The problem vendor finance is not that this is illegal or shady practice, it's that vendor finance is an indicator that perhaps demand is not as strong as the AI hype would suggest. Why would NVIDIA, which is at the centre of a boom where demand is growing exponentially need to support the purchase of its own products?
The comparison to GM is a false equivalence - GM is in a mature industry where vendor finance is a means of expanding your pool of potential consumers. Of course, vendor finance in a mature industry can and often does push to far, where finance providers sacrifice credit quality to unsustainably promote sales, storing up future losses.
I would also note that NVIDIA is also providing vendor financing slightly less deliberately via its trade receivables which have been growing faster than its revenues and have spiked recently.
To argue that legality is the concern is to argue against a slight misunderstanding of the problem. The reason why this is a bubble indicator is that in a business model as unsustainable as the current AI model (where not a single downstream firm is making a profit), the fact that the poster child of the industry is increasingly depending on vendor financing to support demand for its chips suggests that maybe even at the centre growth may not be quite as high as anticipated. It's important to remember that current valuations are based on continued, unprecedented, parabolic growth, not the kind of growth that vendor financing normally fuels.
If you point to a startup like Uber supporting demand in a similar way at early stage I would argue that Uber was establishing market share in an established industry, whereas NVIDIA is the dominant player in a market which, if it is to grow at the rate valuations depend on, will see unprecedented expansion.
The reason why they're a false equivalence is stated in the quotes you cited (which makes me wonder why you cited them?). In one case a mature industry where rapid growth has tailed off is looking for ways to increase demand (or as stated by another comment to reduce negative carry by offloading inventory). In the other, you have unprecedented growth and extremely restricted supply, why would they use funds to increase demand when demand is so strong and rapidly growing? Where vendor financing would be normal in one, it is unusual in the other.
My wife worked in auto finance for 35 years. GM and other auto companies finance divisions would routinely offer special rates and terms for cars in their line up that were not selling, including interest free loans. At times the auto finance divisions would undercut banks and S&Ls across the board. The finance divisions are governed by different rules than banks. They are subsidized by their parent companies and they make loans based on maintaining the solvency of the parent company.
A line of cars that are just sitting in a warehouse is costing a carmaker a lot in interest. Auto finance companies will take a hit on their books to make loans to move them if it benefits the entire company's balance sheet.
“The problem vendor finance is not that this is illegal or shady practice, it's that vendor finance is an indicator that perhaps demand is not as strong as the AI hype would suggest. Why would NVIDIA, which is at the centre of a boom where demand is growing exponentially need to support the purchase of its own products? “
Because who is going to lend OpenAI hundreds of billions to buy chips in NVIDIA? OpenAI will end up having down financing rounds if it keeps trying to cash raise from investors.
Caveat -- I don't like LLMs. They are clearly derived works of pretty much everything written, and the only reason this wholesale theft is tolerated is financial FOMO, i.e. everyone wants a cut of the loot.
But anyway, the biggest question for investors in AI at this time is whether AI can become a real business, where the fees paid by customers have a chance of covering the expenses of training and operating these models. The expected spend on AI data centers over the next 5 years is about $4-5T, while right now, OpenAI's revenue (not income) is annualized at about $13B, or about 60X lower. Their income, of course, is seriously negative.
This is nothing like GM's financing car purchases. Those purchases will be paid for, plus interest over a fixed small number of years, by customers who believe the cars have real value to them. And GM will make money both from manufacturing the cars and from financing their customers.
In comparison, OpenAI and their ilk don't have *any* customers paying a price that comes close to covering OpenAI's expenses, and they don't even project profitability for the next 4-5 years (1). Seen from that vantage point, Nvidia and AMD are just trying to keep the party going a bit longer, perhaps until their executives can cash out more of their options/RSUs.
(1) -- and that's assuming the AI companies don't eventually have to fork out $$$ to the authors whose works they've incorporated into their models.
They never will have to fork out one cent (except to their lawyers) because LLM training is a derivative work which is, and always has been, fair use. The judges who have ruled on this so far agree.
Certainly there has been two very recent cases where a judge ruled that LLM training under some circumstances is fair use. But saying that training "always has been fair use" while vacuously true, ignores the fact that these rulings are like 4 months old by just two individual judges, one of whom suggested that if plaintiffs can show harm to their market, they may revisit things. There is very little case law here.
