On hardware costs, if it produces a large, sustained amount of demand, and if there are fixed costs (e.g. R&D) that can be shared between hardware used for it and other things, it may substantially reduce hardware prices in the long run for other users.

Suppose, to take an example, that there is demand for, oh, completely pulling a number out of the air, 4 times the amount of high bandwidth memory for AI that there is for 3D video cards and video game consoles. That's on a sustained basis, not just our initial AI buildout. There is going to be some amount of fixed costs that have to be done at Micron and Samsung and the like to figure out how to design the product and optimize production.

That's going to mean that AI users likely pay something like 80% of the fixed costs for HBM, which may very well lower costs for other users of HBM.

In late 2025 and 2026 there is a huge surge in demand for hardware. There's a shortage of hardware, and factories don't get built out overnight. So prices skyrocket, pricing out many users to the point where demand at the new price point matches the available supply. But as production capacity increases, that will also ease.

I do get that it's frustrating if someone wants to build a system right now.

But scale matters a lot, and this may enable a lot more scale.

The reason I can have a cheap Linux desktop at home isn't because there are masses of people buying Linux desktops, but because there are huge numbers of businesses out there buying Windows desktops and many of the fixed hardware development costs are shared. If those businesses running Windows desktops suddenly disappeared tomorrow, I probably couldn't afford my home Linux desktop, because suddenly I'd need to be paying a lot more of the fixed costs.

What did AL do this time‽