But in this case I very strongly disagree with his conclusion regarding Palantir. He seems to suggest that Palantir is a possible significant player in the first wave of AI because they are embedding themselves in the enterprise and doing the legwork of integrating data to make it useful and that the resulting benefits will outweigh the enormous costs.

Here's the problem(s). First of all, enterprise data is extraordinarily complicated as it evolves over many years and has the fingerprints of the prevailing technologies of the day all over it. It is nearly impossible to rip and replace or integrate those systems in flight because there is too much at stake and each system is endlessly complicated. That's to say nothing of the processes themselves which tend to evolve alongside the technology and are often inefficient because of the reliance on the dated technology. So Palantir has two challenges- overhauling data and processes.

Second, if the solution is integration work, what is special about Palantir? They hire the best analysts? C'mon. If it's a matter of talent that will become democratized with enough cash. There are loads of SI's that do this today. Are they just bad at it? If there isn't something special about Palantir then they have no edge.

Most damning though is that if you can integrate the data you don't really need AI because there are all sorts of tools available that would work amazingly well with the proper data. The challenge is always integration and making use of disparate data. There are hundreds of thousands of people who's job is solely being the glue between systems. Robotics could solve this but they are very inflexible, expensive and limited in functionality. Put differently, in this scenario AI is not the secret sauce it's the integrated data.

There is always this impulse when something new comes along to assume the existing way is dumb and those who use it or conceived it were ill-informed and this new thing or team will fix it all very easily. But once you get into it you realize there are an infinite number of considerations that make it very difficult to make any meaningful change.

I think the idea of agents is a great one but making them useful consistently will be very difficult and the question becomes does the benefit outweigh the enormous cost of building and training them. I suspect that with this enormous AI buildout you will end up with a lot of unused capacity that is later repurposed for an unintended use that nobody predicted. That is largely how bubbles eventually play out.