The objection to thermodynamic computing held for ten years on operational grounds. The chips would be unstable, the noise would have to be calibrated per device, the toolchain rewritten from a substrate the field had spent five decades suppressing. Underneath the engineering anxiety sat a metaphysical premise about computation owing the engineer determinism. Every datasheet from TSMC's last three nodes already concedes what physics taught in the first lecture. The transistor at sub-3nm is a probabilistic device. The silicon lives inside thermal noise no fab can fully suppress. Every frontier-model GPU burns kilowatts to whip electrons into a binary discipline they no longer want to keep.

Extropic's Z1 ships from the other side of that wall. 250,000 p-bits sampling from shaped distributions on the same silicon footprint the industry currently drinks the watts to refuse. The thermal fluctuations the GPU spent its lifetime damping are the computation. Frontier inference was already sampling from probability distributions, and those distributions are the thing the underlying hardware emulated in fixed-point math at five orders of magnitude in extra joules.

Verdon spent the years between Google Quantum AI and the first Z1 writing the framing in essays and writing the chip in parallel. Both arrived on the same calendar. The 2027 inference bill the hyperscalers are signing is the cost of running probabilistic workloads on hardware built to refuse probability. The substrate stopped fighting itself.