There has never been a bigger gap between what a small group of people seem to know and what the market is doing with its money. From up close, a lot of that money looks like it's chasing a delusion.
To be clear, the delusion has nothing to do with AI itself. I'm extremely AI pilled. The models are improving fast, I think the pace picks up from here, and if anything the market as a whole isn't AI pilled enough.
The delusion is about moats.
There's a recipe going around. Find a hot corner of the market, wrap a foundation model in a thin layer of UI, fill the deck with the right buzzwords, and announce a breakthrough. These companies have no network effects, no real defensibility, and no answer for what happens when the model underneath them gets better. And they're raising hundreds of millions of dollars, sometimes off demos that don't even work.
I'm watching all of this from the inside. I'm a founder myself, building in exactly the kind of market this money chases, so the easy way to dismiss this essay is as a founder annoyed that other people are raising more than he can. But that gets things backwards. Rounds like these are not hard to get right now. If your corner of the market is hot enough, the money is sitting there waiting for a deck. I know because I feel the pull myself. The temptation to raise big, announce loud, and figure out the product later is real, and resisting it is a choice I have to keep making. Raising was never the hard part. Building something that still deserves to exist in three years is.
None of this means VCs are stupid. Their business is buying lottery tickets. A fund needs a couple of 1000x outcomes to survive, so it has to back ideas that sound crazy, and most of those ideas are supposed to die. That part of the model is fine. What worries me is the difference between a moonshot and a company parked in the path of a bulldozer. A moonshot dies if the future it bet on never arrives, but these companies die when it does.
I wrote about this a year and a half ago. The moats available to software companies have changed dramatically in the last three years. Foundation models keep absorbing capabilities, and every capability they absorb deletes a category of product. What's left lives in the business model itself: network effects, legally sensitive data, consequences in the physical world. Those were the constraints I used when deciding what to build. My advice to anyone starting a company right now: read Richard Sutton's Bitter Lesson, fully absorb it, make sure you thoroughly understand it, then read it again. It keeps winning. Proprietary datasets matter much less than they did a year ago, and custom models will eventually lose to general ones, probably within the year. If your whole company is a thin layer on top of someone else's model, all of that progress is headed straight for you.
So where does a gap this large come from? I think a lot of it comes down to how few people understand how deep learning actually works, how these models get trained, and what a transformer is doing underneath. If you aren't in the weeds, you end up running about a year behind the people who are, and the last few years have made that lag painfully easy to see.
Memory is a good example. Anyone who was actually keeping up a few years ago, reading papers, training models, comparing notes with people doing the same, knew that demand for memory was about to spike dramatically. An August 2024 paper on scaling test time compute laid it out clearly, and DeepSeek reinforced it that December. Once models started getting smarter by thinking longer at inference time, the memory math was obvious to anyone who understood the architecture, and yet the market needed almost a year to digest it. Compute went the same way. Only in the last few months has the market seemed to grasp the real scale of demand and the scaling laws driving it, even though scaling laws have been well understood in the field for six years now.
This is what I mean when I say the market isn't AI pilled enough. None of this was hidden. It has all been sitting on arXiv for anyone willing to do the reading, and most of the market simply hasn't done it. Which leads me to a prediction. The venture firms with the best returns over the next decade or two will have AI researchers as actual partners in the firm, people who deeply understand the fundamentals and can reason about what foundation models will absorb next. Too many investors and founders today have a shallow picture of what's happening under the hood, and a shallow picture makes it very hard to make good strategic decisions over a long time horizon.
The other part of the gap is cultural. Somewhere along the way, raising money turned into a status symbol, especially among younger founders. It used to be something you did reluctantly because building the thing took capital you didn't have. Now companies raise huge rounds so they can tell everyone they raised huge rounds, and the capital itself is almost beside the point.
I got a taste of this recently. I was evaluating a tool for my own stack and got on what was supposed to be a demo call. Instead the salesperson took me through a pitch deck. Half of it was about what a prodigy their founder is. The other half was about the board, who was on it, why they're all so impressive. I kept waiting for the actual product to show up. It never did.
I don't care who sits on your board, how much you raised, or where your founder dropped out of school. I'm a customer. Show me the product.
Somewhere we started treating dilution like an achievement instead of a cost, when bragging about a big round is really bragging about how much of your company you had to sell. The numbers worth bragging about are the boring ones: churn, retention, how annoyed your customers would be if you disappeared. And I don't get a pass on any of this. I plan to hold this essay against myself before anyone else does. When I do raise, I want it to be because the product needs the fuel, and I want the round to be the least interesting thing about the company. If I ever brag about one before I can brag about the product, someone please send me this essay.
The companies that survive the next few years will be the ones that picked problems foundation models can't simply absorb, and then solved them well enough that customers stayed. Everything else is a press release.