It is an indication of how mainstream the AI madness has been of late that chip makers are now getting into the financing business, most notably Nvidia.
Instead of simply pairing up its own products with the customers who love them best, Nvidia is designing financing packages which will allow those customers to build data centres.
This is a little uncomfortable, as it could put Nvidia in the position of financing demand for its own products.
Nvidia has seen the light – or perhaps all this has been planned from the start. The company has designed a few data centres for its customers to use.
It has signed deals with land and buildings developers, including SB Energy, to lease a site in Ohio which – and this is the really clever bit – will be a data centre Nvidia designs and controls.
This will be leased exclusively to Nvidia’s partner, OpenAI which is the current media darling in the AI sector, for a pile of processors. SB Energy is happy because it has signed up a guaranteed customer.
And Nvidia is doubly happy as it now has a revenue stream off its land and buildings partners designed to make sure they remain happy too. Everybody wins and nobody has to pay more per processor.
Nvidia’s view - ‘We’re in that business because data centers are expensive to build and, you know, somebody has got to build them. This is an example of how we are helping very promising projects get financed.’ Certainly there isn’t anything inherently wrong about financing customers, although some skeptics might cite the adage about lending money to a friend. One problem is that lending money - aka investing - exposes you to the success or otherwise of the project, but as a maker of chips you don’t have a financial stake in whether or not the equipment you supplied works out any better than promised. It seems very likely that if you’re providing financial aid you will have more reason to see your customers prosper, but that is not necessarily misconduct - there are all sorts of ways of doing this, from overt help to obvious bluffing, that don’t amount to gaming the numbers for a processor to make it look better than it is.
The concerns over competition aren’t unfounded. Amazon.com (AMZN), Google (GOOGL), Meta (META) and Microsoft (MSFT) are all developing their own processors while still being major customers of datacenter chip leader Nvidia. Some are even selling the processors they’ve developed as alternatives to Nvidia.
Who Are the Customers?
Some of them are even less established cloud providers scaling their facilities which is noteworthy in that lenders are not as willing to provide financing to these less-established cloud providers as they do not have the borrowing power of the giants such as Microsoft and Google. Anyway while there are plenty of other reasons to be supportive of building a cloud ecosystem that is not all dominated by just the largest tech companies in the world, doing so may create an environment for more investment in competing AI businesses. These companies push to have open-weight models and do more experimentation on the datasets they release and also what the model building frameworks are that is not available is that these companies most likely won’t ever productize but some seems to think it matters. Regardless of their chances of success (maybe half) it does mean that there is a large increase in the number of independent businesses that could eventually be buyers. From Non-Conventional Lenders
It’s entirely possible that Nvidia was looking into some of these down market investment opportunities before dipping into lending.
What I can understand from all of this is that the deal must include Nvidia receiving some or all of the revenue from whatever capacity is sold above the price it is selling the equipment at. And, yes, I understand that this is speculation based on limited information but this is the way these sorts of things work — we just won’t know the details of every part of the deal. The essential point of the deal is that the cloud provider wants to build a facility but it is expensive to borrow the money to do so. By forgoing some future revenue, Nvidia is guaranteeing the value of the equipment, which helps convince lenders that financing the whole project makes sense. The only requirement of Nvidia is to sell them the processors, which they will sell and the cloud provider will use to sell computing services. The final requirement is that customers actually want to buy those computing services. Making the loan safer is not the same thing as ensuring the library will be fuller and quieter.
There would appear to be legitimate case for such a support and it may prove beneficial in some instances. In particular if it allows a customer to demonstrate to lenders that its model will produce a reliable income and it isn’t a leap of faith. This may be the case at least initially, if the customer goes on to buy its own expansion without the support from the supplier. If however it shows the customer requires an increasing amount of support with each stage of the expansion this suggests that expanding was never intended to be a bridge to greater financial health. In either event it would be unwise if the lender accepted such a guarantee without fully understanding what this signifies about the arrangement. Assuming they do understand the arrangement and there is no real question mark against their judgement, there is nothing wrong in principle with it. It’s just important to ensure that an arrangement designed to fulfil this role has shifted risk and it is impossible to tell whether this is so from the construction announcements.
But as we noted with some uneasiness yesterday, there’s an overlap among the three separate business relationships. Nvidia has made an investment in CoreWeave, sells it processors, and has also agreed to purchase any unused capacity that CoreWeave has. If demand is weak, these three relationships can’t be completely separated. This makes the three relationships part of either a genius partnership or a genius group of investment bankers who have, over the years, raised money from institutional investors around the world (sovereign wealth funds, insurers, pension funds) to finance independent vehicles who buy hardware and build the infrastructure to deliver the hardware as computing services. They might then contract with one of the chipmakers to provide technical assistance and guarantee shortfalls of value of the eventual equipment.