Here are the rules according to Google's AI summary, of when a copyright violation is fair use:
1 -- Purpose and character of the use: Is the use for non-profit, educational, or research purposes, or is it for commercial gain? A "transformative" use that adds new meaning or expression is more likely to be considered fair use.
2 -- Nature of the copyrighted work: Creative works like fiction and music receive strong copyright protection, while factual works may have "thinner" protection.
3 -- Amount and substantiality of the portion used: How much of the original work was used? Using a small amount is generally more likely to be fair use, but if the portion used is the "heart" of the work, it may not be considered fair use.
4 -- Effect of the use upon the potential market for or value of the copyrighted work: Does your use harm the potential market for the original work? For example, uploading an entire textbook to a website would likely harm its market.
IANAL, but WRT #1, we're definitely talking about commercial gain. I'm not sure whether you'd call AI 'transformative,' but some of the uses are just summaries of the copyrighted material.
WRT #2, AFAICT LLMs are trained on opinion, facts, fiction, pretty much everything.
WRT #3, well, the whole work is used.
WRT #4, LLMs do a lot of damage to the market for the original work. Lots of sites aren't getting click throughs any more, because the AI summary contains the core of the site's material.
So, even with today's laws, I'd say LLMs shouldn't be considered fair use. The only argument on the other side would be that an LLM response is 'transformative,' but in a lot of cases, the LLM response is pretty much just quoting directly from the original text.
On top of that, the laws can change. When Internet Radio just started to be a thing, the copyright laws (the DMCA) changed things so that unlike analog radio, digital radio had to pay per-stream performance royalties, because the copies you can make from a digital stream are so much better than you get from recording from an FM signal. Similarly, Congress might change the law to protect content creators from having their work essentially stolen.
Noah has written about this before, but in this piece he doesn't mention what seems like the biggest problem with these circular deals, and the AI spending boom in general - where is the cash coming from to build all the stuff that's getting built?
As I understand it, the AI boom has a big debt footprint. Non-bank private credit (venture/asset-backed loans, private credit funds, SPV financing) has grown rapidly and is now financing everything from GPU fleets to purpose-built data centers. Example: CoreWeave raised a multi-billion dollar debt package to scale data centers. Project finance and private credit seem like they're central to hyperscaler and specialist cloud buildouts. It seems likely that the circular deals are a big enabler for all this debt finance.
IMO, the riskiness of all this will come down to whether the counterparties holding the debt can handle the losses when they eventually come. The problem is it's unclear if anyone understands either the magnitude of the debt or who's holding it.
This is a good summary of the situation and makes good sense. I find the OpenAI deal with AMD most interesting. AMD is the closest thing to a competitor that NVidia has, but without CUDA their chips are best used only for inference. AMD now is desperately trying to keep up and OpenAI made them an offer they couldn’t refuse, and even though AMD will be able to sell lots more chips, OpenAI will benefit most because they will get the upside on the stock warrants they got in return.
That’s a really good clarification. I am still unclear what kind of progress do we need to see in those companies buying all this Ai and compute. I can’t see consumer uptake being the driver of future profits. Interestingly the current earnings season seems to be going well. I note you mention some revenue increase in last six months. If sustained will that justify the forward PEs ?
Bubbles are something to worry about, But when it comes to AI my primary concern is their vast energy consumption and how it's driving up energy prices for consumers. I'd like to see AI companies be forced to provide their own energy sources.
I don't like that the burden of paying for it falls on poor and average Joe consumers. Not to mention that rising energy costs hurt every single other business.
While I think it's arguable that we need to win the AI race for national security, then it should be treated like any other national security concern and be subsidized by the government.
Isn’t the big backlog right now getting permits to put new electricity sources onto the grid? My understanding is that a lot of data centers are building out local power supplies so that they don’t have to get grid connections and can skip that backlog, but that the rise in prices isn’t because electricity demand is outstripping the construction of supply, but rather its outstripping the *connection* of supply.
An AI bust could crash the stock market and reveal ex ante that absent data center construction were already in a tariff-induced recession, but I’m less worried about it causing wider economic contagion because it the boom doesn’t seem (though I don’t really know) to be heavily leveraged of financed by the banking sector. As such an AI crash doesn’t imply a broader credit crunch which is always the really scary scenario. I could be wrong and somehow Chase or whomever could be highly leveraged in AI but that doesn’t look like the case at present.
