The Model-Agnostic AI Platform Betting That No Single Lab Will Win

Stanislas Polu argues that building a model-agnostic platform is crucial because it decouples the product from any single AI provider. He compares buying tokens and product from the same lab to buying machines that only work with one energy provider, which would be risky if that provider were unreliable. This flexibility allows Dust to switch between models as they improve, rather than being locked into one lab's ecosystem.

Stan left OpenAI after three years because he wanted to return to building products rather than doing research. He describes research as a cycle of intense highs that last only five seconds before needing to scratch the surface again. He gave up stock options worth more than Dust's entire valuation at Series B, but does not regret leaving because he prefers the tangible impact of product development.

Dust was founded in early 2023 with the thesis that LLMs would fundamentally change how people work. Initially, applying LLMs to the workplace was considered a niche, but Stan was convinced it would be disruptive. He notes that as of late 2025, work has already radically changed for developers, and he expects other functions to follow, making work feel unrecognizable to past generations.

Stan predicts that within two to three years, work will be so transformed by AI that it will no longer resemble what we consider work today. He draws a parallel to how our great-grandparents would view modern desk jobs as 'not work.' He is optimistic that this shift will make work more efficient and less laborious, though he admits precise predictions are impossible at this stage.

Stanislas Polu acknowledges that building Dust in France rather than the US adds friction, especially for fundraising and hiring. He notes that when a company is performing well, VCs do not care about location, citing examples like Lovable and Spotify. The decision was driven by a desire to build something back in France for reasons of national preference and sovereignty, which he views as an additional constraint but a worthwhile trade-off for the greater good.

Polu argues that verticalized AI products had an advantage when models were weaker, requiring scaffolding to be useful. As intelligence becomes commoditized, the only defensible moat is a network effect—value derived from multiple users interacting on the same platform. Without that, a verticalized product is vulnerable because the underlying intelligence is easily replicated, echoing advice from the 2010 SaaS era.

Dust originally used seat-based flat pricing to encourage usage, but over the past six months, agent loops and model token consumption have exploded, compressing margins. Polu states that credit-based pricing is now unavoidable for AI products because users will max out any flat allowance. He warns that margins will keep compressing under current conditions unless pricing adapts to usage.

Polu points out that AI labs like Anthropic and OpenAI capture massive margins at the token level, as seen in SpaceX's S1 filing. He believes that open-source models catching up will eventually put downward pressure on those margins. Meanwhile, product-layer companies can still capture margin by building well-differentiated, well-crafted products, citing precedents like Box and Dropbox that sold on top of infrastructure with zero-margin competition.