Downgrades

10-year Treasury yield

Recently, interest rates for government debt have been rising worldwide. As I write the yield of the 10-year US treasury is over 5%. It had been below 4.5% for over a year when, in March, it started a steady rise. This is despite (and probably because of) Scott Bessent's efforts to force it lower, which JP Morgan compared to 'paying your mortgage with your credit card'.

Source

A big part of the reason is governments persistently borrowing to cover deficits, with no plan to reduce them. For example, the Trump administration has borrowed more than the growth in US GDP. Projections are that by the end of their term they will have borrowed $1.40 for every $1 growth in GDP. Half of that GDP growth is from the AI bubble. Remove it and the administration has already borrowed almost 2.5 times the GDP growth.

But another part of the reason is that governments are facing competition for the available funds from other highly-rated borrowers. Hyperscalers such as Microsoft (AAA) and Alphabet (AA+) can no longer pay cash for the immense sums they believe they need in order to keep up with their competitors in the race for AGI. So the bond market is facing a significant increase in demand, which naturally increases interest rates.

Below the fold I look into how this competition for funds is going.

It is a fact of life that the more debt you carry, the worse your credit rating and thus the more expensive it is to add to your debt. This is why the hyperscalers are so anxious to have their massive borrowings kept off their balance sheets and attributed instead to "Special Purpose Vehicles". They hope in this way to distract the rating agencies enough to prevent being downgraded. A downgrade would increase their interest expenditures, thus reduce the coverage of their interest payments, and potentially lead to a further downgrade and a death spiral.

This is also the reason Michelle Chan et al report that Anthropic and OpenAI bankers push for top-tier credit ratings post-IPO:

Achieving an investment-grade rating from Fitch, Moody’s and S&P soon after going public would be a remarkable feat for the two lossmaking AI labs, unlocking big benefits for the companies and their infrastructure partners including Oracle and Nvidia. …

Analysts at rating agencies are waiting to see the results of their IPOs before reaching a decision. The two companies remain unprofitable and have shown little sign of generating positive free cash flow. They also face growing risks, including the popularity of Chinese open-weight models.

“We still treat OpenAI and Anthropic as deep in speculative grade . . . they are in the red,” said another senior credit analyst.

An investment-grade rating of BBB- or above would allow them to borrow from institutional lenders such as pension funds.

Now Toby Nangle violates Betteridge's Law of Headlines with Are credit rating agencies getting fed up with hyperscalers?:

S&P reckons a full half of the economic growth coming from the US private sector was linked to AI-centric activity over the past year. And they assume the top six US hyperscalers will collectively spend more than $7tn in the five years through 2030. So the fate of the US economy, not to mention the stock market, credit, infrastructure and real estate, is increasingly tied to the AI show staying on the road.

Nangle quotes S&P:

Every time we take a deep dive into this sector, we find that capex is rising faster than we anticipated, financings are becoming more complicated and less transparent, and that returns on investment will take years to realize.

Why would the rating agencies be concerned?

The credit rating process involves analysts diving deep into the workings of companies, typically with access to a ton of non-public information. But the authors write that, despite AI return on investment being critical to the rating judgments, “the big six hyperscalers don’t quantify their returns on investment on AI”. Maybe, the report’s authors speculate, this is because it’s difficult to work out. Or maybe, “simply, entities just choose not to share that data.”

The reason these companies "choose not to share" their "return on investment on AI" with the rating agencies is left as an exercise for the reader.

Source

Nangle has a set of interesting charts based on S&P data. This one shows the ratio of debt and "debtlike commitments" to EBITDA (earnings before interest, taxes, depreciation, and amortization) for the six largest US hyperscalers. Each company is labeled with their current rating. Horizontal dotted lines indicate the trigger level for this ratio that could cause a downgrade. Note that:

• Oracle's projected 2027 and 2028 commitments are so close to the trigger for downgrading from their BBB- rating that even a small drop in earnings would render their bonds junk. This would force institutional investors to sell them, making future borrowing and rolling over existing debt as it comes due much more expensive.
• Amazon's projected 2027 and 2028 commitments look as though they would cause a downgrade from AA, but it is more likely that Amazon can increase earnings fast enough to prevent this.
• Alphabet has a lot of headroom before their AA+ rating is at risk.

