AI.

AI. My two cents on the macro.
For the last half year I’ve been pretty deep in the тусовка, sometimes even around the people who decide stuff and build the industry — including funds, and big model makers. So, a few observations.
tldr: the cycle is wrapping up. Better not build startups right now.

What we’ve got:
1/ infra. Here it’s simple: the ones printing money are building the most expensive place where that money will work. Overall the hardware re-leasing model is super low-margin. But when you’ve got ads behind you (where margins are insane), you can invest. Capacity is wildly short, honestly. Even finding a rack is hard (like, getting a b200 node for a year).

2/ most venture money is poured into the big labs. The logic is simple. Training models is very expensive, like very expensive. Mistakes, accordingly, are also very expensive. We pour money where the chance of a mistake is minimal. None of the big labs do frontier research as their main focus. They scale what already works. Top up RL, data, and so on. Plus every lab is trying to get as close to the user as possible (Anthropic is going prosumer, like Microsoft once did — to me that’s the best comparison).

3/ researcher comp can hit $10–50M a year, because not many people have seen how to train a 1–10T model. That knowledge costs a lot. Because mistakes cost a lot — see llama4. Hence VCs hope maybe they’ll make something useful, but also they’re the only ones whose chance of making something is above zero. Hence $100M seed rounds. And yeah, if anything you can sell them back to the places that make models (since hardly anyone can do it).

4/ given the size of the checks, you get a situation where an average startup needs to show insane outcomes, because why give $10M and get back $20M when the numbers over there are 10x higher. Moreover, AI startups have this cute property: most companies have very low revenue retention. Meaning the typical cycle is roughly: a customer takes all AI services to try (all competitors as a bundle) and picks 1 or even zero, deciding to build their own. Your revenue in posts grows, but everyone understands that in a year you’re fucked. This is a real situation. Notice how lovable and co went quiet. Yeah, because the growth changed and it’s not enough anymore. And investors initially gave you a crazy valuation.

4/ and here the only solution is brewing. Grow revenue on paper ultra-fast and try to unload previous-round investors onto new ones — or better, onto retail — i.e., IPO. And that’s a dangerous game. And for those raising at $2–10–50B valuations there isn’t much choice. And at those stages it’s already harder to sell so everyone’s happy.

5/ given that big lab revenue is heavily subsidized, an IPO will unload investors, but the market will look at it all and say “weeeeell not that great a business,” which will create a cooldown. And where do they pull money from, right — the riskiest segment, i.e. early-stage venture. And the odds that consolidation kicks off only went up. The big will get even bigger, and the small won’t have a chance (see pts 3–4).

So there’s this meme going around the Valley that Ilya Sutskever’s SSI is a hedge fund. Because he knows exactly what’s going on on the ground, and he’ll definitely be able to make money on all of this. Such is life.