I don’t think the effects on the hardware market of open weight models triumphing has been reflected upon enough. It’s not clear that it will be less impactful if every enterprise decides to build for predominantly on-premise inference. In fact, before the question is ultimately decided, we’re probably heading towards a few years where hobbyists, enterprises and ‘hyper scalers’ are all competing for certain parts in common. Ram booked through 2027 sounds about right.

Companies are avoid risks, and at this stage, sometimes, I feel that all cloud providers are just buying RAM that they would have bought anyway. Now it's ensure that no company will be willing to pay x digits just for cards. One of my clients is stuck in "we are doing things on premise but we are too cheap to spend a few 100k in cards but we don't want to go on clouds.