Those of you with long memories may recall how one of the biggest casualties of the financial crisis were the financial guarantors. Companies like Ambac and MBIA grew out of their traditional business of guaranteeing payment of municipal debt to guaranteeing mortgage securities like CDOs.

We now have the AI version of financial guarantee. Except this time, it’s not being written by insurers or even banks.

No, this time, it’s being written by the largest tech companies in the world – and they aren’t charging an actuarial sound premium for it or holding any reserves for the risk!

$250 B-I-L-L-I-O-N!

You may have seen the recent headline that NVIDIA is providing a $250 billion – that’s with a B – guarantee to data center lenders in order to finance a giant data center in Ohio for OpenAI.

One more time, to make it super clear – that’s not $250 million, it’s $250 BILLION!

What does this mean and why was it necessary? OpenAI wants to build the world’s largest data center. The good news is Softbank is willing to operate the data center and the US government is willing to provide land to construct it on.

The bad news is OpenAI doesn’t have the money to pay for the data center. So, instead, it’s going to lease it. The problem with that is Softbank doesn’t have $250B either. So, they have to borrow the money.

The problem with that is lenders aren’t stupid and don’t want to lend $250B on spec hoping OpenAI will be able to pay Softbank rent so Softbank can pay their lenders.

This would appear to be an unsolvable circular loop. Lenders need security of some sort and neither OpenAI or Softbank have it. What to do?

Nvidia To The Rescue

OpenAI could ask the government to guarantee the project but even the US Treasury isn’t so reckless as to provide a $250B letter of credit. Plus, the Feds are already giving them land and permits to construct their data center faster. You don’t want to be greedy and ask for a LOC too!

But lucky for OpenAI, they know of one other person with almost as much money as the US government. Jensen Huang, the CEO of Nvidia!

Maybe they could convince Nvidia to convince the lenders to give them the money? Of course, Nvidia would do this out of the goodness of their hearts and not because of a quid pro quo that OpenAI will buy tens of billions of dollars of Nvidia chips to operate the new data center!

And guess what? The kindly, saintly Jensen agreed to help out. And boy did he help out! As mentioned, he has guaranteed the entire $250B to the lenders. What a guy!

Specifically, if OpenAI can’t pay Softbank rent, Nvidia will step into its shoes and pay instead to make sure the lenders get paid in full. In the insurance world, we call this a financial guarantee.

Financial Guarantee Insurance

So how does financial guarantee work when written by an insurer? Let’s take the traditional use case of municipal debt.

Your local town would issue some debt for say the water system. As part of the issuance, they would ask an insurer to guarantee the debt. Why? Because it lowered the issuance cost as bond buyers could sleep better at night knowing they didn’t have to worry about the underlying credit of every little town in their portfolio.

There was an arbitrage as the municipality saved say 0.25% on the coupon and they gave maybe 0.1% to the bond insurer (it’s been a minute since I’ve thought about bond insurance so I can’t remember the exact numbers anymore but that’s a decent approximation I think).

Thus, on a $250B issue, Nvidia should be charging call it $250M of premium per year. Let’s say this is ten year arrangement and that’s $2.5B of cumulative premium. As best I can tell, they are charging $0.

Of course, they will sell maybe $100B of Nvidia chips to OpenAI which will make them much more than $2.5B. Now, reasonable people can argue OpenAI would buy those chips regardless if customer demand exists for them, but maybe Nvidia doesn’t think customer demand exists for them???

Regardless, Nvidia is not charging an insurance premium. You know what else they’re not doing? Holding capital against the risk!

Again, I can’t remember how much capital bond insurers had to hold relative to exposure but let’s call it 1%. That’s $2.5B of required capital!

Now, you can argue Nvidia can easily afford to set up a sub with $2.5B of capital to manage the risk. But they haven’t! No finance company could get away with “don’t worry we have enough capital, why make us set any aside?”.

Evaluating the Risk

Also, let’s be honest, this isn’t equivalent to a municipal bond risk. A sewer bond guaranteed by the government is a far lesser risk than a spec data center. The appropriate capital charge is probably $10B or maybe even $25B.

I mean, realistically, if Chinese models supplant OpenAI and OpenAI runs into financial troubles, who is stepping in to use the world’s largest data center?

Maybe somebody would but likely at a far lower rent. So it’s not hard to imagine a scenario where Nvidia could lose $100B. You can argue that’s only like a year of income so they could recover from that and it wouldn’t bankrupt the company, but it certainly would be painful.

Think how reinsurance stocks act when they lose a year of earnings to a storm! Oh, and on top of that, in the scenario where OpenAI can’t pay the rent for their data center that means Nvidia isn’t earning $100B a year anymore. It might be half of that…or quarter…or a tenth!

As recently as three years ago, Nvidia’s earnings were in the $5B a year range, in which case a $100B financial guaranty loss is twenty years worth of earnings lost!

So yeah, it’s a catastrophic risk to take.

