The Beginning of the End for Big AI
Local models, broken pricing, and the slow shift from centralised frontier labs to controlled systems at the edge
Welcome to Drift Signal. I’m Nicolas, Head of Research at Vsquared Ventures, and this is my personal newsletter on markets and macro.
First, some recent work elsewhere that may be of interest: last Sunday, Marieke Flament and I published a new edition of our newsletter Currency of Power dedicated to Tether. We look at how it has evolved from a stablecoin issuer into a global financial institution sitting at the centre of dollar liquidity, and why most of traditional finance is still misreading what it has become. Read the full piece here: Tether Is Not a Stablecoin Company.
Now, in today’s edition, I react to Alex Karp, CEO of Palantir, spilling the beans on frontier model companies and why I think it signals an imminent turn toward local models — and, correspondingly, the bursting of the Big AI bubble.
My argument is that the economics of frontier models, the rise of strong open source alternatives, and the shift toward data and control at the edge are starting to pull the system in a very different direction.
1/ This is it. Mark my words. This is the signal that it’s time to run for the exit, because the ‘Big AI’ bubble is about to burst. Alex Karp says businesses depend on a handful of frontier labs, pay for tokens that create little value, and risk handing over the data that makes them valuable. Here’s what he suggests everyone should do instead: run good enough models on your own data while keeping control of the weights, the data, and the compute. What we see is customers unhappy about paying too much for things they do not understand, while suppliers are still losing billions. At some point this has to correct itself, and listening to Karp, I get the sense that the correction is about to happen.
2/ So what is Alex Karp essentially saying? Again, from my perspective, he says everyone has become too dependent on a few frontier labs, which together constitute ‘Big AI’ (my words), and that this has gone too far. Being dependent means three things here:
You use Big AI’s products, and they send you hefty bills for ‘tokens’, which, frankly, nobody understands. Somehow, tokens have become the unit for billing software, with usage based pricing replacing traditional SaaS models. Yet nobody knows how many tokens they’re about to use, why one task burns ten times more than another, or why they suddenly run out. But apparently this is now a perfectly sensible way to buy software 🤔
You upload your data, not really knowing what Big AI will do with it, potentially giving away your “alpha”, to use Karp’s word. The very thing that makes your business valuable ends up sitting on somebody else’s platform, under terms that are constantly evolving. You’re told your data is safe, your IP is protected, and nothing funny is going on. Trust us.
You depend on whatever those companies decide. If they change their prices, retire the model your product depends on, change the rules overnight, refuse to work for the Pentagon, or their chatbot wakes up one morning and decides it’s MechaHitler, there’s not much you can do about it. Is that really considered a foundation for enterprise software?
3/ What is the remedy to this? Local models! It does not matter whether your model is frontier, because in most use cases you do not need frontier, you just need something that works and is good enough. The key point is that value sits in the data, not the model itself, and some organisations have so much valuable data they would rather keep it for themselves than ship it off to a frontier model.
Suddenly, the clever bit is no longer buying access to the biggest model on the planet, but putting a capable model next to your own data and keeping both under your control. And if you are a responsible and accountable business leader, you probably want to master the costs, the data, the compute, and the handling of all of that, rather than discover next month’s AI bill after the fact.
4/ Local models will first happen at the periphery: individuals, startups, researchers, and tinkerers using Chinese open source models, Hugging Face, or whatever comes next. That’s where people experiment, break things, and discover what actually works. That is how most computing shifts happen: at the edge first, then slowly into the core.
By contrast, Big AI has always been an anomaly. Instead of bottom-up experimentation, we got a small number of companies setting the rules, the prices, and the access. Maybe that was inevitable for the first wave, but I expect the second wave to look rather different. It will be about putting capable models in the hands of people who own their data, control their infrastructure, and would rather not rent intelligence by the token forever.
5/ A local model does not mean you are doing everything yourself, and that is exactly what Alex Karp is selling. Palantir’s “ontology” makes it possible to run whichever model makes sense on your own data, while you remain in control. You can swap models, keep your data where it belongs, and decide how much compute you need. The intelligence becomes part of your infrastructure, rather than somebody else’s product.
You still depend on a service layer, but it looks more like systems integrators such as Accenture or Infosys. They build, integrate and operate systems for you, but they do not own the systems themselves. The supplier can be replaced, even if it is costly. That is different from being locked into a black box that sets prices and rules from the outside. At least with this model, you can see what is in the box, even if you still need someone to help you run it.
