# Palantir CEO Alex Karp says many believe we are in an AI bubble because ... *UninformedInvestors* ยท https://www.youtube.com/watch?v=UFD9ilfXCfM ## Summary Palantir CEO Alex Karp argues that buying large language models off the shelf is not precise enough for regulated work like underwriting, and that value comes from building a software layer to orchestrate them. He says the AI 'bubble' is really a lag where some AI works and some doesn't, and companies that deliver working products barely need a salesforce. He also explains that Palantir's growth is constrained by scarce, highly cleared technical talent and training bandwidth. ## Key points - Karp says Palantir's adoption is accelerating beyond its capacity. - He argues off-the-shelf large language models are a commodity and not precise enough for regulated tasks like underwriting. - He believes value comes from building a software layer, like Palantir's ontology, to orchestrate and manage language models in a language the enterprise understands. - Karp suggests the AI 'bubble' is really a lag: some AI works and some doesn't, and companies now ask how to make it work for them. - Palantir barely has a salesforce, and it keeps getting smaller, because working products sell themselves in a low-trust AI environment. - Exporting Palantir's work is hard because it requires people with the highest security clearance who are also technical, and such people are super scarce. - Training those scarce people takes time, and customers have to really believe in the work. - He cites the battlefield context, where most people assumed AI would not work but now it does. ## Notes ### Off-the-shelf LLMs are not enough - Karp opens by saying Palantir's adoption accelerates beyond its capacity. - He argues that buying large language models off the shelf and putting them on your stack "won't work" โ€” they are a commodity and not precise enough. - Example given: you cannot do underwriting or other regulated work with a raw large language model. - The adoption problem today is that people have tried things that can never work, then wonder why they fail. ### The software layer creates value - Karp predicts people, especially in America, will try to build something like Palantir's ontology, maybe by hand. - Once you build the software layer to orchestrate and manage language models in a language your enterprise understands, you can actually create value. ### Are we in an AI bubble? - Karp reframes the bubble question: if anything, we are in a lag where there is a lot of AI and some of it works. - In the battlefield context, everyone assumed AI would not work, but now it does. - The question has shifted from "does it work?" to "how can we get it to work for my country or my company?" - Companies now compare: it worked for that company, it didn't for mine โ€” what are you doing? ### Why Palantir barely needs a salesforce - Karp notes Palantir barely has a sales force, and it seems to get smaller every time he sees it. - This is not about saving on unit economics; it reflects a low-trust environment in AI where people have tried lots of things that haven't worked. - If you have delivered things that do work, the product sells itself โ€” in both commercial and government. ### Constraints on growth - It is very hard to export Palantir's work because they must train people and have limited bandwidth. - In government contexts, building something like Project Maven into your architecture requires someone with the highest level of security clearance who is also technical. - Most technical people do not pursue the highest level of clearance, so that resource is super scarce. - Training takes a while, and customers also have to really believe in the work and think it is important. ## Transcript [00:00] Well, I mean, our adoption accelerates beyond our capacity. >> So, it it the um I think if you just buy large language models off the shelf and try to do any of this, it won't work. >> It's a commodity. >> It's and well, also it's not precise enough. It like you can't do underwriting with a you can't you couldn't do any of these things that are regulated. So moving to like everybody there's not like the the problem with adoption at this point is people have tried things that just can never work like buying you borrow a large language model you put it on your stack and you wonder why you know it's like not working and so what you're going to see especially in America is people trying to do something [00:46] like what we do withtology maybe by hand because once you build the software layer to orchestrate and manage the language language models in in a language that your enterprise understands, you actually can create value. And I don't I I think like there's like there's a lot of discussions like are we in an AI bubble? What is the meaning of the bubble? Well, it's like I I think what if anything we're just in a lag where there's a lot of AI some of it works. Again, if you go back to the battlefield context, like it most pe everybody in the world assumed this would not work, but now it does work. And so now the question isn't does it work, but how can we get it to work for my country? And this is exactly [01:27] what's happening in >> companies. It's like, oh, this company it worked, mine didn't. What are you doing? >> Like for example, I mean, if you want to just make it like parochial to us, Palanteer barely has a sales force. In fact, it seems to be getting smaller every time I go see them. And it's like and it's simply because you're laughing because you know it says Palanteerians here, but it's it's getting smaller and smaller and smaller. And um and it it's not because we're trying to save on the unit economics. It's actually because it is a low trust environment in AI. People have tried lots of stuff. A lot of it hasn't worked. But if you've delivered stuff that does work, why do you need a salesforce? like you you you just it [02:10] sells itself. >> Well, you have to say, "Hey, don't talk to us." >> And then on that's that's that's that's on commercial in government >> at this point. It's really we don't like we don't we it's it's it's very hard to export simply because we have to train the people and we have limited bandwidth. >> I was going to say that's your limited bandwidth, the training once somebody takes on your your software. Well, in the government contest, every every country like obviously has the equivalent of a security clearance, right? So to use our to build to build ma project maven or maven into your architecture, you're going to need somebody with the highest level of clearance that also is technical and [02:52] most technical people unfortunately are not going after the highest level of clearance. So there are very very few people like that. So the that that resource is super scarce. Uh and then they have to be trained and that that takes that can take a while and and and then you also have to have like anything you really have to believe in this and really think it's important and you know not everybody fits in that category.