Article · AI translation
IT Home reported on November 22 that François Chollet, founder of the open-source deep learning project Keras, posted on X that either you believe the buildout of AI infrastructure is over, AI is merely a short-lived bubble, and demand for computing power has peaked; or this is a rare buying opportunity.
François said GPU spot prices are still rising, capacity is hard to secure, HBM and DRAM production capacity through 2029 is essentially fully booked, and chip and memory supplies are expected to remain tight into the 2030s. Computing power used to train frontier models roughly doubles every 5 months, and model capabilities are improving more than twice as fast as before reasoning models emerged. Profit margins in AI inference can reach 70% to 85%, demand is still growing rapidly, and each new generation of models brings new applications.
IT Home noted that Musk replied to François, saying it is extremely difficult to power, cool, and keep computing equipment of this scale running reliably.
(The AI industry) could face a severe funding shortfall next year. If the industry-wide interval between paying for computing power and equipment generating revenue reliably is delayed by several months in total, hundreds of billions of dollars in additional working capital alone may be needed. Where will that money come from?