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

Stan Houlé, former Stripe and OpenAI engineer, explains his transition from AI research to building Dust, a horizontal platform focused on applying large language models to workplace productivity. He left OpenAI despite substantial equity because he preferred product-focused work over the iterative cycle of pure research. Dust's core strategy centers on being model-agnostic, allowing users to leverage intelligence from multiple providers rather than being locked into a single lab's models—a critical differentiation point against OpenAI and Anthropic.

Houlé addresses the challenge of building alongside frontier labs, noting that while the labs' market education benefits the ecosystem, Dust differentiates through collaboration features, multiplayer AI capabilities, and intentional model agnosticism. He uses an energy provider analogy: just as buying machines from an energy provider whose plugs only work with that provider's energy creates dangerous lock-in, requiring customers to buy tokens from the same lab providing intelligence is problematic long-term.

On work disruption, Houlé reflects that while he correctly predicted AI would fundamentally change work, he underestimated the pace. After three years of relatively stable work patterns, significant changes only began in late 2024 and early 2025. Regarding funding, he acknowledges that frontier labs absorb substantial VC capital, but expects IPOs to remove them from venture competition, improving market conditions for startups.

Dust's fundraising philosophy emphasizes raising conservative amounts at reasonable valuations—a "No GPU before PMF" mantra that shaped their €5M seed despite industry trends toward $100-200M AI training raises. He balances this with admiration for Mistral's aggressive growth. On geographic strategy, building from France rather than Silicon Valley added friction but felt important for sovereignty and vision alignment.

The discussion covers how vertical AI products face defensibility challenges as models commoditize. Houlé argues that network effects become the only sustainable moat when intelligence becomes commoditized, advising founders to focus on defensible competitive advantages rather than pure intelligence delivery. On pricing, he explains the shift from seat-based to credit-based pricing as token consumption exploded with longer agent loops and improved models, noting that margins compress without usage-based pricing. He points to frontier labs' estimated 70-80% margins compared to 9x pricing differences for equivalent latency models, suggesting open-source competition could recalibrate markets.

Houlé concludes with advice for entrepreneurs: find the core vision that sustains you through the inevitable grinding difficulty of building, as daily company-building can be brutal without that deeper purpose.