Startup School Paris

Stanislas Polu on Why No Single AI Lab Will Win

Stanislas Polu· Co-founder of Dust; former OpenAI engineer at Dust
·~23 min·English·Y Combinator
LLMAgentsAI CompanyBusiness Strategy
TL;DR

Dust co-founder Stanislas Polu argues the durable AI company is the one that never bets on a single frontier lab — staying model-agnostic, horizontal, and defended by network effects rather than scaffolding.

01The Core Bet

Model-Agnostic Is the One Thing Labs Can't Copy

A lab that sells you both the product and the tokens can never be model-agnostic, and that, Polu argues, is the one structural edge it cannot copy.

We're also uh model agnostic which I think is a is a very interesting aspect of that and that's something that the labs will not be able to do.

— Stanislas Polu, Startup School Paris
Key Insight
The tell is that Dust's usage still concentrates on one provider at a time — so agnosticism isn't load-balancing, it's preserving the right to switch the instant a different lab becomes best. What he's really buying is optionality, and a lab that sells its own tokens structurally can't hold it.
02Origin Story

The 5-Second High That Made Him Leave OpenAI

Research, Polu says, hands you a breakthrough high that lasts about five seconds before you're back to scratching the surface, so he left OpenAI to build a product instead.

you find something interesting gets you like super high super high but it lasts for 5 seconds literally 5 seconds and after 5 seconds you're like oh yeah that was a you keep scratching again.

— Stanislas Polu, Startup School Paris
Key Insight
The stock math makes the decision look irrational — he says the OpenAI options he gave up were, at one point, worth more than Dust's entire company, though he notes that is no longer true. That is what makes it revealing: he optimized for the shape of the daily reward, not the size of the eventual exit.
03Future of Work

In Two to Three Years, Today's Work Will Look Like 'Not Work'

Just as our great-grandparents' physical labor makes today's desk work look like sitting and chilling, Polu bets that in two-to-three years today's work will look like barely working at all.

if they were looking at us today they would be like you're not walking here you're just sitting in a chair looking at a screen all day chilling around that's not work

— Stanislas Polu, Startup School Paris
Key Insight
The framing does the persuading: by making 'what counts as work' a moving line rather than a fixed thing, Polu turns agent disruption into inevitable continuity rather than rupture — the rhetorical move that makes a scary transition sound like just the next step.
04Right Thesis, Wrong Timing

The Plateau He Bet On Still Hasn't Come

Polu was right that AI would upend how we work but wrong to expect an early plateau, and even if models had frozen months ago, he says there were still decades of deployment left to do.

even if that technology had plateaued like a couple months ago we would still have like like decades of deployment uh and and innovation to do for it to be adopted in every pieces of places where we do work but uh but that that plateau is not happening at all

— Stanislas Polu, Startup School Paris
Key Insight
What he got right and wrong splits along a clean line: the direction (work gets reinvented) he nailed in 2022; the timing (when the curve would flatten) he missed. For a founder that's the survivable error — being early on direction is recoverable, being wrong on direction is not.
05Platform Strategy

Everyone Said Verticalize; Dust Went Horizontal

While the market chanted 'verticalize,' Dust built a horizontal platform, and now every product is converging on the same shape: a productivity suite for human-agent interaction.

it feels like every product is converging towards the same thing which is some sort of uh new uh uh productivity suites uh like Microsoft has been doing or Google Drive has been doing but for human agent interactions

— Stanislas Polu, Startup School Paris
Key Insight
The convergence he describes is a trap he's aware of: if every product becomes the same horizontal suite, coverage stops being a differentiator and the fight relocates to sensibility — collaboration, multiplayer AI, model-agnosticism. He isn't claiming horizontal wins; he's claiming that's where the next fight is fought.
06The Moat That Survives

As Models Get Good, Defensibility Falls Back to Network Effects

Verticalized products won by wrapping weak models in scaffolding, but as intelligence commoditizes, Polu says defensibility falls back to the 2010-era answer: network effects.

we're going back to the advices that we used to receive in 2010 almost which is network effects. Find a network effects. Find a thing that is defensible

— Stanislas Polu, Startup School Paris
Key Insight
The uncomfortable implication for most AI startups: the glue code that made them useful in 2023 is a depreciating asset, not a moat. Polu is quietly dating the entire 'AI wrapper' category and telling founders to find defensibility that survives, in his words, perfect intelligence coming in.
07Follow the Margins

Latency Betrays a 9x Gap in What Labs Charge

Because latency is a proxy for a model's true cost, Polu compares a frontier model to an equally-fast open model and finds roughly a 9x price gap, implying the labs run 70 to 80 percent margins.

the difference is like almost 9x. So that means that there is very good chance that those labs are margin kind of creating margin of something like 70 or 80% uh serving us those frontier models

— Stanislas Polu, Startup School Paris
Key Insight
The move worth stealing is methodological: he can't see the labs' cost sheets, so he uses latency — fixed by model size and hardware — as a public proxy for cost, then reads the price gap as margin. It's how an outsider reverse-engineers a black-box P&L, and it's why he bets open weights will eventually compress it.
08The Human Close

The Only Fuel That Survives the Glass-Eating

Building a company is 'shitty as hell' with a few highs and many lows, Polu says, so the only thing that sustains a founder is a vision they genuinely want to fix.

Find the the thing you want to fix the vision you have because building company is shitty on a daily basis like shitty as hell.

— Stanislas Polu, Startup School Paris
Key Insight
It's the same optimization he made leaving OpenAI: pick for the daily experience, not the outcome. A vision you'd wake up for is what makes the median day — which is bad — survivable. The exit doesn't sustain you; the reason does.