on the other hand, your customer set exists on some power law, where for a set of problems you’re solving, there’s a small number of customers with the hardest version of the problem and if you can solve it, you get the entirety of the market, because everyone else will have this problem, they just don’t know it yet, some are living in the future — then you have to go from specific to general.

The journey from specific to general is very hard is very hard, engineers hate the notion of wasted work. But you get from one to two to three, and you’ve earned enough surface area to understand the general form of the solution that’s going to work for the fleet. It’s a heterodox way of thinking about strategy, and it requires thinking about the shape of the engineers who do this work differently.

Engineers exist on a broad continuum — this is a caricature, intentionally so, to be illustrative — on one end you have the hacker, the MacGyver, the person whose dopamine hit comes from solving this problem in the world.

They’re the best duct-tape users in the world.

SS: Absolutely! Never mind, it’ll only work once or twice, but, “This thing is true because of me”. They have a knack for figuring out which problems are worth solving, which is hard, let’s call it the first 80% of the problem.

There’s a different skill, call it the artist, the other end of the dialectic, the engineer whose dopamine in life hit comes from how beautifully built the thing is. Never mind it solves no problem anyone cares about, “Don’t you see how elegant it is?”, this dialectical tension between two extremes, and how you get to the second 80% of the work. Each personality undervalues the contribution of the other, so it’s held together as a dangerously unstable equilibrium that has to be managed.

I think we’re up to 96% if I’m doing the math right, so what’s the last 4%?

SS: No, 80% + 80%, the whole point is the math doesn’t math, because everyone thinks they’re doing the majority of the work.

So the answer is, someone sees a problem, doesn’t just understand the problem and what to fix, but how to fix it systematically so that it’ll be a sustainable fix not just for that customer, not just that time, but going forward, for all customers.

SS: Yeah, and earning the opinion on where the abstraction should sit in the stack so it can actually generalize because building an end-to-end solution is overfit to one customer. So what are the parts you’re actually going to generalize, and what parts do you let be customer-specific? It’s engaging with an inherently messy process.

I think a lot of folks are jamming a sparkling form of sales engineering, they take the FDE role and say, “We have really smart engineers who can go work with customers”, but there’s no way of metabolizing the pain the FDEs are experiencing into product, if you don’t have that, you don’t have FDE — it’s just sparkling sales engineering.

Two questions to follow on this. When you talked about the Gaussian distribution of customer problems that lends itself to, “Build one thing that broadly solves a problem for everyone”, in many respects that’s consumer software. The best consumer software companies and consumer software leaders do not rely on focus groups, because you can’t get a focus group on eight billion people — you have to have your own sense and drive the software in a particular direction and believe in your vision. There used to be a phrase, you’re old enough to remember this, the consumerization of IT. Was the consumerization of IT, in your view of the world, the application of this “one solution for everyone” mindset, when engineering IT back in the day was line-of-business apps, it was people actually building software for their organization? In the SaaS-ification of everything, the spreading out of R&D across as many customers as possible, that is the issue that you’re pushing back against, and that’s actually what happened with the consumerization of IT. It wasn’t just about user interfaces, it was about one-size-fits-all.

SS: I’d agree with that. But I think the correct point is an unstable equilibrium between the two. There’s a constant need for the masses to revolt against the high priests who have their perfect conceived notion of how things are supposed to work but they don’t work.

Part of the problem is there was an economic rationale for this in that developing software was exceptionally expensive, so you needed to spread it across everyone, and because you had leverage over enough customers, you could do an 80% job that was better than a dedicated individual. I think the argument for AI is actually no, you can do small-scale engineering projects that solve 98% of the problem instead of 80%, and get back to a “many flowers blooming” approach much more akin to the ’70s and ’80s in terms of enterprise IT.

SS: I think that’s right. And again, you’ll find the pendulum keeps swinging. There’s a proliferation where people go all-in on Codex or Claude Code in the enterprise, a proliferation of apps that don’t speak to each other — almost like attritional software. You built it to solve a problem, you either kill it or you accumulate debt over time.

So what constrains that entropy? That’s what we think ontology does. The ontology provides that constraint point you can govern, so these things can interact, you can have a lot of freedom to vibe-code an app, and the cost of it operating in your enterprise is low, the cost of throwing it away is low. So you can empower your builders without giving up your ability to have a cohesive operating system for the enterprise.

Here’s a defense, or a question, about the AI labs putting forward their versions of FDEs — or was it “sparkling sales representatives”? Isn’t the actual FDE the model itself? It ties into your overall concern, the way for them to solve your problems is to actually absorb all of your data. So yes, the people are not the FDEs, the AI is the FDE, and that’s the actual FDE you’re worried about?

SS: I’m actually not worried about that now, this goes back to my experience of, “Let’s just have AI FDE be the entry point for everyone in our platform”, it’s all a mean reversion. There’s a lot of value in getting the problem decomp right, and if you get only an okay problem decomp, you leave a lot of the alpha on the table. That’s probably the part I’m least worried about going away.

