Post

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/

wheresyoured.athttps://www.wheresyoured.at/wherere-all-the-ai-chips/If you liked this piece, you should subscribe to my premium newsletter, and you can subscribe on the following links: $70 a year, $18 a quarter, or $7 a month. In return you get a weekly premium…

236 reposts939 likes

Getting a bulk discount for pallets of GPUs, warehousing em for a few years to drive up demand and then selling them to increasingly desperate gamers? Now that's just sound business strategy if you ask me. This is legal financial advice.

This sounds *exactly* like WorldCom warehousing Nortel gear, to fool Wall Street analysts, to maintain the illusion of growth. It worked until it didn't...

See also "inventory loading" aka "garage qualification" in MLMs. The idea that these giant firms are acting like LuLaRoe or AdvoCare is ... both disturbing and at least a little humorous.

In May, I asked where all the data centers were, with no gigawatt-scale facilities actually completed, and a few months later would report that Microsoft only had 2.2m GPUs installed - or around 1.993GW of capacity. Bloomberg confirmed two weeks ago. www.wheresyoured.at/wherere-all-...

Microsoft has around 2GW of AI compute, yet it's announced on each of the last three quarters that it's brought a gigawatt of capacity online. It's very clear that it's misleading the public about its capacity - and data centers are massively delayed. www.wheresyoured.at/wherere-all-...

Since the beginning of 2022, Microsoft has spent $265bn+ on capital expenditures, yet has only brought around $50bn to $60bn of GPUs online. It's unclear why it keeps saying it has gigawatts coming online, but it isn't reflective of its AI buildout. www.wheresyoured.at/wherere-all-...

Microsoft should, as one of the single best-capitalized and most-experienced data center developers, be thriving. Yet based on information from sources and my analysis, it likely has between $50bn and $100bn in warehoused GPUs and AI-related IT equipment www.wheresyoured.at/wherere-all-...

In which case their reported depreciation becomes a straight-up fraud; if you depreciate an asset over six years when it regularly burns out in two...

Not necessarily. If the chips can still be used in lower use tasks, like inference, even if those chips can only work at a small fraction of their original throughput, they'll still be on the legal depreciation schedule.

Tulip bulbs, on the other hand, are very easy to plant and require only water and sunlight to grow. Would you consider investing in my tulip business, Ed?

You look too young to remember the fiber boom and bust of the late 90s, but this feels like that. 90% of the overbuild with fiber sat DARK as internet traffic growth initially failed to match capacity, bandwidth prices collapsed and everyone went belly up.

Ed talks about the dark fiber comparison all the time and AI infra buildout isn't that for a number of reasons, principle among which is that there is no forthcoming growth ramp, there's no internet-like future for gigantic ultra expensive LLMs after this bubble pops

That fiber eventually became useful and used at full capacity. The fiber does not become obsolete. A GPU is laughably out of date after 3 or 4 years. So if they're 3 years late building the data center they need newer GPUs then.

Good to have someone agrees with me on that. I've been saying this since this craziness began and every 6 mos Nvidia has a new version lol.

Very difficult because these ones are directly integrated into massive server racks, they were built from the start to run "AI" not for normal computing.

Every insurer and broker feasting on it is incentivized to keep the good times rolling, so I imagine they won’t talk, but I wonder if they would be the ones who have the goods on whether these data center projects they’re feasting on are actually starting/moving

I think about this in the context of all those loans being collateralized by the chips; those don’t get done if the assets aren’t insured. Are those sitting in the builders risk placements? With NVIDIA? Really wonder if London is noticing like $100B of limits stacking on a random warehouse in Asia

So, I’m just skimming AI news, but I wonder if the current call for the feds to intervene with a “slowdown” is motivated by the fear of each company collapsing due to this extreme over-extension? Or are we talking about two distinct things?

I guess it's a bit of a technicality given that the GPUs are presumably paid for & on the books, but - are these actually physically warehoused by Microsoft & co, or are these future production/delivery contracts?

This week’s earning calls are a pivotal moment - is the deluge of AI slop really worth the mountain of debt? And why should investors keep pumping in more money to inflate the bubble?

I wonder how much hardware from nVidia and others are just paper. How much of this can be no actual physical equipment being delivered, just balance sheet pumping?

yes. Nvidia releases a new and better chip every year and compute farms have to weigh buying new chips vs using their old chips which provide less computing performance per watt.

What are they going to do with all of these single-purpose chips that will be obsolete by the time they have a building to put them in? They have to take the write-down, obvs, but what actually happens to the physical stuff? Does it get dumped on poor countries ? Who bears the eco cost of this waste

This is part of why NVIDIA is pushing the idea of compute as an asset class -- you can extend the depreciation schedules of the GPUs even if they're not cutting edge performance, you can borrow against the compute you own even if it's not deployed, and you can keep buying as much as they can make.