Post

Business Insider confirms it: Amazon has warehouses of GPUs that can’t be installed, and is buying them so far in advance that a supply chain team said they “didn’t have a clear forecasted path” to installing them. www.businessinsider.com/amazon-220-b...

Ed Zitron@edzitron.com

Free newsletter: I estimate that ~50% of AI chip sales - $200bn to $300bn+ - are sitting in warehouses, with NVIDIA selling hundreds of billions of GPUs years in advance. Data centers take way longer to build than hyperscalers are leading us to believe. www.wheresyoured.at/wherere-all-the-ai-chips/

584 reposts2.2K likes

Please, I want the fire sale to come so I can buy a H100 for $50 and tinker with it. AI stuff is cool to play with and I’m not giving my data to the next Zuck to sell off when they go bust. Let the coders and computer nerd have they toys to home lab with!

They’re not really useful for anything other than AI work — they can’t output graphics and they’re worse at higher-precision math than cheaper alternatives because they went all in on fp4 and fp8 performance — so I wouldn’t count on a broader market impact

More likely scenario, IMHO, is that they get dirt cheap for hobbyists to put in home labs and maybe NVidia starts cranking out graphics cards again in an attempt to recoup costs on pre-purchased TSMC wafer starts. Maybe GPU prices get good again…?

The biggest problem is that they aren't usable for gaming because they have minimal raster units. Someone tried with a Hopper datacenter part and managed to get it to display but it performed worse than a iGPU from the mid 2010s

Most likely they would part them out. If this whole thing comes crashing down, there will be shops (and maybe NVIDIA themselves) that would be willing to chop up these boards for PARTS (RAM/GPU/etc) and create something that works. Really depends on how far they crash.

The line in the screenshot is the story: room for racks, no power to turn them on. Running models myself taught me the same thing at small scale. The chip is rarely the bottleneck. A GPU waiting on a pallet isn't capacity, it's inventory losing value every time a new generation ships.

... And by the time they have hardware to install them *in*, they're going to have been obsoleted by whatever new development in tech has come up in the interim.

That's what boggles my mind. Like even with 'Net360' or whatever crazy shit NVIDIA is offering: why are people still buying GPUs? Eventually that train's gonna end, right?

They don't want people getting GPUs to their own models and games. They need to make sure all compute goes to them. You will own nothing, because you will rent it from them.

An important piece of the story, I suspect. Ordinary people with unique ideas for developing local AI models / cheap replacements for over-priced or over-monetized technology (subscriptions, surveillance, forced dependence) are a threat to current predatory business strategies.

Business Insider also reporting that Amazon projects it would take delivery of $3 billion a month in GPUs “every month for the foreseeable future,” all while not having anywhere to put them, as part of an “early rack receipt” program. Extremely strange!

I’ll take the bet that they’re more concerned about keeping them out of customer hands than actually installing them. Oh you want you want your own AI models hosted in racks you own? Too bad. By the way you can pay us by token for that

2025 the RTX PRO 6000 Blackwell was 8k. Now it's 16k. This isnt difficult to figure out why, and it sure isnt AI demand.

We need to start stealing pallets of these fuckers in big, intricate heists, and spin up a cottage industry of jailbreaking/repackaging them to some state that makes them usable for the average joe.

I’m waiting for a really crazy revelation like the boxes actually being empty and Nvidia hasn’t actually shipped anything to these companies at all.

No shit this is how executives think, they don’t consider the cooling and the power and of course you need to buy fiber optic modules and cables and all that shit is back ordered for ever because of all this shit I enjoy my job most days though heh

everyone is obsessed with the silicon but nobody is talking about the power availability. you can have all the h100s in the world but if you can't plug them in they're just very expensive paperweights.

As Ed has noted, Nvidia probably forced them to do this by saying they'd be denied future GPU allocations if they didn't lock in purchases now and then Amazons own greed took over and did the rest

This is totally consistent with the idea that they're all spending/building so much because they *know* there's not room for all of them to win. It's not a technology race; it's a "force competitors to drop out" race. Hoarding resources is very much part of that strategy. Capitalism at its worst.

And they go obsolete about every 3 years. When the ndw model doubles previous ones. This is a modern tulip bulb scheme lololol these chips drop value faster than a lambo driven off the lot.

So this seams like it's creating a BIG house of cards. Functionally on paper the hyperscale's are ok as long as their being rewarded (in increasing share price) for increased CAPEX spending. They can use the capital to buy more GPUs and the market takes it as a GOOD sign.

Technically the GPUs don't ACTUALLY depreciate because: 1) they're not actually USED 2) they're ARE actually increasing in PRICE due to demand With the side benefit of not needing to spend money on electricity to power them. That said, their also not making any revenue. The moment they

plug them in they then have to depreciate those assets... so are the GPUs actually worth MORE sitting in a warehouse now? Are the hyperscale's ACTUALY depreciating the unused GPUs? Or are they marking them at "fair market value"? And what happens the moment their not rewarded for increased CAPEX?

ok so everything is expensive, they made all these junk gpu that cant be recycled, and there is no end in sight regardless so who knows when production on usable memory and such will ever be back to normal. cool cool depressing cool.

The gap between buying GPUs and putting them to work seems crucial here. Is there any public disclosure that separates installed capacity from equipment still waiting for power? That would help distinguish a temporary delay from a lasting drag on returns.

So fucking cool how every piece of tech is twice the price now so these seven companies can hold onto pallets of these things and do literally nothing with them. Capitalism!!!

American version of Soviet collapse where instead of warehouses full of endless weapons it’s warehouses full of computer chips being sold on the cheap to any sketchy buyer that shows up

From a "real" economist [ @JoeBrusuelas ]: "Today’s AI build-out is similar, which is why a bullish outlook—as opposed to what in my view is an incorrect dystopian outlook that has gained favor—needs to prevail." realeconomy.rsmus.com/market-minut...,

realeconomy.rsmus.comMarket Minute: This time is not differentWhenever a transformative technology is introduced, there will be malinvestment. Eventually, the bond market imposes discipline.

By the way, Microsoft seems to be hiring(Or advertising for positions) for their three data centers that they are building in Brazil, but I didn't find updates for something like eight months. No one seems to be talking about them.

I ask this question as an AI hater, but I don't really understand why this piece of it is bad? If you believed you needed to build these data centers for real and sustainable reasons, wouldn't it be reasonable to lock down GPUs even tho the buildings to put them in were taking longer than expected?

Well for one thing it's making everything with ram in it much more expensive. For context the ddr4 (which is obsolete) in my gaming pc that I paid just over 200 dollars for less than two years ago now costs over 1000 dollars. The same amount of ddr5 is over 4000 dollars currently.

Oh yes, I understand why this is bad as part of the excesses of this whole corrupt bubble; it's enragingly wasteful on so, so many levels. But a lot of Ed's work shows underlying weakness of the companies that should undermine the hype, & I'm trying to see if this aspect fits in to that angle

Obviously misleading investors is bad, but I don't understand why they would do that about this particular information, which doesn't seem that damning from the perspective of someone who believes the hype. The companies stocking up could be seen as evidence of their long term commitment

By the time they’re ready to be used they’ll be outclassed by something vastly more energy efficient. Fucking dumbasses.