And do these not all have established business models that are ridiculously profitable, and huge cash reserves?
How the heck are THEY vulnerable, and the "bare" frontier labs like Anthropic and OpenAI with no preexisting business models or profits, but equally ridiculous borrowing are supposedly not?
Google, Meta, Amazon, MS have huge cash-producing businesses they can just fall back on. Of course the stocks will get hammered in the short term, but they are still sitting on cash volcanos.
Pure AI plays are what are most at risk. In the bad case, all the CoreWeave type companies have their assets acquired on the cheap by the existing big players. In the armageddon case, companies like Anthropic and OpenAI also end up in the discount bin.
Hyperscalers are FAANG.
No, by Hyperscalers (HS) I was referring to the pure-play companies doing the massive build outs of mega-scale data center farms for FAANG and frontier labs (FL). FL / FAANG own some of their own mega-scale data centers but they don't want their business to become too invested in real-estate and GPU depreciation. While it might lower costs to invest BIG capital in vertical integration (owning the whole value chain), the margins in that part of chain are too low to support FAANG stock prices.
So they are using a mix. Some of their own but mostly leveraging OPM (other people's money) through Hyperscalers for pure-play AI. FAANG view data centers necessary for their own core product needs (like AWS or Google Search / Gemini) in a different financial bucket than pure play AI scaling. The fact they are able to resell their excess DC capacity for pure AI to OAI/Ant so profitably right now is an opportunistic windfall, not a long-term plan. In fact, it's actually off-plan in a sense because their financial plan was not to invest tons of capital to sell AI compute on "margin-plus". It's just that the "margin-plus" margin is incredibly frothy right now. Their strategists don't think that will still be true in five+ years.
That's why their long-term mode is leaning to doing a mix and largely relying of external Hyperscalers for non-core AI capacity. Their strategy is generally, "Own our own AI compute for minimal baseline needs for our branded products (we sell to the final end-user)." Then they want to use Hyperscalers for non-core and excess capacity because HS are borrowing billions on debt markets using FAANG contracts to secure those loans. I explain why and how we can be in both 'rapid growth' and 'increasing bubble danger' at the same time in more detail here: https://news.ycombinator.com/item?id=50430395.
NVIDIA will take a huge hit and we will probably end up with anti circular financing laws.
Frontier labs are losing money and are most likely to go bankrupt and be absorded by bigger companies.
Hyperscalers will take a huge beating but they are ultimately profitable and they already control the frontier labs. So after a hard crash they will be fine. A lot of their debt is actually offloaded onto pension funds, insurance companies, etc. Those risk bankruptcy.
I think we're essentially saying nearly the same thing and agree it eventually leads to a similar outcome. The difference is in part saying it in different ways and in part addressing the flow and feed backs at different points. Ultimately there are three key players in an ecosystem flow that supplies end-user demand:
[NVidia (NV)] <-> [HyperScalers (HS)] <-> [Frontier Labs (FL)]
TSMC is upstream of NV but they're diversified enough to be sheltered, and end user demand is the ultimate gating factor. Most people agree that in the long-term (10+ yr) AI compute services will be a very big viable business.
> Frontier labs are losing money and are most likely to go bankrupt and be absorded by bigger companies.
Anything's possible but I don't think that's likely. It really depends on who is left 'holding the bag' on committed contracts and callable loans. While all the Frontier Labs have essentially taken what nets to equity-secured loans from NV, AFAICT it's not collateralized in way that NV (nor HS) can force FLs into bankruptcy. Creditors with capital were willing to take FL equity for cash, making them investors, not creditors. Post-IPO that will change. FL sales will continue to steadily grow but not at rates required to make their equity worth what investors thought it would be worth in the time frame they expected/needed.
> Hyperscalers ... already control the frontier labs.
Through what legal or financial mechanisms is this control? If end-user demand continues to grow super-linearly, FLs will need more compute from HS than they get. In that case, HS continue to have pricing power due. If future end-user demand growth rates taper from super-linear to just strong linear, FLs will need less compute than they've projected. HS pricing power will erode.
Hyperscalers ... are ultimately profitable
HS are collateralized to their bankers, GPU and power suppliers in ways that CAN force bankruptcy. When demand softens, prices will fall. HS have bought, and committed to buy, more GPUs, power and real estate sooner than they are going to need it from people who CAN force them into bankruptcy. The HS have secured their loans by pledging their GPUs, committed power contracts and real estate/physical plants as collateral.
When cash flow is high and the growth rate of cash is increasing, lenders will loan $2 to $4 for every dollar of 'sellable asset' the loan is secured by. The HS are already extending their GPU depreciation time frames to SIX years which inflates the forward book value of their GPUs - so they can borrow more at cheaper rate. When the growth rate of cash flow starts to taper from super-linear to just very good, lenders start to demand more collateral per dollar (or capital costs go up). Both are already happening.
Google, Amazon owns Anthropic shares. Microsoft owns OpenAI shares. Fairly sure some of the others also have Anthropic and OpenAI shares. And I mean large chunks, at least 5% up to 30%, I think.
I agree with the rest but I think hyperscalers will just get drastic haircuts.
> hyperscalers will just get drastic haircuts.
I think everyone may get drastic haircuts (though I'm still hoping for a 'soft unwind' type slow-mo crash, I don't think the odds are good). Hyperscalers are what probably breaks first and worst because more of their money is from creditors vs investors, combined with, more of their creditors are financial speculators / banks who have no vested interest in propping up the AI market (like an NV, FL do). They are only loyal to their cash.
Yes. And valuation. I think the leading frontier labs will be fine in the long-run and NVidia will be fine. Some of their investors who bought at too high of a valuation, expecting to flip it in too short a time frame will lose a chunk of their money when they have to sell for less.
The hyper-scalers in the middle are the most vulnerable. They borrowed billions at increasingly high rates to lock up GPU and power capacity. They are 'ground zero' of a lot of the 'circular money' deals. Big banks who financed hyper-scalers initially, expecting they could package that debt up and sell it on to diversify, are already struggling to find buyers for the debt, causing its value to drop, which then forces the banks to cover with more collateral, and that's how the snowball starts down the mountain.