Anthropic vs. the U.S. Government

What are you most worried about? What has changed between 2023, when we last talked, and 2026, today? Are the models more capable than you expected? Less capable? How do you view the world differently now than you did before?

SS: The models are more capable. I don’t know if I had an expectation — it was so hard, in a world before AI, to truly have an expectation, but it’s amazing, I’m delighted every time I see a new model release and get to play with it and get to see what’s possible.

What I’m more convicted of is that there’s not a one-for-one overlap between human intelligence and machine intelligence, and the more we’re willing to build things that make sense for machine intelligence alongside the humans, the more value there will be. I see this in subtle ways. This might seem orthogonal, but it’s very relevant — I have this parts manufacturer for the submarine industrial base, and it used to take them two weeks to do their production plan during which the line was down. With AI, it took less than ten minutes, and now they have all this excess demand they didn’t have people to go do so they hired a third shift.

There’s an army of blue-collar laborers who have their job because of AI, not because they use it, not because they think about it, they’ll never tie that back together, it’s the opportunities in our economy I don’t think people are going after hard enough.

Just one question, I just mentioned the Anthropic thing — a lot of this feels, at the end of the day, feels like a lot of this conversation is ultimately about Anthropic. You were very close partners with Anthropic — the whole using Anthropic in the U.S. national-security apparatus thing was through Palantir. Do you regret that? When the whole showdown happened with the Department of War, were you stuck in the middle, or decisively on one side or the other?

SS: Ben with the tough questions. Yes, we were stuck in the middle. I don’t regret it. We still work with Anthropic, we work with both frontier labs. This speaks to AI sovereignty right on the head which are who gets to decide how these things are used? I do think there’s a question of epistemic humility over these issues, which I’ll get to in a second.

The customers have to decide if they’re going to be okay with the terms of use of a model or not, and what will the market accept ultimately. I’m not sure it’s going to be sustainable for model companies to have long, nuanced arguments over, “How will you use this thing?”, it’ll be much more binary, do you want to work with this institution or not? You have to trust your customers, to certain degrees around these things.

It was probably not handled excellently by many parties involved, but it did surface an important issue here, when the intelligence is coming from one place, what are the conditions do you need to know you can depend on it in a mission-critical environment — whether that’s the factory floor or a government organization?

Did that change what models you’ll work with — or is that still up to the customer?

SS: I want that to be up to the customer. What it did change is we spent a huge amount of energy and effort on being able to safely swap models out, and being able to exercise that sovereignty. You need that just as much to go between two frontier labs as between a frontier lab and an open-weight, post-trained model that you’ve created. Creating that scaffolding gives customers assurance, if you thought it was a two-year journey to do one of these swaps, the decisions you’re making are effectively a one-way door. If it takes two minutes to do safely against an eval set for a specific set of workflows, you don’t have to be hung up about this in the same way. I think I’ve created a lot of trade space for my customers as a consequence.

What was a bigger effect in developing this capability — that episode with the Department of War/Defense, whatever you want to call it — or was it that whole Fable “No more zero data retention, I promise it’s for safety” angle? Was that the one that actually woke people up?

SS: The second one woke up the broader commercial world, because until then it seemed like some narrow government dispute, “What do I care about that?”. The first one allowed us to build the product that we needed to enable that so when the second one happened, we were able to give our customers choice, to make decisions inside the OODA loop of “Holy shit, I just woke up to this reality, what do I do?”.

Did that apply to you? “Holy shit, we just woke up to this reality — what did we do?”

SS: Oh yeah, a hundred percent.

So was that actually the best thing that happened to you?

SS: I’m not sure if it was the best thing that happened to us, but we felt like we had seen where the world was going, and the sorts of trade space we would need.

There’s a clarity that’s come, I think, from Alex [Karp] and from Satya [Nadella] in particular, who have basically been saying the same thing. Maybe I should ask you, is Microsoft actually your competitor in the long run? Both of you are talking about this concept of being harness-makers — the ontology, Microsoft Graph, whatever you want to call it, having the data in a controllable way, building your own hill-climbing capability, so you’re absorbing your own learnings rather than giving them away to a model. But it felt like I wasn’t getting that message clearly from either of you until about six months ago, and it’s like no one was really paying attention to this until then.

SS: I’ll give you the inside baseball. Before Alex’s CNBC interview, it was the end of the quarter, he’d probably visited sixty customers, and every customer in private was telling him this, but none wanted to say it in public. It was very clear that there was this underlying frustration and belief that no one was willing to put their finger on. That’s part of why it had so much resonance, and why he delivered it with so much conviction and certainly, if you look at what we’re saying, we see the world in very similar ways to Microsoft.

AI Religion

Number four, I told you we skipped this one, has effective altruism turned AI safety into a secular religion?

