The Model-Agnostic AI Platform Betting That No Single Lab Will Win
In a Nutshell
Stan Huleu founded Dust as a model-agnostic AI platform because tying products and tokens to a single frontier lab creates dangerous dependencies—comparable to building a factory that only runs on one energy provider. The company bet early on a horizontal platform strategy despite market pressure to verticalize, and continues to resist abandoning model flexibility even as usage patterns shift dynamically. Frontier labs' high margins (70-80%) and vertical market entries are compressing pricing power for pure-play AI companies, making credit-based pricing and defensible network effects essential for survival.
These notes were generated by AI and may contain inaccuracies.
Model agnosticism is a key differentiator that frontier labs cannot replicate. Purchasing both product and tokens from the same provider creates dependency risks. The analogy compares this to building a factory where machinery only works with one energy provider—if that provider fails, the entire operation collapses. This was demonstrated when European buyers of Russian gas faced supply disruptions, while nuclear energy from France remained accessible.
Stan Huleu joined Stripe early and spent five years there during an acquisition year. He later joined OpenAI when research could be conducted as an engineer, spending three years working on large language models and mathematics. He is not a trained researcher but gained significant experience during this period.
Stan left OpenAI to return to building products rather than conducting research. Research involves scratching surfaces for weeks or months, experiencing brief highs when finding something interesting that last only five seconds, then continuing the process. He wanted to focus on product development rather than the research cycle.
Dust was founded with the motivation that large language models were ready to fundamentally change how people work. In 2022 and early 2023, applying LLMs to the workplace was considered a niche. The company operates on a rapidly evolving technological substrate, unlike the previous 20 years when JavaScript and PostgreSQL provided a stable foundation.
Stan predicts that in two to three years, work will feel fundamentally different from current perceptions. Just as great-grandparents would view current desk-based work as not real work compared to physical labor, future observers will view today's work as not real work. The technology will continue evolving without plateauing.
The founders were correct that work would be disrupted by AI technology. This disruption is only now becoming visible at a three-year timescale, with developers experiencing radical changes since November 2025. The technology is proving extremely disruptive and restructuring for how work is conducted across all functions.
The founders were overly optimistic about an earlier plateau in AI capabilities. Even if the technology had plateaued months ago, decades of deployment and innovation would remain for adoption across all work contexts. However, the plateau is not occurring, and work is being actively redefined.
OpenAI and Anthropic are viewed as both beneficial and challenging competitors. They educate markets efficiently, driving convergence toward horizontal productivity suites for human-agent interactions. This resembles the three-body problem concept—a powerful force that covers everything while remaining barely visible.
Dust pursued a horizontal platform strategy early when verticalization was the dominant advice. Verticalized products offer simpler go-to-market strategies with clear target audiences, while horizontal platforms face the challenge of people asking what specific use cases exist. However, the market is converging toward less depth and more coverage in product integration.
Model agnosticism remains a core commitment. While usage may concentrate on one provider at any given time, this shifts dynamically based on which model performs best for specific work. The company will not abandon model agnosticism. This approach avoids dependency risks similar to being locked into a single energy provider.
Frontier labs absorb significant funding market energy, as VCs prefer investing $1 billion in one lab rather than $100 million across ten different companies. Dust's seed and Series A were relatively easy to raise, but Series B proved more challenging. Building from France during a period when it was less fashionable from an American perspective contributed to fundraising difficulties.
Despite backgrounds from OpenAI and Stripe as second-time founders, Stan and co-founder Gabriel chose reasonable valuations for seed and Series A rounds. This decision was influenced by observing friends get burned during 2020-2021 by raising too much money. The strategy follows the mantra "No GPU before PMF"—they raised $5 million at seed when others were raising $100-200 million for training, recognizing they were building at the product layer.
Everything would have been easier from a company-building standpoint in the United States. However, the founders wanted to build something back in France due to national preference and sovereignty considerations. When companies succeed, location becomes irrelevant—VCs don't care about founder locations for successful companies like Lovable or Spotify.
When labs like Anthropic enter vertical markets (such as legal tech with Legora), verticalized products lose their scaffolding advantage as models improve. The value of verticalized products will depend on creating defensible network effects rather than just delivering intelligence. Without network effects, defensibility becomes challenging as intelligence commoditizes.
Dust transitioned from seat-based to credit-based pricing due to exploding usage from longer agent loops and increased token consumption. Flat pricing becomes unsustainable as margins compress under current conditions of improving models and increasing usage. Credit-based pricing is necessary to maintain margins when users want to maximize AI utilization.
Precedents exist where companies successfully sell products on top of infrastructure that captures margins while competing with zero-priced offerings from infrastructure providers. Open source catching up to frontier models creates pressure on lab margins. Analysis of latency between Claude Code and GLM 5.2 shows roughly equivalent performance, yet GLM 5.2 costs approximately 9x less on platforms like Fireworks, suggesting frontier labs maintain 70-80% margins.
The most important factor is finding the thing you want to fix—the vision that drives you. Building a company involves daily difficulties with highs and many lows. Without a compelling reason to wake up each morning, the work is unsustainable. It's acceptable to reinvent this vision while building, but having that driving objective is essential despite the challenges involved.
Keep Y Combinator in your library
Save the videos and channels worth coming back to, and find them again in one place.
