Does that 100% include the gear that gets stolen in transit?

'Stolen', meaning 'delivered to wrong address'?

I mean, the wind can start blowing the other direction. And they can go overboard on the deals. But in the past few years, yes Nvidia has predictably sold out.

It's obviously impossible to tell without more insight, but the large AI labs are very much out compute and could use more of it, or they wouldn't be renting DCs from their competitors.

The neoclouds ensure there will be consistent demand of Nvidia GPUs even after everyone big moves to their own custom ASICs

Out of compute or out of chips? I swear I heard somebody saying Amazon had a whole datacenter worth of GPUs not even pluugged in sitting in storage because they were pressured into to purchasing them early by Nvidia and datacenter contruction was significantly delayed.

> had a whole datacenter worth of GPUs not even pluugged in

It could be these chips [1] which they are trying to offload to someone else.

[1] https://shattered.io/amazon-8-billion-nvidia-chip-leaseback-...

But that also just shows that the demand for nVidia’s products exceeds supply. The reason that nVidia was able to pressure people with take-it-or-leave-it deals (if that’s what did happen) is precisely because everyone involved expected (and correctly, as it turns out) that there would be plenty of other customers ready to take all the places in the queue.

Running out of compute isn’t the same as “can buy more GPUs and use them”, else AWS and MSFT wouldn’t have billions of dollars of GPUs in warehouses that they have no datacenters in which to rack them.

GPU supply isn’t the constraint; datacenter space is. And that means Nvidia might not normally be able to sell all that they’d like even with seeming high demand.

I have customers begging, pleading for more GPUs. There is still strong demand.

Yeah now that they have engineered artificial scarcity.

How can we even talk about this without talking about price?

For AI data center chips, they pretty much are constrained by how much they can produce (TSMC is fully booked and RAM is scarce / expensive).

They are constrained by the chips they can power, at least in US