Articles like this plus observations that aggregating over hw types could create a separate/independent credit cycle than the dynamics of the Nvidia driven one (c.f. @LukePiette) resemble the early days of the online ads market If you were a big publisher (e.g. Conde Nast) you ran your own exchange and effectively supported credit to ad buyers who came direct to you. But then the remnant/secondary/non-publisher market grew and you found ad networks and demand side providers who aggregated remnant/secondary ad space selling people bundles that replicated the click through rate / false discovery rate of the primary publisher venues. Eventually, the market turned into a barbell/bimodal structure with FB/Google/TikTok being one end of purchasing spectrum and a number of other public ad networks taking the rest. Given the recent DoJ frontrunning cases against Amazon and Google, however, combined with the much harder, collateralized credit needed for AI, you probably will have a different market structure for AI. But there are definitely lessons from ads and crypto — especially in terms of who takes most of the value and who doesn’t (hint: how hard their technical solution was to implement doesn’t correspond with long term value capture)