Invisible timer

Algorithms and training data are commoditizing fast (MIT-licensed MatterGen, open-source MACE/NequIP/Allegro). The parts AI cannot touch — physical synthesis at scale, multi-year qualification regimes, regulatory paperwork — are immovable and dominate time-to-revenue.

Key data

  • 80%+ of AI-predicted candidates show crystallographic disorder (Fritz Haber Dec 2025)
  • A-Lab's "36 successes" turned out to be known phases or impure mixtures
  • 9 of 10 random GNoME entries already in ICSD; 18,138 contain radioactive Pm/Ac/Pa
  • Cumulative new-EV-battery chemistry = 18-36 months, $millions
  • Citrine + HRL Al 7A77 → NASA = ~2 years (cleanest precedent)

Matter Loop strategy

Defensible position — exclusive throughput agreements with specific contract synthesis labs and certifying partners. AMS-QQ-A Atlas — codify the graph as a proprietary asset.

Evidence

"No functionality has been demonstrated for the 384,870 compositions… they cannot yet be regarded as materials."
"The problem is not coming up with new materials — it's coming up with new materials that have desired properties and can be cheaply synthesized."
"Cumulative new-EV-battery-chemistry qualification: AS9100 + NADCAP + AMS-QQ-A + UN 38.3 + IEC 62660 = 18-36 months, $millions."

All insights →