The case for moving now
Why deep-tech founders
can't wait on AI.
Most founders building in the physical world are watching the AI noise, not using it — stuck in the “let’s-see” zone. The gap isn’t the technology. It’s knowing what to apply, and where.
The world has already moved
While the frontier debates, it also ships. AI is already inside the hardest problems in deep tech:
- Drug discovery AlphaFold 3 (DeepMind & Isomorphic) predicts how proteins, DNA and drug molecules interact — ≥50% better than prior methods. Insilico Medicine went further: AI picked the target and designed the molecule for a lung-fibrosis drug that showed real efficacy in a Phase IIa human trial — the first AI-originated drug to work in patients.
- Materials Microsoft’s MatterGen designs brand-new stable crystals to order: specify the property, it generates the material. One design has already been synthesized in the lab.
- Fusion DeepMind + EPFL gave an AI direct control of a live tokamak’s magnets, adjusting them 10,000×/second to hold the plasma. AI controlling real fusion hardware.
- Weather DeepMind’s GraphCast beats the world’s best physics forecaster on ~90% of measures and makes a 10-day global forecast in under a minute (vs hours on a supercomputer).
- Chip design Synopsys’ DSO.ai has run 300+ commercial tape-outs; customers report ~3× design-productivity gains.
- Robotics Physical Intelligence’s π0 is one foundation model controlling many different robots across real tasks — atoms moved by agents.
- Autonomy Waymo: tens of millions of paid driverless miles, ~90% fewer injury crashes than humans — corroborated by peer-reviewed and insurance-claims studies.
- Aerospace NASA uses generative design to evolve spacecraft parts up to ~3× stronger and two-thirds lighter, in about a week.
India’s on the board too: SatSure runs ML on satellite imagery with ISRO’s commercial arm; Agnikul 3D-printed a single-piece rocket engine, cutting build time ~10 months → ~3 days. The frontier isn’t only in San Francisco.
These aren’t demos. They ship. The question isn’t whether AI matters to hard tech — it’s how fast you move.
The real unlock: your best people, back to their real work
Your scarcest people — senior engineers — spend most of their day on work they shouldn’t: reports, reviews, docs, tooling. AI removes the busywork around engineering, not the engineering. And the time it frees compounds — they use it to build more tools.
Proven up close
An electric-aircraft company: CFD post-processing 2–3 hr → 5 min; cross-discipline design review (days → minutes); one engineer got 4–5 hr/day back.
“What I was doing was only 5% of what I can do now.”— Senior CFD engineer
A construction-tech company: one daily site video → automatic status + report; drawings you can query; half a day/day of ops time recovered.
Prospect research automated (~2 hr/day/rep); five tools → one; monthly target raised 4 → 10.
“Why is it 10,000? Why can’t it be 15,000?”— CEO
Every team became an AI team; engineering headcount flat — 45 vs a planned 150 — output compounded.
Where AI bites in deep tech
- Compress simulation / analysis loops (CFD, SIL/MIL) — hours not days
- Specialist-perspective agents for cross-discipline review
- RAG over standards / regulatory corpora — instant cited answers, drop costly licenses
- Auto-documentation & test / cert evidence as work happens
- Engineering → investor / ops visibility layers
- Engineers building their own tools — the compounding loop
This is what the community is for.
A room of deep-tech founders figuring out exactly this — together.
Join the community