"Hyperscalers, by contrast, look like winners. Their cloud revenue (a stat Wall Street uses to measure AI demand) grew 50% in the second quarter. They have huge contract backlogs and the capacity to absorb higher borrowing costs."
And if that revenue was coming from end users, it would be great. But it isn't---most of it is coming from the very same debt that is the subject of the article. There's a huge difference in my financial outlook if my biggest customer is borrowing all the money they are spending with me vs earning it in the marketplace. In the latter scenario, my business is tied into the success of the my customer in the market (harder to grow), in the former it's just a question of how much money they can borrow (much easier to grow, especially if I'm Nvidia and I can guarantee their debt).
I see the business press make this fundamental mistake all the time, starting off by drawing a boundary around the AI industry as a whole and correctly assessing the staggering levels of debt, but then switching perspective to individual sectors and completely forgetting about where the revenue actually originates, and treating it the same as organic growth.
The markets are just in the Twilight Zone at the moment. The CDS on some of these bonds are now priced assuming the bonds will fail (or at least more likely than not). Per headlines on Friday, Anthropic would likely be rated as “junk” status from a bond risk standpoint were it to go public.
From a financial standpoint this is all looking like early 2008. The risk assessment bits of the market seem to have concluded this is all totally f*ked. But money keeps getting pumped into things that the market has decided will fail because the only thing that stops it from imploding is pumping more money in.
Like 2008 once key players get their armor on, expect things to suddenly unleash in a way that gets real ugly real quick. This is essentially what the article is saying. This all only hasn’t collapsed because money lenders haven’t pulled the plug yet. When that happens, and with every day this looks like a question of when not if, it will be a $hitstorm the likes of which we’ve not seen in a while.
Methinks they're calling it a bit too early. Earnings volatility combined with the upcoming midterms, some profit taking after a substantial rally and possibly investors freeing up capital ahead of some major IPOs could explain a good portion of the drop. There have been some disappointing earnings and forecasts that added fuel to the fire, and the increasing cost of financing AI infrastructure is certainly a legitimate concern, but I don't think we've seen enough evidence of a fundamental deterioration in demand to conclude that the whole thing is coming apart. Unless the bad news continues to stack up, I don't think this is The End.
As much as it'd thrill me in one of the worst Pyrrhic victories to see the AI bubble burst, I think it is way too early to claim that a single bad day or a single bad week means it's collapsing.
The compute they sold is based on the idea that the amount of money people spend on inference will continue to go up dramatically.
If we’re in a gas shortage, all that matters is who has the gas, true. But the issue is that we’re burning 50% of the gas that’s made every day, mean while the other 50% is being stockpiled based on the idea that everyone is going to drive 10x farther next year.
Is that true? Aren't they going to use those datacenters for training and robotics as well? Not just chatbot inference? GPUs aren't Bitcoin miner ASICs, they can do a wider variety of things.
Gasoline in a car is one analogy, but another one is data on the Internet. When the Internet first broke out of universities, there was dial-up. In the late 70's the 300 baud modem was being sold to consumers. It would take roughly three hours to download a novel of 350 pages. It would take roughly two hours to download a single poor quality 256 kilobyte raw 8bpp image file. Jpeg hadn't been invented yet. Now, that image file will take less than ten milliseconds. Roughly a million times faster.
The things we transferred changed too. Today, we stream multiple 4k video with audio streams up and down, so a regular home connection won't suffice for more than a few users. Wirth's law; what Intel giveth, Microsoft taketh away; also applies.
Sometimes digital trends do tend to 10x, or in the case of ISP speeds, we went from 300 baud to gigabit in ~40 years; 3.3 million times faster.
I have no idea what the future holds, but analogies only go so far.
> Is that true? Aren't they going to use those datacenters for training and robotics as well?
Sure, those things too. I was meaning “inference” as a shorthand for any ai workloads.
Whatever analogy you want to use: gas, data transfer, it doesn’t matter. My point is that all the capacity is being hoarded for future use making it feel like a scarce resource now.
