In October, OpenAI secured a 4 billion dollar revolving credit facility from J.P. Morgan and several other banks. I was surprised when I heard this because OpenAI is a young company with no earnings. Shouldn't all their capital come from investors? Let's run some numbers.
From first principles
Let's do an Expected Value (EV) calculation, first from the perspective of an investor and then from the perspective of a lender. We'll pick some arbitrary parameters first, then refine.
Putting our investor hat on, the possible returns for investing $1,000 into OpenAI look like this:
Cost: $1,000
Case 1 (90%): OpenAI goes bankrupt. Return: $0
Case 2 (9%): OpenAI becomes a big successful company and goes 10x. Return: $10,000
Case 3 (1%): OpenAI becomes the big new thing and goes 100x. Return: $100,000
Our expected value is:
\[\begin{align} EV &= -1000 + 0.9 \times 0 + 0.09 \times 10000 + 0.01 \times 100000\\ EV &= -1000 + 0 + 900 + 1000\\ EV &= 900 \end{align}\]
The EV is positive, so this is a good investment. Obviously, there's a 90% chance of it going to zero, so if this were our only investment, it would be an insanely risky one. But provided we can do many investments like this and provided their failure cases aren't correlated, this would be a profitable strategy.
What happens if we instead put our lender hat on? Using the same probabilities as above, the possible returns for lending $1,000 to OpenAI at 5% interest look like this:
Cost: $1,000
Case 1 (90%): OpenAI goes bankrupt. Return: $0
Case 2 (9%): OpenAI becomes a big successful company and goes 10x. Return: $1,000 + 5% interest = $1,050
Case 3 (1%): OpenAI becomes the big new thing and goes 100x. Return: $1,000 + 5% interest = $1,050
Lenders don't benefit directly from the success of the company. Whether it barely scrapes by but manages to repay the loan or becomes the greatest company ever and easily repays the loan, it's all the same to a lender. So, we can merge cases 2 and 3 into:
Case 2+3 (10%): OpenAI doesn't go bankrupt. Return: $1,000 + 5% interest = $1,050
This makes our EV in the lending case:
\[\begin{align} EV &= -1000 + 0.9 \times 0 + 0.1 \times 1050\\ EV &= -1000 + 0 + 105\\ EV &= -895 \end{align}\]
The EV is negative, so we'd end up losing most of our money on average. Lending on these terms doesn't make sense.
There are two numbers we made up in the above calculation: the probability of bankruptcy and the interest rate. Let's leave the interest rate fixed at 5% and see what the probability \(p\) would have to be for us to break even.
\[\begin{align} EV &= -1000 + p \times 0 + (1 - p) \times 1050\\ EV &= -1000 + 1050 - p \times 1050 \\ EV &= 50 - p \times 1050 \\[0.5cm] & \text{Set EV to 0} \\[0.5cm] 0 &= 50 - p \times 1050 \\ p &= \frac{50}{1050} \\ p &= 0.0476 \end{align}\]
So, we'd break even if the probability of OpenAI going bankrupt was only about 5%. In practice, we'd want it to be lower than that so that we made a profit and so that we had a margin of safety in case our assumptions were wrong.
This 5% failure rate seems very optimistic to me, but this scenario is basically the one the consortium of banks got into. Concrete details on the deal are sparse, but this CIO Leaders article claims the interest rate was "SOFR + 100 basis points". The overnight SOFR rate is about 4.1% in October, so this puts OpenAI's interest at about 5%.
From market data
The problem with the above expected value calculation is that it's very idealized. The shape of it is correct, but the real world is too messy to be accurately represented by just a couple of parameters. I think it would be very difficult to build a model with enough predictive accuracy to be useful and I suspect there just isn't enough publicly available data to plug into it to make it work.
Luckily for us, banks exist! We know the banks have the better model and the non-public data and we know they came up with about 5% interest. So, let's work back from that and see what we can learn.
We're talking about a loan here and that's very similar to issuing bonds. So, we should be able to look at the bond market and find companies in similar financial health (from the perspective of a creditor). One problem is that we only know the overnight rate for OpenAI of about 5%, but bonds on the market will have longer maturities. We need to calculate what what yield a longer maturity loan would require and we can do that by looking at US treasuries.