In the past, we’ve seen deals that seemed great and turned out not to be. The difference between NVIDIAs potential with and without a partnership might be compared to partnerships that required cash, or worse, debt from shareholders. While those deals are littered across Wall Street’s history, the aftermath of the financial crisis has taught everyone not to expect anything that touts future value without an upfront cost associated with it. Ultimately, how coverage, duration and triggers are set will determine how expensive it ends up being, and it should be the only real judgement of an investor’s enthusiasm.
It does not really care how much capacity its customers will need it explains, in theory. Uncertainty ? Well it will just sell more capacity. In practice, it has a narrow customer base, and does care very much how much capacity these customers will need; this is its ambition of high margins. This should be seen as a gamble, rather than a default commitment; and statements about how it feels about smaller customers is pure marketing. Look for anything in its sales data that it has no choice but to respect, rather than its own commitments; these will be obligations rather than anything that would look like committments. Consider any comments about the gaming and mining fans as Huang with his fingers crossed.
Finally, it’s possible that Huang is right — just not for the reasons he states. As we wrote recently, CoreWeave has found business for older Nvidia hardware, which strongly suggests that there are workloads (notably some sort of inference) which don’t have an immediate alternative. If that’s true, though, the market for older processors must be pretty good, and we’d expect to see those being turned around quickly (upgraded) from current states. It’s quite possible that the low depreciation period which cloud providers like Microsoft and Amazon use is having an effect here, but that is also the point that Michael Burry is raising. This leads to our conclusion: Huang has the benefit of a big reputation with years of evidence in the rear view, but this specific claim has a lot of missing evidence. Each of his bits of evidence could (with some fairly easily implied assumptions) fall apart, at least if you assume that the short term of continued demand for older equipment is really just a market flooded with more capacity. We also expect that there will be more evidence that shows what the continued demand is — specifically because of the lack of clear alternative to newer processors. If we can’t identify those, then it’s only cheaper, more specialised alternatives that we’re looking for.
Bankers running the numbers on Nvidia’s latest and improved products have two distinct reasons while cheap and free for now. One set of clients are interested in developing new ways to lend more money on more profitable terms. The other is new technologies primarily in developing and refining artificial intelligence along with imbuing robots with the ability to work alongside humans more safely and efficiently. Cisco Systems and Lucent Technologies had voracious appetites in the telecoms boom and bust because they were financing a lot of customers who wanted to buy their gear. Just as and if banking gets ahead of itself in a bullish story about the AI phenomenon coming true instead of likelihood it won’t, like the nearly guaranteed chance Cisco and Lucent made the same mistake of thinking demand for their products would justify their returns — the telecoms boom and bust was a case of stock prices getting ahead of themselves not companies having to absorb losses on loans they should have never made. Banking is already factoring in cheaper computing and new technologies will generate demand and returns that will overcome the disappointment of whatever the AI phenomenon will come out to be.
Nvidia seems much more certain of _how_ it’s going to lose on these facilities than GGP is of its $300bn figure. The other big difference is that for Nvidia, the losses only get worse if things go well.
GGP could find itself in an upside scenario where demand is so strong it has trouble keeping the new facilities full. In that case it wouldn’t suffer depreciation obligations on an actual customer shortfall, and the only thing it might do is deploy some extra capital on buying services.
For Nvidia, though, the scenario where demand completely vanishes just piles on the bad news. It doesn’t just mean that its growth was huge, it also means that new facilities operators built _under the bet that_ they would be able to sell their capacity at profitable prices. Network demand might also come in such a scenario, but there are definitely situations where operators are stuck with more computing services than they have customers for.
CreditSights analyst seems not overly worried that Nvidia will see losses arise from these substantial guarantees, and neither am I — were an economic downturn severe enough to create losses on such guarantees to occur, Nvidia would be just fine financially (with an oversized cash pile and liquid investments) and still doll it out of its operating business pretty profitably. Furthermore, one isn’t guaranteed anything to begin with, and any contrary view assumes guarantees are free simply because the company doesn’t have to put up any cash immediately.
To be sure, Nvidia’s guarantees aren’t shrinking by any means. As CreditSights analyst Andy Li points out, So far, people are betting against a company that hasn’t broken. But it’s a question of how long you can push the pedal until something breaks.
Li is only suggesting cautions in regards to Nvidia’s accumulating obligations, not that current levels are unmanageable. To that point, rivals are responding to Nvidia’s guarantees with some head-scratching options of their own. AMD, for instance, has offered to sell equity stakes in exchange for large chip orders. Broadcom has agreed to cover a possible payment shortfall in a financing arrangement involving Google and Anthropic. Once competitors start helping customers finance their purchases, refusing to join in means risking the order as well.
The revealing moment will come when an existing customer wants to expand. If it can finance the next facility on the strength of its operating business, the earlier support will have helped establish something durable. If it needs a larger guarantee from Nvidia to make the numbers work again, another chip order will arrive with another obligation attached. The sales announcement will look much the same either way.
References:
Thank you for this excellent analysis. I wrote something of much lesser importance: https://debradouglas007.substack.com/p/whos-your-banker?r=ft54k&utm_medium=ios