If the demand grows like they say it will, then hoarding was a smart choice and the people who own the GPUs will make lots of money. If the demand doesn’t grow as fast and as much as they say, then hoarding was a bad call and the house of cards collapses and a lot of people will lose a lot of money.
To use your data transfer analog - during the dotcom bubble a ton of fiber was built. But after the bubble popped most of it went unused for years. Some of that “dark fiber” is still unused today, 25+ years later. Afaik it would have been a lot cheaper and more stable for the economy to lay that fiber more incrementally.
1. GPUs are reportedly increasingly just getting put on shelves because there aren’t data centers to host them, which is creating a setup for a financial mess (rapidly depreciating assets purchased, paid for, but not utilized),
2. And massive chunk of the compute “demand” is coming from circular financing (ie AWS is “sold out” in large part because it gave startups like Anthropic money to buy the thing it’s now telling the market is “sold out” so there’s some shenanigans behind those headlines.)
I do agree though that compute is really the only place to make money in AI in the long run. Models are an increasingly a commodity with little value above the cost of compute to run them. However selling compute will be with razor thin margins, not the fat AWS margins the market got used to in cloud, so AWS and others will have to deal with that uncomfortable reality.
I guess GPUs are interesting because unlike houses, we've suddenly found a massive amount of new utility for these things. If my house suddenly started producing its own furniture, it might be a bit more valuable. Which isn't to say people haven't found new things to do with their houses - COVID made having a home office a lot more valuable, for example, but nothing quite as dramatic as GPUs suddenly able to produce their own code.
In 2008 we had both a massive under supply of housing in many cities plus crashing prices. These things can very much exist at the same time if, as if happening now with GPUs, the financial engine fueling the ecosystem is out of whack.
"Hyperscalers, by contrast, look like winners. Their cloud revenue (a stat Wall Street uses to measure AI demand) grew 50% in the second quarter. They have huge contract backlogs and the capacity to absorb higher borrowing costs."
And if that revenue was coming from end users, it would be great. But it isn't---most of it is coming from the very same debt that is the subject of the article. There's a huge difference in my financial outlook if my biggest customer is borrowing all the money they are spending with me vs earning it in the marketplace. In the latter scenario, my business is tied into the success of the my customer in the market (harder to grow), in the former it's just a question of how much money they can borrow (much easier to grow, especially if I'm Nvidia and I can guarantee their debt).
I see the business press make this fundamental mistake all the time, starting off by drawing a boundary around the AI industry as a whole and correctly assessing the staggering levels of debt, but then switching perspective to individual sectors and completely forgetting about where the revenue actually originates, and treating it the same as organic growth.
The markets are just in the Twilight Zone at the moment. The CDS on some of these bonds are now priced assuming the bonds will fail (or at least more likely than not). Per headlines on Friday, Anthropic would likely be rated as “junk” status from a bond risk standpoint were it to go public.
From a financial standpoint this is all looking like early 2008. The risk assessment bits of the market seem to have concluded this is all totally f*ked. But money keeps getting pumped into things that the market has decided will fail because the only thing that stops it from imploding is pumping more money in.
Like 2008 once key players get their armor on, expect things to suddenly unleash in a way that gets real ugly real quick. This is essentially what the article is saying. This all only hasn’t collapsed because money lenders haven’t pulled the plug yet. When that happens, and with every day this looks like a question of when not if, it will be a $hitstorm the likes of which we’ve not seen in a while.
Methinks they're calling it a bit too early. Earnings volatility combined with the upcoming midterms, some profit taking after a substantial rally and possibly investors freeing up capital ahead of some major IPOs could explain a good portion of the drop. There have been some disappointing earnings and forecasts that added fuel to the fire, and the increasing cost of financing AI infrastructure is certainly a legitimate concern, but I don't think we've seen enough evidence of a fundamental deterioration in demand to conclude that the whole thing is coming apart. Unless the bad news continues to stack up, I don't think this is The End.
It’s “How much longer will the bond market sustain the nonstop spending spree?”