SS: Clearly I think so. One of the things I’m worried about most, this goes back to the systems-engineering critique is that you have this weird rhetoric — the argument is essentially, “AI is going to create utopia, but of course there’s some chance, between 0 and 20 percent, of ending humanity”, so dystopia. These seem contradictory—

They’re not contradictory — they’re two sides of the same coin. It’s a rejection of any middle pragmatic path towards absolutes. It’s extremism, just manifested in two different ways.

SS: I’m preaching to the preacher here. It rejects all human agency in what’s happening right in front of us and therefore the obligation toward systems engineering. It influences how you even think about testing these things, what you think would have to be built around these things.

I don’t begrudge anyone their articles of faith, but when your articles of faith are represented to the world as technical critiques, it’s very dangerous and risky. People think, “Who could know more than this person who works on this thing?”, and that would be true if it were a technical critique, instead of a religious perspective that comes out of a unique worldview. It’s a combination of a millenarian cult with a messianic complex. It’s, “This thing is going to end humanity, except if I do it, we have some chance of working out”.**

You are preaching to the choir — I’ve written about this. You have to reject the frame completely. You don’t argue the relative chances, the expected value of utopia versus dystopia, you question why those are the only two options that might come out of this.

But here’s the question: for many years people would say Palantir is maybe akin to a religious organization. Today defense tech is hot, but when Palantir came along it was not hot, you went there because you were a true believer. We may have talked about this last interview, we’ve certainly talked about it at some point, and the power that granted you as an organization to push through all these problems, to figure out all these issues, because you believed you were doing something important and meaningful. As you look back, how important was that belief system — the shared belief system — to Palantir’s success?

SS: I think it was crucial. Part of the reason I can see this so clearly is that cult recognizes cult. We were a very different sort of cult, in the service of the West, in the service of democracy, a belief in state capacity, that the legitimacy of our institutions depends on the people who been elected to run them having the tools to run them in a way that serves the will of the people or their shareholders. That national security and prosperity are two sides of the same coin, national security is not an end unto itself, it’s a means to prosperity and human flourishing. You’ve got to get these things right, particularly in a post-terrorist world where every one of these events was a one-way ratchet toward less freedom and less security somehow. So yes, it was crucial to getting us to work, that’s why cult recognizes cult, although they’re very different cults.

Is your critique that this is a religious movement, or that the religion is wrong?

SS: That the religion is wrong.

So this is a religious war. There’s a War on Terror callback for the people who know.

SS: It is a religious war, but I’m not interested in fighting the religion head-to-head, I’m interested in making sure that people listening to the doomsaying have an understanding whether they’re hearing a technical critique — the presumption everyone has — or a religious perspective on a future world that may exist, and then let them decide.

How is it the case, though? It’s really striking, I’ve been picking at you and teasing at you that all your arguments, which I believe are heartfelt, do happen to perfectly align with Palantir’s business model. It is in your interest for companies to understand and believe that the data layer is what matters, they need to control that, that the models need to work in service to them, not throw your exhaust to the model, you’ll get the median distribution anyway, etc. I think that’s all valid, and it also perfectly aligns with what’s good for you.

By the same token, the remarkable thing about Anthropic — which just filed their S-1, I haven’t had a chance to read it yet — but all their fervor, doomerism, and the things they put forward that you’re critiquing for not being technical, happens to perfectly align with their business model. It aligns with what they’re selling, and, interestingly enough, often with their customers’ interests — their customers don’t view Anthropic as a competitor, “Better they’re trying to create God than trying to compete with my SaaS application”. Is there a certain extent of game-recognize-game here?

SS: The part I’d say is we’re all still founder-led institutions with a huge amount of agency over our business strategy and what we’re going to do. And the best way to run that is to align your strategy to things you actually believe in, all these things are all so hard. So I think the business is downstream of the fundamental beliefs you have.

So at the end of the day, what decides who wins? Is it who believes harder, who works harder, or will fundamental market forces come to bear? You’re making a lot of arguments that companies that don’t pursue the median, that preserve their businesses, are going to succeed in this world, and in which case, isn’t everything going to be fine?

SS: Unless we engender non-market forces that are reactionary. Like wrong safety regimes, a drive toward nationalization, we have data-center moratoriums that are engendered through a doomerism belief structure, we’re not able to rationalize and execute on delivering the benefits of AI to the average person in a way that lets them participate in the economic upside. All these forces, everything from a replay of the French Revolution to turning our back on the fundamental progress of capitalism are the greater risks.

Shyam Sankar, thanks for the interview questions.

SS: Thanks, Ben. Thanks for having me.

This Daily Update Interview is also available as a podcast. To receive it in your podcast player, visit Stratechery.

The Daily Update is intended for a single recipient, but occasional forwarding is totally fine! If you would like to order multiple subscriptions for your team with a group discount (minimum 5), please contact me directly.

Thanks for being a supporter, and have a great day!