The most worry to me is to see that the numbers of decision makers are few and that their biased view of the yet unproven AI industry supremacy is also unproven.
Also, apart from heads over heals, why on earth is AI the next to only investable «industry»?
Good Analysis... I would add that AI value (as shown by revenue) has not caught up with AI investment.... yet. If it catches up, no problem. However, even if it does not, there will be pain, but real value is being created. This sort of bubble is not as bad as a financial bubble.
When the average person sees a graph like that - they are inferring that fake money is being "created"; when a researcher looks at a graph like that - they are inferring that the companies are highly correlated.
As someone who lived through a number of spectacular failures of financial manipulation, from the original dot com bubble and the collapse of financial engineering marvels like Enron to the financial ouroboros of the subprime mortgage market, I think that this time it really might be different! I think that many of the important pioneers of this field are acknowledging the limitations of LLMS and realizing that AGI is not immanent so I don’t feel that it is as much hype as it used to be. And the AI based tools are still very useful and continuing to be adopted. So I don’t think as a technology oriented person that it’s a classic bubble like the dot com era where you had a lot of companies that had no realistic way to profit! And let’s not forget that a lot of these companies like Oracle actually survived the dot com bubble. I think a lot of these people are industry veterans and I have more confidence that this AI investment is going to go away. Thanks for a very thoughtful piece about this important technology!
The 1999 DotCom crash simply preceded the financial crash in 2007/8. At some time the revenues from AI will not be sufficient to pay the interest on all the borrowing and repay the loans. It took WW2 to restore the economies of most Western countries after the 1929 crash and the effects of the financial crisis of 2007/8 are still with us, disrupting politics, encouraging right wing politics like Trump and Farage. Perhaps someone will see the light and encourage the AI companies to merge with the chip makers and take over the data centres.
The problem vendor finance is not that this is illegal or shady practice, it's that vendor finance is an indicator that perhaps demand is not as strong as the AI hype would suggest. Why would NVIDIA, which is at the centre of a boom where demand is growing exponentially need to support the purchase of its own products?
The comparison to GM is a false equivalence - GM is in a mature industry where vendor finance is a means of expanding your pool of potential consumers. Of course, vendor finance in a mature industry can and often does push to far, where finance providers sacrifice credit quality to unsustainably promote sales, storing up future losses.
I would also note that NVIDIA is also providing vendor financing slightly less deliberately via its trade receivables which have been growing faster than its revenues and have spiked recently.
To argue that legality is the concern is to argue against a slight misunderstanding of the problem. The reason why this is a bubble indicator is that in a business model as unsustainable as the current AI model (where not a single downstream firm is making a profit), the fact that the poster child of the industry is increasingly depending on vendor financing to support demand for its chips suggests that maybe even at the centre growth may not be quite as high as anticipated. It's important to remember that current valuations are based on continued, unprecedented, parabolic growth, not the kind of growth that vendor financing normally fuels.
If you point to a startup like Uber supporting demand in a similar way at early stage I would argue that Uber was establishing market share in an established industry, whereas NVIDIA is the dominant player in a market which, if it is to grow at the rate valuations depend on, will see unprecedented expansion.
> it's that vendor finance is an indicator that perhaps demand is not as strong as the AI hype would suggest
> GM is in a mature industry where vendor finance is a means of expanding your pool of potential consumers
So in both cases it's a means of selling product you otherwise wouldn't sell. This is a perfectly legitimate equivalence.
The reason why they're a false equivalence is stated in the quotes you cited (which makes me wonder why you cited them?). In one case a mature industry where rapid growth has tailed off is looking for ways to increase demand (or as stated by another comment to reduce negative carry by offloading inventory). In the other, you have unprecedented growth and extremely restricted supply, why would they use funds to increase demand when demand is so strong and rapidly growing? Where vendor financing would be normal in one, it is unusual in the other.
My wife worked in auto finance for 35 years. GM and other auto companies finance divisions would routinely offer special rates and terms for cars in their line up that were not selling, including interest free loans. At times the auto finance divisions would undercut banks and S&Ls across the board. The finance divisions are governed by different rules than banks. They are subsidized by their parent companies and they make loans based on maintaining the solvency of the parent company.
