A physicist who once expected AI to remake the economy now argues the opposite. Connor Blake, writing in the Substack newsletter Boson Cutter, says the double-digit growth forecast by AI optimists will not arrive, because the world runs too slowly to absorb it. His essay is an argument, not a measurement, and it should be read as one.

Blake’s background explains his starting point. In theoretical physics, he writes, the limiting factor was always his own knowledge, so watching software and math problems collapse in hours felt staggering. In December 2025, when Claude Code paired with Opus 4.5 made agents feel truly useful, he feared a wipeout of white-collar jobs and bought “insurance” against it. He now says those predictions of rapid labor transformation were very wrong, and he promises a one-year retrospective in January.

His core tool is a simple model of a process made of sequential steps. Speed up one step to infinity and the whole process still moves only as fast as the slowest remaining step. Blake builds a hierarchy in which each step is faster than the last. Removing the worst bottleneck then yields only a constant-factor gain. Exponential gains require smashing a fresh bottleneck, again and again. He points out that this is Amdahl’s law, long used in computer science to study the limits of parallelism.

The practical test he offers is a widget factory whose defect rate suddenly doubles. Superintelligent agents could find the faulty step and even the contaminated precursor quickly. But fixing it still means ordering parts, shipping samples to labs, installing hardware and running the line to confirm the cure. Blake estimates a month for that cycle today and thinks about a week is the floor, even with perfect machine decisions around the clock.

Scaling, he argues, hits the same wall. Even in oil, a multitrillion dollar industry with every reason to automate, he thinks a 3x speedup would be hard. He cites two recent chokepoints: turbine blades made from single crystals, needed by gas power plants, and indium phosphide wafers used in optical links. Both appeared merely while building out computing. A true industrial boom would make niche magnets, bearings, sensors and alloys the next chokepoints.

Information is another limiter. Blake says AI researchers work where every run repeats perfectly and more experiments just mean more containers. Biology offers no such luxury. Experiments vary, informative results arrive slowly, and some answers, such as long-term drug side effects in people, are inherently expensive to obtain. Those costly bits, he argues, set the pace for everything around them.

His historical evidence is the strongest card. Electricity spread, communication costs fell from weeks by courier to nearly nothing, computers made once-impossible calculations cheap, and transport costs dropped by an order of magnitude. US agriculture went from 51 percent of the workforce in 1880 to 1.2 percent in 2022. Yet GDP per capita kept growing at about 2 percent a year for nearly 150 years. If those general-purpose technologies left the trend line flat, he asks what would make superintelligence different.

The counterargument is plain: this technology may differ in kind, because it targets the thinking that every bottleneck-removal effort depends on. Blake does not fully rebut that here, and he says follow-up pieces will take on it directly. His claim also rests on his own factory scenario and a stylized model rather than data on AI deployments.

For investors and operators, the stake is pricing. Valuations and capital plans that assume near-term growth far above the historical 2 percent are betting that intelligence is the scarce input everywhere. Blake’s wager is that shipping time, regulation and lab throughput are scarcer, so the sensible forecast for the next ninety days is steady adoption gains in software and little change in physical industries.

Adapted from “On the Brutality of Amdahl’s Law” by Connor Blake, published in Boson Cutter (bosoncutter.substack.com).