Toke use exploded because people like Google released updates that turned my phone's power button into an AI button that I have literally never used. Every web search is now an AI bot and we have even seen "Linux use" go from 2.5% to 8% and most of that increase is just AI bots skimming the Internet.
Outside of scientific and research purposes, AI products are almost universally hated. Kids are only using it to get out of home work. There are the AI girlfriends and who didn't see that one coming.
But here is something that everyone seems to forget that preaches about AI token numbers. More advanced models use more tokens. If you take away the AI button on phones, copilot and web searches the same people are paying for AI. The cost of compute hasn't changed and bloating up AI models that have basically platued in performance. Applied AI hasn't grown, it's the same people doing the same work only the models use more tokens now than they did a year ago.
The real trend in AI is buying a couple of racks and running free open source models locally. It's not the datacenter ready hardware that's making a splash from Nvidia, it's the racks with 8GPUS in it that companies like SuperMicro are selling
