Foundation models for materials need proprietary experimental data → data is the moat.

Fedus thesis (Periodic) drives fundraising. BASF empirical proof.

• BASF doesn't buy from startups; partners with academia.
• MACE-MP-0, NequIP, Allegro, MatterGen (MIT + training data), PLaID++ — open-source.
• Citrine 13 yrs / $76M / no raise since 2023 — canary.

Bifurcation

When it breaks — bifurcation: generic foundation-model layer becomes a free utility; value migrates to (a) workflow/UX layer that wins reluctant-scientist adoption, (b) integration layer operating on customer data in customer environment.