6/ None of this is happening because Alex Karp woke up one morning and had an epiphany. If anything, he is merely recognising that something has been bubbling beneath the surface for quite some time:
On one side, customers are paying eye watering sums for AI products whose pricing bears little relation to the value they receive. At the same time, Chinese open source models have become astonishingly good, making “good enough” look increasingly attractive. If your local model solves 95% of the problem at a fraction of the cost, the economics start to speak for themselves
On the other side, the frontier labs are still setting fire to cash on an industrial scale. They are spending tens of billions building ever larger models and ever bigger data centres, all while the cost of chips, power and infrastructure keeps rising. That is a perfectly reasonable strategy if someone eventually earns an attractive return on all that capital. It becomes rather less convincing if customers increasingly decide that they do not need the biggest model after all.
This tension cannot hold forever, and we have seen this film before. As Sameer Singh wrote here and here, the 1970s taught us that ever larger muscle cars, supersonic airliners and giant supertankers all eventually collided with economic reality. Bigger is impressive, until it is simply too expensive. I don’t see why AI would be different.
7/ What happens when local models start to replace a world where a small number of companies try to run everything through Anthropic and OpenAI? We will not need all of those data centres, and that is good news because they are extremely expensive to build and run. Indeed, there is little reason to fund a race to AGI if most real work can be done with smaller models close to the data. At that point, two things can happen:
One path looks a bit like Batillus-class supertankers and Port d’Antifer. Massive capital gets locked into infrastructure built on the assumption that demand will grow in a straight line and that ever larger models will always be worth paying for. Then the economics change. Demand shifts to smaller, local or distributed models, and the giant build-out starts to look misplaced. Not useless overnight, but slowly harder to justify, harder to fill, and harder to fund at the same return. The result is stranded capacity.
The other path is more dynamic. Compute becomes something closer to a market in its own right. Instead of a handful of companies owning the full stack and renting it back out in fixed ways, you get a more fluid system where people run their own models in their own cloud instances, and buy compute when and where they need it. Capacity is not pre-committed years in advance to a single provider’s roadmap. It is allocated more like a spot market, with price signals doing more of the work that planning used to do.
In that world, data centres do not disappear. But they stop being monuments to a single bet on ever larger models, and start looking more like shared infrastructure that feeds a wider range of uses. Whether we end up closer to Port d’Antifer or closer to a real compute market depends on how quickly the shift to local models actually takes hold.
8/ What will be left? Some big customers and entire sectors will still pay for frontier models, think pharma, drug discovery, clinical research, and other areas where getting the answer slightly more right is worth a great deal of money. Financial services will likely sit in the same camp for a subset of use cases, especially where small differences in prediction or timing can move large sums. Defence and intelligence will also keep a strong appetite, because in those domains you do not optimise for cost per token, you optimise for outcomes.
But even there, it is not really “market price” in any clean sense. Pharma is already a world of subsidies, insurance systems, and indirect payers. Defence is a world of procurement cycles and budget allocations that have little to do with marginal cost. Frontier AI in those settings becomes another layer of specialist infrastructure, paid for because the alternative is worse, not because it is efficient in any normal commercial sense.
So yes, frontier models survive. They just stop being the centre of gravity.
9/ About that: if one of the frontier model companies wants to survive, I do not think the right path is to try and secure ownership by the US government — which, by the way, is another sign that things are about to go in a very particular direction. Instead, I think the only stable path is to become more like Google: find a business model that generates so much profit that everything else becomes optional, and only then start playing with the really expensive toys.
Google acquired DeepMind, and DeepMind went on to build AlphaFold, which solved a long-standing problem in biology by predicting protein structures at scale, including most of the known human proteome. That work later formed part of the basis for the 2024 Nobel Prize in Chemistry awarded to Demis Hassabis, John Jumper, and David Baker.
Sebastian Mallaby has also written about this arc in The Infinity Machine, which treats DeepMind less as a product company and more as a long-running experiment in what happens when you give serious capital to serious people and let them push at the edge of what computers can do.
In sum, the serious science and the expensive experiments sit on top of a business that already prints money. Without that base, the rest does not survive long enough to matter.
10/ Whilst OpenAI seems to be in a very tricky situation (I wrote about it), Anthropic could become something closer to a next-generation Google. Claude Code can become the cash engine, in the way AdWords once was for Google: a way to fund everything else, while the real work happens in the background on improving the model itself.
All in all, the companies that survive will not be the ones trying to act like standalone frontier labs that set the rules for everyone else, but the ones that turn their models into infrastructure sitting inside much larger business systems, where the economics are stable enough to support long cycles of research and build-out.
That is the shift. From a small number of companies calling all the shots, to a more fragmented world where models are embedded, localised, and controlled closer to the edge. Big AI is not disappearing so much as being absorbed into a more distributed system. That is good news not only for Palantir, but for builders more broadly. Now the real work moves to the edge.
Sign up to Drift Signal if you don’t want to miss the next issues 🤗
From Munich, Germany 🇩🇪
Nicolas
This is a great post. Thank you.
sewage in, sewage out.
the law of entropy.