According to Bloomberg, the three month treasuries have a yield of 3.94%. One year ones have a yield of 3.58%.
Figure 1. Treasury Yields for US government bonds. The table shows the following yields: 3 months is 3.94%, 6 month is 3.81%, 12 month is 3.58%, 2 year is 3.50%, 5 year is 3.62%, 10 year is 4.03%, 30 year is 4.62%. This describes a smile that initially goes down, goes back up to starting level at 10 years, then continues upwards.
One way of thinking about corporate bonds is that they're basically treasury bonds plus some premium to account for the risk of default. This default spread seems to be about \(5\% - 3.94\% \approx 1\%\) in OpenAI's case. By this logic, OpenAI's one year debt would have a yield of about 4.6%.
Can we find some one year bonds with a yield of 4.6%?
Figure 2. A sample of corporate USD bonds expiring in one year or less, sorted by their mid-yield to maturity. (Source: Saxo Bank)
Some bonds in the vicinity of what we're looking for are:
4.99%: HCA Inc. (US healthcare provider with credit rating BBB),
4.73%: Ziraat Katilim (Turkish bank with credit rating B+), and
4.24%: Citigroup (US bank with credit rating A).
In fact, scanning the sample above, it's mostly banks with BBB and A ratings. So, the consortium of big banks seems to have lent money to OpenAI at the kind of rates they themselves are borrowing at.
Looking at just a few bonds is interesting, but anecdotal. It would be better if we had some statistics across the whole bond market. Helpfully, Prof Damodaran goes through the exercise of calculating just such statistics (archive link) every year, most recently this January.
Figure 3. To quote the author: "This is a table that relates the interest coverage ratio of a firm to a 'synthetic' rating and a default spread that goes with that rating. The link between interest coverage ratios and ratings was developed by looking at all rated companies in the United States. The default spreads are obtained from traded bonds. Adding that number to a riskfree rate should yield the pre-tax cost of borrowing for a firm."
Looking up OpenAI's default spread of 1% in that table, we see it's at the level we'd expect for an A- or BBB firm (same as with the anecdotal search earlier). This normally corresponds to an interest coverage ratio of 3.00-4.24. However, OpenAI's actual interest coverage ratio is negative because their earnings before interest are negative.
This doesn't make sense: any way we look at it, OpenAI is getting the kind of interest rates only much more established and profitable firms would be getting. So, my initial surprise at hearing about this is justified, but there must be an explanation because the big banks wouldn't make such an obvious mistake.
Making this make sense
OpenAI is not a profitable company. It's also a private company, so we don't get to see audited financials, but we still know some things. This Reuters article claims OpenAI is going to generate $3.6 billion in revenue this year, but the costs will lead to a loss of more than $5 billion.
There's also speculation that their revenue next year will jump to $11.6 billion. However, there's no speculation about what their earnings will be because they're currently selling their services below cost and there isn't really any story as to how they'll turn this profitable.
The banks are lenders in this scenario, so they don't really care about how many users OpenAI gets or how huge their revenue becomes. As lenders, all they care about is getting paid back and it really doesn't seem like OpenAI will have the earnings to do that. But maybe earnings aren't what matters here.
If OpenAI can't pay its debts, it goes bankrupt and the creditors seize the company. Importantly, they seize it from the equity holders. Who are these equity holders? According to this Digital Information World article, the owners are Microsoft (28%), OpenAI non-profit and employees (52%), and other investors (20%).
So, the hypothetical is OpenAI runs out of money. They have revenue, but since their costs are higher, they don't actually have anything left over. They can't make interest payments on their debt, so they go bankrupt, and the banks seize the company from Microsoft. I don't think Microsoft will allow this to happen. Microsoft's earnings for last year were $88 billion, so I think Microsoft will just pay off OpenAI's $4 billion debt in this scenario. And I think the banks know all this.
So, the banks loaning money to OpenAI at an A- interest rate doesn't make sense, but effectively loaning the same to Microsoft with its AAA rating does, and that's what's actually happening here.