Three AI Rounds Closed on One Day. None of Them Was a Frontier Lab.

On September 15, three AI companies closed funding rounds worth $414 million between them. Not one of them was a frontier model lab.

Factory took $200 million at a $5 billion valuation, roughly triple its previous mark, for coding agents that run without a human in the loop. CADDi raised $114 million at a $1.2 billion valuation for software that reads drawings and supply chain data for manufacturers, out of Tokyo and Chicago. Delos Data, founded by former Intel engineers, raised $100 million to build networking chips that move data between mismatched AI data centres.

That same week, the frontier was arguing about whether to stop. Three lab CEOs called for pacing development. OpenAI pushed its IPO to 2027, with Sam Altman telling Fortune that listing this year would be ill advised given safety concerns. Japan's Nikkei closed down 518 points as investors sold AI and semiconductor names.

Markets are treating these as one story. They are not.

## A pause has a narrower blast radius than the selloff implies

Ask what a slowdown would actually bind. It binds pretraining runs at a handful of labs with the compute to do them. Yesterday we argued that nobody currently has the standing to enforce even that.

What it plainly does not bind is a coding agent company, a Japanese manufacturing platform, or a chip startup selling into data centres that are already under construction.

None of those three businesses needs the next frontier model to exist in order to have a product. They need last year's model to be cheap and reliable. That is a different bet, and in a slowdown it is the better one.

## Check which layer got sold

This is the uncomfortable read of the Nikkei drop. Investors sold semiconductor and AI names on slowdown news. That is coherent if you believe AI value accrues to compute and to the labs. It reads worse if value is migrating toward the application layer, which is exactly where September 15's money went.

A genuine pause at the frontier would be good for applications, not bad. Stable model targets, falling inference prices, and no need to re-architect your product every six months are the conditions application companies have been asking for. The firms a pause actually damages are the ones whose valuation is a bet on the next capability jump.

Three deals on one day is an anecdote, not a trend line, and this week is noisy for obvious reasons. Take it as a question to test rather than a conclusion. But the question is sharp: if the slowdown is real, does your AI exposure sit above or below the frontier?

## What to do with this

**When you see an "AI selloff", check which layer got sold.** An index falling on slowdown news tells you about sentiment toward compute and labs. It tells you very little about companies selling applied software to manufacturers and engineering teams. Those got funded the same week the index fell. Our deal tracker breaks rounds down by type if you want to check the split yourself rather than trust the headline.

**Ask one question of any AI company you are evaluating: does this need the next model, or last year's model at a lower price?** Companies in the second category get cheaper inputs and a stable target in a slowdown. Companies in the first category are holding a call option on capability that a pause would write down. Right now the market is pricing both the same way, and that is where the mispricing lives.