During the internet boom, lot of fiber capacity was build, and 98% was not used. It is funded heavily by debt. OTOH, AI Capex is out of operating cash and frontier model providers are running out of capacity, i.e., capacity is fully utilized, in fact, Anthropic claude is so famous that routinely developers are complaining that Anthropic hits token limit quickly…
This doesn’t mean there won’t be a bust.. or every stock going down along with it, not just the AI stocks, or other sectors like Utilities which benefited from the AI boom…
The Blue Owl Joint Venture: Meta partnered with funds managed by Blue Owl Capital in a joint venture targeting up to $27 billion in development costs. Through this deal, Wall Street investors sold billions in bonds to finance the construction. Meta will not hold the bulk of this debt on its own balance sheet; instead, the company will lease the facility and pay rent to Blue Owl.
Meta owns just 20% of the joint venture. Blue Owl-managed funds own 80%.
Through a special purpose vehicle (SPV) arranged by Morgan Stanley, the project issued $27 billion in A±rated debt and $2.5 billion in equity, anchored by PIMCO ($18 B) and BlackRock ($3 B).
Meta keeps operational control but not the debt. It leases back the finished campus under a long-term operating lease, converting what would have been CapEx into predictable OpEx.
That’s true, they are not “borrowing” in the same way that the dot-coms did, but they are borrowing just the same. They just happen to be borrowing from their future cash flows, and indebiting themselves for terms sometimes said to be longer than the useful life of the centers themselves.
The first point might have been true very early on, but AI data center build out is now mostly being funded by debt. The hyperscalers use a lot of financial technologies to keep that debt off the balance sheets of the parent companies, but it’s debt all the same.
As for capacity being utilized, I think there the concern is that the demand might go away once the frontier model providers start charging enough for their services to cover those massive capex costs. It’s easy to have an oversubscribed product if you’re charging pennies on the dollar; it gets somewhat harder to do that once you start raising prices/imposing ads sufficient to cover your expenses. Once these massive capex investments start getting passed through to end users in the form of higher token prices, will the user flow still be there at these levels?
A DC useful life is anywhere between 30 to 40 years. It is not DC funding itself I am worried but lot of funny assumptions were going around GPU/CPU depreciation math. That’s the first one that will blow up because of depreciation accounting and debt.
There is a difference. If you look at their off balance sheet debt vs their yearly FCF, they can pay those debt within a year or two using their FCF. Also, the first 20% or so risk is taken by players like “Blue Owl”, and I am not expecting the DC’s will not be used, rather they will be used, may not be at full capacity, wiping the equity players, but can service debt and meets META’s of the worlds need.
Lot of folks questioned how can $AMZN make any money on the delivery.. here we are. Separately, while the price goes up, the cost could come down. The economic model may very well survive. Today, at least, for anthropic, they can increase the price, but they are constrained in their growth because of the availability of compute.
We are already here… $MSFT has canceled its claude license because of high usage/ cost, and $UBER has burned their entire budget in 4 months, and the CTO is warning the company in an internal mail… there are other areas like GitHub moving away from flat pricing to usage based pricing (because of excessive use due to AI)…
AI is evolving much more rapidly, we are getting new data points everyday… It is difficult to stay tuned to the news… may be it is about time I need an agent
Few random thoughts…
If any of you worked with wordstar, the first word processing, it ran on less than 64k RAM that includes OS too, today you see a simple word file runs into multiple MB’s and word will take anywhere between 300~400M. Already we are seeing every new model is using more and more compute. So compute usage will only grow. To that extend DC’s may be utilized.
The ever increasing compute usage will require newer cooling, network technologies. Will any of these make DC’s redundant? My guess is they may require costly retrofit, but will not obsolete… assuming you get breakthrough technology that is vastly different from existing solutions.
OTOH, where the real dangers are, in my view, in new age hyperscalers finances. They are already making aggressive depreciation, useful life assumptions, some are taking the first 20% hit on these DC buildouts. So they are exposed to double whammy of faster obsoletion and either saddled with the cost of retrofit, or mild demand destruction…
Interestingly many of them also carry nose-bleeding valuation. Right now shorting them is getting in front of a freight train, when things turn around, it will unravel very fast, hopefully I have the courage, and ability to short some of these names. It is always easier to ride the wave then short it… but some are attracted to heavy punishing trades.
