Jason Kelly on autonomous labs and the coming software model for science

Published 2026-03-24 - Runtime about 58 min - Watch on YouTube

Jason Kelly’s thesis is blunt: biotech has been bottlenecked less by ideas than by manual lab work. If reasoning models can run experiments through autonomous labs, science becomes a software-like loop of parallel hypotheses, shared data, and cheaper iteration, with implications for drugs, industrial biology, and national competitiveness.

What Matters

• Ginkgo began in 2008 but did not raise capital until 2014; Kelly says the early company bootstrapped on grants and services because it was “uninvestable.”
• His framing is split cleanly: design biology better with AI, then make lab work cheaper and faster; he has shifted toward the second half.
• In Ginkgo’s OpenAI project, a reasoning model plus robotic lab optimized cell-free protein synthesis, running 30,000 experiments and beating state of the art by 40% after six rounds.
• Kelly’s key claim: the breakthrough was not superhuman creativity, but logic plus throughput. The model could run experiments 24 hours a day and share raw results across 100 parallel hypotheses.
• He says science spending is badly allocated: less than 5% goes to reagents, while most of the cost is people, regulatory overhead, lab space, and duplicate equipment.
• His vision is daily data exchange among AI scientists, weekly notebook-style outputs, and a lab where failed results from one hypothesis immediately inform the others.
• Near-term biotech wins, in his view, are still drugs and consumer health: GLP-1s, better biomarkers, weekly blood testing, and longevity products that measure molecular change over time.
• Ginkgo’s own cloud lab now offers experiments as cheap as $39, signaling his bet that “formalized human curiosity” becomes accessible once the lab barrier drops.