A line of cars that are just sitting in a warehouse is costing a carmaker a lot in interest. Auto finance companies will take a hit on their books to make loans to move them if it benefits the entire company's balance sheet.
That sounds functionally equivalent to putting something on sale, which is normal.
“The problem vendor finance is not that this is illegal or shady practice, it's that vendor finance is an indicator that perhaps demand is not as strong as the AI hype would suggest. Why would NVIDIA, which is at the centre of a boom where demand is growing exponentially need to support the purchase of its own products? “
Because who is going to lend OpenAI hundreds of billions to buy chips in NVIDIA? OpenAI will end up having down financing rounds if it keeps trying to cash raise from investors.
Caveat -- I don't like LLMs. They are clearly derived works of pretty much everything written, and the only reason this wholesale theft is tolerated is financial FOMO, i.e. everyone wants a cut of the loot.
But anyway, the biggest question for investors in AI at this time is whether AI can become a real business, where the fees paid by customers have a chance of covering the expenses of training and operating these models. The expected spend on AI data centers over the next 5 years is about $4-5T, while right now, OpenAI's revenue (not income) is annualized at about $13B, or about 60X lower. Their income, of course, is seriously negative.
This is nothing like GM's financing car purchases. Those purchases will be paid for, plus interest over a fixed small number of years, by customers who believe the cars have real value to them. And GM will make money both from manufacturing the cars and from financing their customers.
In comparison, OpenAI and their ilk don't have *any* customers paying a price that comes close to covering OpenAI's expenses, and they don't even project profitability for the next 4-5 years (1). Seen from that vantage point, Nvidia and AMD are just trying to keep the party going a bit longer, perhaps until their executives can cash out more of their options/RSUs.
(1) -- and that's assuming the AI companies don't eventually have to fork out $$$ to the authors whose works they've incorporated into their models.
They never will have to fork out one cent (except to their lawyers) because LLM training is a derivative work which is, and always has been, fair use. The judges who have ruled on this so far agree.
Certainly there has been two very recent cases where a judge ruled that LLM training under some circumstances is fair use. But saying that training "always has been fair use" while vacuously true, ignores the fact that these rulings are like 4 months old by just two individual judges, one of whom suggested that if plaintiffs can show harm to their market, they may revisit things. There is very little case law here.
Here are the rules according to Google's AI summary, of when a copyright violation is fair use:
1 -- Purpose and character of the use: Is the use for non-profit, educational, or research purposes, or is it for commercial gain? A "transformative" use that adds new meaning or expression is more likely to be considered fair use.
2 -- Nature of the copyrighted work: Creative works like fiction and music receive strong copyright protection, while factual works may have "thinner" protection.
3 -- Amount and substantiality of the portion used: How much of the original work was used? Using a small amount is generally more likely to be fair use, but if the portion used is the "heart" of the work, it may not be considered fair use.
4 -- Effect of the use upon the potential market for or value of the copyrighted work: Does your use harm the potential market for the original work? For example, uploading an entire textbook to a website would likely harm its market.
IANAL, but WRT #1, we're definitely talking about commercial gain. I'm not sure whether you'd call AI 'transformative,' but some of the uses are just summaries of the copyrighted material.
WRT #2, AFAICT LLMs are trained on opinion, facts, fiction, pretty much everything.
WRT #3, well, the whole work is used.
WRT #4, LLMs do a lot of damage to the market for the original work. Lots of sites aren't getting click throughs any more, because the AI summary contains the core of the site's material.
So, even with today's laws, I'd say LLMs shouldn't be considered fair use. The only argument on the other side would be that an LLM response is 'transformative,' but in a lot of cases, the LLM response is pretty much just quoting directly from the original text.
On top of that, the laws can change. When Internet Radio just started to be a thing, the copyright laws (the DMCA) changed things so that unlike analog radio, digital radio had to pay per-stream performance royalties, because the copies you can make from a digital stream are so much better than you get from recording from an FM signal. Similarly, Congress might change the law to protect content creators from having their work essentially stolen.
Noah has written about this before, but in this piece he doesn't mention what seems like the biggest problem with these circular deals, and the AI spending boom in general - where is the cash coming from to build all the stuff that's getting built?
