Former OpenAI CPO Weil seeks $150M at $750M valuation for science-data startup

Weil, a veteran product executive who previously worked at Meta and Twitter, left OpenAI earlier this year. He is now pitching his new project. His stated vision echoes a comment he made when OpenAI launched Prism, a workspace that embeds frontier models into scientists' workflows. He described 2026 as the year of AI and science, just as 2025 was about AI and software engineering.

Prism is widely seen as OpenAI's first major productization of its AI-for-science strategy. Users compared it to Overleaf, LaTeX and Codex, focusing on features such as formatting, citation management and LaTeX compilation speed. But its presence in actual discovery, including experiment design, hypothesis testing and data insight, is weaker. The limits may not lie in the product itself but in the organization. OpenAI's resources remain concentrated on general-purpose model iterations, leaving scientific data infrastructure a lower priority.

That structural bottleneck also appears to have shaped the career moves of several top AI figures. Jeff Dean, who left Google after 27 years, co-founded Discovery Loop with Sanjay Ghemawat, Oriol Vinyals and Quoc Le. The company aims to automate the whole scientific process with AI. In a conversation at Startup School 2026, Dean said general-purpose large models often fail in hard science fields such as materials science, chip design and biocomputing. AlphaFold succeeded, he argued, not because of compute scale but because a very small team of core researchers with deep obsession and top skills worked on a specific domain.

Discovery Loop sits at the model layer of the AI-for-science stack. Another startup, Periodic Labs, founded by former OpenAI researcher Liam Fedus and former Google DeepMind researcher Ekin Dogus Cubuk, operates at the application layer. It builds an 'AI scientist plus automated laboratory' model, with robots doing experiments and generating training data. Its investors include a16z, Accel, Jeff Bezos and Eric Schmidt.

Weil's new company would occupy the data layer. It would not touch experiments or train models, but instead collect and structure scientific data scattered across labs, Excel sheets and private databases. The idea is that such data, though richer and higher quality than internet text, is fragmented and unusable for model training. A well-integrated data layer could become a toll booth for the entire AI-for-science supply chain.

The $750 million valuation, according to the report, reflects a heating-up in the AI-for-science track, talent premium and the potential of the data layer. Weil's OpenAI track record and the scientist user network he built through Prism give him rare access to labs. Scientific data, unlike web data, is not public, not standardized and heavily protected by intellectual property and competition. But it is also highly domain-specific: data logic in biology differs from materials, and particle physics formats are incompatible with climate models. Whether a cross-domain 'AWS for scientific data' is feasible remains an open question.

Skeptics say without breakthroughs at the model layer, data alone is not knowledge; without experimental loops at the application layer, data cannot be verified or iterated. Scientists could also choose to hand data directly to automated labs or train their own models, bypassing a middleman. The integration difficulty may be underestimated by capital markets, and the $750 million valuation could become the first bubble in the AI-for-science field to pop. No one has an answer yet, the report says.

The report notes that AI4S is now moving from tool narratives to infrastructure construction. Weil's startup, Discovery Loop, Periodic Labs and more unnamed players will define the rules of this track. Leiphone says it will continue to follow the story.