Could they ignore the noise and hope it goes away? Let’s not forget AIG’s crisis wasn’t created by the actual insured losses. It was caused by not having enough capital to post more collateral when the perception of potential future losses increased.

While Nvidia doesn’t have the same regulatory constraints, if markets perceived the guarantee as being in the money, pressure would build on Nvidia to show it can make good. Rating agencies could downgrade them, they could get cut off from debt markets, suppliers might require LOCs to ensure they get paid, etc.

So while it wouldn’t be as hard a trigger as AIG or Ambac faced in 08, Nvidia can’t treat it like asbestos circa 2000 and hope nobody will notice it has a huge unfunded liability.

The Cat Analogy

Speaking of catastrophic risk, nat cat insurance is a great framework for evaluating the risk Nvidia is taking. Imagine if Ajit wrote a $250B line on Miami hurricane. That would be a bet the company risk.

The difference is a cat 5 in Miami is a 1 in 100+ event. A collapse in AI demand may be a 1 in 10? So Nvidia is taking a lot higher AAL for a lot lower ROL…and holding no capital for the downside.

Worse, this is a cat policy with clash risk. If OpenAI can’t pay the rent on the data center, it means there is a lot lower demand for Nvidia chips. It also means other similar arrangements by Meta, Google, etc. (it’s not just Nvidia providing free guarantees, they just happen to be the largest) will generate losses.

In other words, there is a giant clash event between a data center guarantee loss and a overall loss of income in their core business. It would be like if a cat reinsurer put their entire investment portfolio in cat bonds!

Systemic Risk

What this means is the end of the AI boom won’t be like 2000 where we saw a lot of smaller companies go under but for Microsoft, Intel, Cisco, etc. it was an earnings event but not a threat to their survival.

Because of the risk of large guarantee losses coinciding with a collapse in their core business, these gigantic companies could be in peril. If that happened, it would obviously have ramifications for the entire economy.

The tech industry isn’t as interconnected and as leveraged as large banks, so the damage hopefully wouldn’t be of the magnitude of 2008, but it would be very painful.

Don’t forget how much private credit has been involved in funding the AI buildout. They would have severe impairments.

Look at the stocks of “boring” companies like Caterpillar and Corning who supply data centers. There would be impacts throughout the economy, not just in Silicon Valley.

Speaking of stocks, a lot of people have invested with too much leverage and will face margin calls when markets collapse.

The AI boom has added a point to GDP growth. A collapse would take away far more than a point. It would be a guaranteed recession, worse than the one in 2000.

Listen To The Market

As I noted, Nvidia isn’t the only one doing these deals. Meta and Google have something like $50B each in total exposure. So on top of the clash event, there is also a correlation issue. The entire industry will see its capital impaired at the same time.

It’s worth noting the market isn’t totally ignoring the issue. Nvidia’s CDS (a proxy for their credit risk) has gone from 40bp to 80bp this week. That still is better than most of Corporate America but the change suggests debt investors are worried about the tail risk and, to the extent that Nvidia issues debt, it will be more costly going forward.

I’ll end with one last thought. If a new data center can only get built if Nvidia is willing to bet the company on its success, maybe that is a sign it shouldn’t be built!

6 thoughts on “Nvidia Enters Financial Guarantee Insurance”

  1. And no one ever mentions what the maintenance cap ex is on these data centers. How long do these chips last?

    1. Exactly. There is some serious negative convexity = either chips keep getting better in which case there is more demand but a lot of the current investment is wasted or demand flattens out and the data centers last a long time but the price for using them comes way down.

  2. this scenario highlights the fundamental problem that all the AI hyperscalers have, which is their costs vastly exceed their revenues. Its hard to be certain by how much, but I think 8-10x is the best case. No matter how impressive the models are (and LLM’s in my experience are the computer version of that person who sounds intelligent as they talk with such confidence, but when you actually examine what they say you realise their confidence is masking their stupidity), I’ve seen little evidence that any of them are close to showing that their models can deliver a ROI even remotely close to what they require to reak even. So I think your estimate of a collapse in demand for AI being a 1 in 10 event is under estimating the risk. Time to go short NVDA?

    1. Lol, yeah I’ve used that analogy before. When I ask it things I don’t know, it sounds impressive. When I ask it things I do know it sounds like a damn fool.

      It has improved on that score each year but you do have to be careful in rotely trusting them. My habit is to ask two models and see if I get consistent answers.

      1. when I am using it summarise reports and regulations etc I have taken to asking follow up questions such as ‘does the …. actually state that?’ I would say that about 50% of the time the response is something along the lines of ‘good spot, the …… doesn;t technically say X ………’ It’s failure rate is particularly high when its not just summarising the report but outlining what it means for me.
        Can the models get there? Perhaps!. Can they get there before the music stops? I’m not too confident

  3. The comparison with Ambac and MBIA moving from municipal debt into CDOs is a sharp one. A guarantee of that size without reserves or a priced premium does look a lot like the old monoline business.

Leave a Reply

Your email address will not be published. Required fields are marked *