As much as it'd thrill me in one of the worst Pyrrhic victories to see the AI bubble burst, I think it is way too early to claim that a single bad day or a single bad week means it's collapsing.
The only thing that matters is who has the Compute.
I bought a gaming GPU six months ago for running local LLMs, at outrageous price (2x MSRP), and in 6 months it's price doubled again.
AWS already sold all the compute it has planned for 2026, 2027 and 2028.
The compute they sold is based on the idea that the amount of money people spend on inference will continue to go up dramatically.
If we’re in a gas shortage, all that matters is who has the gas, true. But the issue is that we’re burning 50% of the gas that’s made every day, mean while the other 50% is being stockpiled based on the idea that everyone is going to drive 10x farther next year.
Is that true? Aren't they going to use those datacenters for training and robotics as well? Not just chatbot inference? GPUs aren't Bitcoin miner ASICs, they can do a wider variety of things.
Gasoline in a car is one analogy, but another one is data on the Internet. When the Internet first broke out of universities, there was dial-up. In the late 70's the 300 baud modem was being sold to consumers. It would take roughly three hours to download a novel of 350 pages. It would take roughly two hours to download a single poor quality 256 kilobyte raw 8bpp image file. Jpeg hadn't been invented yet. Now, that image file will take less than ten milliseconds. Roughly a million times faster.
The things we transferred changed too. Today, we stream multiple 4k video with audio streams up and down, so a regular home connection won't suffice for more than a few users. Wirth's law; what Intel giveth, Microsoft taketh away; also applies.
Sometimes digital trends do tend to 10x, or in the case of ISP speeds, we went from 300 baud to gigabit in ~40 years; 3.3 million times faster.
I have no idea what the future holds, but analogies only go so far.
> Is that true? Aren't they going to use those datacenters for training and robotics as well?
Sure, those things too. I was meaning “inference” as a shorthand for any ai workloads.
Whatever analogy you want to use: gas, data transfer, it doesn’t matter. My point is that all the capacity is being hoarded for future use making it feel like a scarce resource now.
If the demand grows like they say it will, then hoarding was a smart choice and the people who own the GPUs will make lots of money. If the demand doesn’t grow as fast and as much as they say, then hoarding was a bad call and the house of cards collapses and a lot of people will lose a lot of money.
To use your data transfer analog - during the dotcom bubble a ton of fiber was built. But after the bubble popped most of it went unused for years. Some of that “dark fiber” is still unused today, 25+ years later. Afaik it would have been a lot cheaper and more stable for the economy to lay that fiber more incrementally.
Except:
1. GPUs are reportedly increasingly just getting put on shelves because there aren’t data centers to host them, which is creating a setup for a financial mess (rapidly depreciating assets purchased, paid for, but not utilized),
2. And massive chunk of the compute “demand” is coming from circular financing (ie AWS is “sold out” in large part because it gave startups like Anthropic money to buy the thing it’s now telling the market is “sold out” so there’s some shenanigans behind those headlines.)
I do agree though that compute is really the only place to make money in AI in the long run. Models are an increasingly a commodity with little value above the cost of compute to run them. However selling compute will be with razor thin margins, not the fat AWS margins the market got used to in cloud, so AWS and others will have to deal with that uncomfortable reality.
I worry they’re going to destroy the chips they can’t afford to deploy to keep them from being used for local inference
> rapidly depreciating assets purchased
GPU's produced in 2020, with a supposedly "3 year life before depreciating to zero" are now sold at twice their price from 2020.
So did houses in 2007. That’s not a good sign at all.
I guess GPUs are interesting because unlike houses, we've suddenly found a massive amount of new utility for these things. If my house suddenly started producing its own furniture, it might be a bit more valuable. Which isn't to say people haven't found new things to do with their houses - COVID made having a home office a lot more valuable, for example, but nothing quite as dramatic as GPUs suddenly able to produce their own code.
In 2008 we had both a massive under supply of housing in many cities plus crashing prices. These things can very much exist at the same time if, as if happening now with GPUs, the financial engine fueling the ecosystem is out of whack.