As I understand it, the AI boom has a big debt footprint. Non-bank private credit (venture/asset-backed loans, private credit funds, SPV financing) has grown rapidly and is now financing everything from GPU fleets to purpose-built data centers. Example: CoreWeave raised a multi-billion dollar debt package to scale data centers. Project finance and private credit seem like they're central to hyperscaler and specialist cloud buildouts. It seems likely that the circular deals are a big enabler for all this debt finance.
IMO, the riskiness of all this will come down to whether the counterparties holding the debt can handle the losses when they eventually come. The problem is it's unclear if anyone understands either the magnitude of the debt or who's holding it.
This is a good summary of the situation and makes good sense. I find the OpenAI deal with AMD most interesting. AMD is the closest thing to a competitor that NVidia has, but without CUDA their chips are best used only for inference. AMD now is desperately trying to keep up and OpenAI made them an offer they couldn’t refuse, and even though AMD will be able to sell lots more chips, OpenAI will benefit most because they will get the upside on the stock warrants they got in return.
That’s a really good clarification. I am still unclear what kind of progress do we need to see in those companies buying all this Ai and compute. I can’t see consumer uptake being the driver of future profits. Interestingly the current earnings season seems to be going well. I note you mention some revenue increase in last six months. If sustained will that justify the forward PEs ?
Bubbles are something to worry about, But when it comes to AI my primary concern is their vast energy consumption and how it's driving up energy prices for consumers. I'd like to see AI companies be forced to provide their own energy sources.
I don't like that the burden of paying for it falls on poor and average Joe consumers. Not to mention that rising energy costs hurt every single other business.
While I think it's arguable that we need to win the AI race for national security, then it should be treated like any other national security concern and be subsidized by the government.
Isn’t the big backlog right now getting permits to put new electricity sources onto the grid? My understanding is that a lot of data centers are building out local power supplies so that they don’t have to get grid connections and can skip that backlog, but that the rise in prices isn’t because electricity demand is outstripping the construction of supply, but rather its outstripping the *connection* of supply.
An AI bust could crash the stock market and reveal ex ante that absent data center construction were already in a tariff-induced recession, but I’m less worried about it causing wider economic contagion because it the boom doesn’t seem (though I don’t really know) to be heavily leveraged of financed by the banking sector. As such an AI crash doesn’t imply a broader credit crunch which is always the really scary scenario. I could be wrong and somehow Chase or whomever could be highly leveraged in AI but that doesn’t look like the case at present.
The most worry to me is to see that the numbers of decision makers are few and that their biased view of the yet unproven AI industry supremacy is also unproven.
Also, apart from heads over heals, why on earth is AI the next to only investable «industry»?
I believe the sheep factor is at work.
"data centers are almost all about AI now" - do you mean new construction of data centers?
Good Analysis... I would add that AI value (as shown by revenue) has not caught up with AI investment.... yet. If it catches up, no problem. However, even if it does not, there will be pain, but real value is being created. This sort of bubble is not as bad as a financial bubble.
When the average person sees a graph like that - they are inferring that fake money is being "created"; when a researcher looks at a graph like that - they are inferring that the companies are highly correlated.
As someone who lived through a number of spectacular failures of financial manipulation, from the original dot com bubble and the collapse of financial engineering marvels like Enron to the financial ouroboros of the subprime mortgage market, I think that this time it really might be different! I think that many of the important pioneers of this field are acknowledging the limitations of LLMS and realizing that AGI is not immanent so I don’t feel that it is as much hype as it used to be. And the AI based tools are still very useful and continuing to be adopted. So I don’t think as a technology oriented person that it’s a classic bubble like the dot com era where you had a lot of companies that had no realistic way to profit! And let’s not forget that a lot of these companies like Oracle actually survived the dot com bubble. I think a lot of these people are industry veterans and I have more confidence that this AI investment is going to go away. Thanks for a very thoughtful piece about this important technology!
This is very problematic but I think the interlocking deals w
Excellent article. No arguments with your positions.
The 1999 DotCom crash simply preceded the financial crash in 2007/8. At some time the revenues from AI will not be sufficient to pay the interest on all the borrowing and repay the loans. It took WW2 to restore the economies of most Western countries after the 1929 crash and the effects of the financial crisis of 2007/8 are still with us, disrupting politics, encouraging right wing politics like Trump and Farage. Perhaps someone will see the light and encourage the AI companies to merge with the chip makers and take over the data centres.