AI frontier slowdown could give second tier leg up

That should insulate so-called hyperscalers like Amazon AMZN.O, Alphabet GOOGL.O and Meta META.O, which make up the bulk of capital spending to some degree. Yet current AI winners like $5.3 trillion Nvidia NVDA.O, whose graphics processing units are vital for training models, and South Korea's SK Hynix 000660.KS, which supplies high-bandwidth memory for such chips, look worse off.

Intel INTC.O, for example, which specialises in another type of semiconductor called central processing units, may be a major beneficiary. Data centres catered to training advanced models generally require 1 CPU for every 8 GPUs; boss Lip Bu Tan has flagged that for iterative tasks carried out by agents, the ratio is one to one. Nvidia challengers that focus on inference, like Cerebras Systems, also stand to gain, as well as Samsung Electronics 005930.KS, whose conventional memory products are more suited for CPUs.