Bob McGrew says superintelligence will stall unless it rebuilds the physical world. He made the case on No Priors: AI, Machine Learning, Tech, & Startups with iPhone math. McGrew said 75% of parts change each year while only 25% carry over.
Researchers are not the bottleneck.
McGrew, CEO and co-founder of Arda and former chief research officer at OpenAI, rejected the hard takeoff story where automating researcher time unlocks runaway speed. He said the limit when he led research was GPUs on hand, not headcount, and he framed progress as moving that stack from 35% of optimal toward 90%, McGrew said, where the last gains buy little. He recalled pitching papers to Ilya and hearing, "So what are the results," because even mathematical research stays experimental until the run finishes.
A car program shows the cost of inflexibility. McGrew said a highly automated car line can cost $200 million and then run the same model for years, which works for autos and fails for phones. Arda had secured $232 million at a $1.5 billion valuation as of October 2026, reported by Bioengineer.org, or roughly the price of one such car line to chase products that turn over constantly.
The numbers do not show Arda can turn a design into a working automated line in days at customer cost. That caution fits a wider worry about physical limits, also raised in Cowen Says AI Growth Will Hit Bottlenecks. One working high-mix line, repriced after changeovers, would change the read.
Can cheap intelligence make $200 million tooling cheap enough to chase 75% churn?




