As far as I can tell, because that is not happening.

The suggestion seems to have come from the claim that to maintain a priority queue, you have to be buying consistently. That might mean buying more than you need to ensure supply when you need it. The assumption that the excess goes into a warehouse seems to have come from someone's fevered imagination.

If there are GPUs built and not yet installed, and people who want to be using them now, money can be made getting the supply to find the demand. Most large scale companies know how to do that sort of thing. If someone has a data center and no GPUs arriving for six months, paying a few million to trade the GPUs that will exist in six months for the ones that exist now gets them an additional six months of operation, while ensuring the original buyer will have GPUs when they need them, plus a few million.

▲ torginus 20 hours ago | parent [-]

Yeah. Don't know anything about data center GPUs, but it's common knowledge you can get like 80% of the performance of a gaming GPU at half the power or even less. That's basically how Nvidia's mobile GPUs are. So if you had a warehouse full of GPUs, you could just go ahead and install them, run them at low power, and still get like 2/3rds of the performance.

▲ bigbadfeline 13 hours ago | parent [-]

> So if you had a warehouse full of GPUs, you could just go ahead and install them, run them at low power, and still get like 2/3rds of the performance.

This doesn't add up. First, it's still hoarding, I'd call it lukewarm GPUs as opposed to dark GPUs. Second, nobody would do that because it needs more space and maintenance, more racks, along with the added wear, tear and power supply losses that come with them. How do you people come up with these genius ideas, justifying the documented hardware hoarding [1]?

Here's another fun case, where some GPU hoarders simply forgot about a truckload of GPU's and didn't notice it was missing for full two weeks after it was stolen. It doesn't look like they need GPU's for "urgent installation" [2].

[1] https://www.datacenterdynamics.com/en/news/microsoft-has-ai-...

[2] https://www.tomshardware.com/pc-components/gpus/six-pallets-...

▲ schiffern 5 hours ago | parent [-]

  >documented hardware hoarding

If you read the actual quote, all they say is that Microsoft isn't bottlenecked by GPUs. The "hardware hoarding" angle is just editorialization by the writer of the article.

  >“The biggest issue we are now having is not a compute glut, but it’s power – it’s sort of the ability to get the builds done fast enough close to power,” Nadella said. “So, if you can’t do that, you may actually have a bunch of chips sitting in inventory that I can’t plug in. In fact, that is my problem today. It’s not a supply issue of chips; it’s actually the fact that I don’t have warm shells to plug into.”

  >His comments echo those made by Microsoft’s CFO Amy Hood on the company’s Q1 2026 earnings call last week, where she told analysts that access to compute hardware has not been a constraint for the hyperscaler

Nothing about GPUs sitting on a shelf somewhere.

	▲	torginus 2 hours ago \| parent [-]
		Which is the point I was making. One example I managed to find is: https://www.techpowerup.com/301649/nvidia-geforce-rtx-4090-w... They get 95% of performance in a game with 60% TDP. I'm not sure how this translates to datacenters, but I suspect You can get at least 75% of the perf from a datacenter GPU running at half power, so it makes sense to install them anyway, and run all your stuff at lower power, as you'd still get 1.5x the perf.