Some additional thoughts, I am bearish on OpenAI, Spacex, Anthropic IPO’s. Initially because that is lot of Trillions coming to market, and it will trigger selling to fund those purchase, but I am also now added, that once these companies become public, there are going to lots of disclosure, and that is going to be “interesting”. Right now, investing public is generally aware that these AI companies are running huge loss, but to see $10, $20, $30, $50 B operating loss will be very dramatic, I assume.
That’s why, while I am bullish on AI as a technology, but I think there is going to be lots of economic, financial dislocation before we get to the promise land.
This makes me wonder, particularly about “scaling”. This is not scaling like Microsoft’s “write the software once, sell it a million times”, or Facebook’s scaling of adding users at almost zero incremental cost; this is “adding capacity” and have to pay for capacity as it is used.
Clearly there is some hypothetical level where “enough users” pay “enough dollars” to make it profitable, but each additional user or each additional use) does not come free, which is what I think of when we talk of “scaling.”
I have to build this huge plant first and get to a point - both in hardware and in user demand - where those axes cross into profitability, but this is not one of those cases where once that happens the line goes vertical and it’s all cash in my pocket (unless I have overbuilt for the eventual demand by 1000% or something).
[I’m not saing you can’t have a terrific business making things and selling it to people for a higher price: Apple does it with phones and IBM does it with services, but in all those cases the business built with measured demand and appropriate expenditure - and both still have a “secret sauce” in their respective OS, reputation, etc. none of which I see happening with these AI giants for some time - perhaps Google’s halo excepted.]
You are seriously underestimating the anthropic or even OpenAI moat. OpenAI has 1 Billion Monthly Active Users is amazing achievement. And Anthropic has much better AI product than OpenAI!!! Anthropic reputation in my view is far superior to IBM, in the last 25 years…
BTW, they need compute capacity, but that doesn’t mean they have to build the capacity. They will buy it from $AMZN, $MSFT, $xAI’s of the world.
Yes, because 94-95% are using their product for free. That’s not to minimize the magnitude of that accomplishment - it’s amazing to have a billion users of any product, free or not. But they’re spending, and losing, a shirt ton of money providing the infrastructure and compute that’s servicing those free customers. At some point, that “free” has to change - and the question is whether there’s going to have the ability to start converting those free users to paying customers at a high enough rate for the massive capex spend to make sense. That’s an even more difficult proposition than building a massive customer base with “free.”
Thanks for making my point. If they have additional users they need to buy more “compute”: either rent it from somebody else, or build it themselves. Neither of those is “almost free additional users” for “scaling” in the sense that I have always heard it used.
The beauty of AirBnB is they built the software, everybody else provided the capital and the labor. Tesla’s Robotaxi was supposed to be “everybody else buys the hardware and software from us, then: presto, Robotaxi fleet.” Now it looks more like Hertz: Tesla will own (or lease) the metal and the software, build over time brick by brick, take care of the back room functions, clean the cars, insure everything, blah blah blah. Big difference.
Radio & TV are a good example: build a small plant once (heck, a radio station can be 3 rooms: a studio, the receptionist, and an office for the sales guys). Then you add listeners/viewers at exactly zero cost. Now that’s scaling!
I have a niece who writes lesson plans and teaching aids for teachers and sells them on “teachers pay teachers”. After two years she stopped teaching because she was making so much “producing once” and selling dozens of times. It’s the Microsoft model, and so far, at least, I don’t see AI building with the same zero acquisition cost for new subscribers, having no supply chain encumberments, etc.
Again, not saying it won’t be a business, just not so crazy as some that have gone before.
The other issue that I keep thinking about is that they have external constraints regarding adding capacity. Back in the software only scaling world, all you needed to do is hire a few hundred more software folks, or if you couldn’t, just import them from Asia. Today, more actual hardware is required and more power and more RAM is required and all of those have external constraints. No matter what you pay, you can’t get enough. And indeed that is why the prices of those things keeps getting bid way up.
The capacity constraint is now clear from SpaceX filing. xAI build significant capacity at record time. Looks like Grok is not being used that heavily and that excess capacity is used by Anthropic and Google!!! $GOOGL is paying almost $1 B a month and Anthropic capacity utilization has some gymnastics involved, where SpaceX can count the revenue now, and Anthropic doesn’t pay now!!! This kind of helps both companies in their IPO quest.
Now, everyone is jumping into the equity raise… so $GOOGL beat everyone. This is going to be one hell of a fund raising time, anywhere between $300 to $500 B is going to be raised.
