BlockBeats news, November 22: François Chollet, founder of Keras, stated that if the AI infrastructure buildout is not yet over and compute demand has not yet peaked, then the recent decline may be a rare buying opportunity. GPU spot prices are still rising, compute is difficult to lock in, HBM and DRAM capacity through 2029 is basically fully booked, and chip and memory supply is expected to remain tight into the 2030s. Training compute for frontier models roughly doubles every 5 months, and the pace of model capability improvement is more than twice as fast as before the emergence of reasoning models. Margins in the AI inference business can reach 70% to 85%, demand is still growing rapidly, and each new generation of models brings new application scenarios.
Musk replied that it is very difficult to power, cool, and keep such large-scale compute equipment running stably, and the AI industry may face a severe funding shortage in 2027. He believes that if, across the industry, the time from paying compute costs to equipment being able to generate stable revenue is delayed by several months in aggregate, working capital alone could require hundreds of billions of dollars in additional investment, and he asked where that money would come from.