Source

An alternative way of looking at the situation is to ask how much more could these companies borrow, or how could their EBITDA decrease, each year before they would be downgraded. This chart answers the question. Note that:

• For 2027 through 2029, only Alphabet can borrow more than $100B each year.
• On the current projections Amazon will, and Oracle is extremely likely to, be downgraded. Nangle makes an English understatement:
  

  For double-A-rated Amazon to move to high single-A wouldn’t be much of a disaster. But in Oracle’s case, getting downgraded means getting junked. And with around $117bn of index-eligible US dollar bonds, that would be quite the event.

• The limited headroom for increased borrowing starts next year, which is when the "take or pay" commitments with data centers start costing serious money.

Source

The third chart provides yet another way to look at the situation. It plots the companies' projections for EBITDA against the threshhold for downgrades (dotted line). Note that:

• Amazon has to exceed and Oracle has to meet their EBITDA projections if they are not to be downgraded.
• All these companies' projections for their EBITDA growth from 2025 to 2029 are astonishing. Eyeballing the chart we have:
  
  • Alphabet: 2.5x
  • Amazon: 2x
  • Meta: 2.6x
  • Microsoft: 1.8x
  • Oracle: 3.5x
  • SpaceX: 15x
  Nangle's comment that these represent "pretty punchy EBITDA growth" is another understatement. For example, Alphabet is saying it will add around $225B in EBITDA by 2029. I'm an Englishman, so I can say that it isn't clear where an additional $225B would come from.
• Alphabet and Meta together are in effect projecting that by 2029 companies will increase their advertising budgets by $385B/year. Of course, this isn't going to happen. All these companies are expecting the EBITDA growth to come from the "returns on investment on AI" that they "choose not to share" with the rating agencies.
• Ignoring the ludicrous SpaceX projection, the other 5 companies' "punchy" projections assume 2029 increased EBITDA of around $900B, which is less than half the estimate of over $2T/year in additional revenue needed to cover their borrowings.

Is it plausible that in 2029 OpenAI and Anthropic will together have more than $1.1T in revenue to make up the difference?

Then there are the additional risks:

from circular financing, through to risk of overbuilding. We’ve included a link to Joachim Klement’s Substack, in which he discusses the risk from open models just to round things out.

The report also features a circular financing taxonomy, which includes residual-value guarantees, take-or-pay agreements, chip financing, lease liabilities, power purchase agreements, backstop guarantees, lease guarantees, and direct equity investments.

Nangle quotes S&P:

interconnected financing structures could amplify volatility if demand weakens unexpectedly. The failure of one entity would have implications for entities that have lent it money, have guaranteed the value of its assets, or are just expecting payment for goods delivered.

The broader question is whether circular financing is creating leverage collectively that is individually manageable but could become highly correlated should demand fall sharply. . . . 

The scale of overlap and interconnectedness is vast. In a downturn, even the best capitalized and most profitable firms may incur substantial pain.

Source

Ed Zitron's Concentration Risk points out another big risk to the hyperscalers' EBITDA projections that S&P seems to have missed:

Anthropic and OpenAI’s compute commitments, in my mind, should be seen more as debt obligations than “contracts,” because they (as take-or-pay agreements) function in much the same way, requiring the company to pay whether or not they need the capacity.

For now, everything looks awesome. Microsoft, Google and Amazon have all had big bumps in revenue from AI lab compute spend along with massive, ever-swelling revenue backlogs — over $1.5 trillion worth to be specific. More than half of that backlog is attributable to Anthropic and OpenAI, which, as I’ll say again and again, isn’t a problem because the money is yet to stop coming in.

Eyeballing the chart, something like $1.1T of the revenue backlog for Microsoft, Oracle, Google and Amazon is revenue that they expect to get from two companies that have yet to make a profit. That is a concentration risk.

Source

But the two companies are themselves exposed to a concentration risk:

Per data from fintech firm Ramp, 80% of OpenAI and Anthropic's enterprise revenues come from 1% of their customers, a number that hasn’t improved over the last three years. Ramp’s lead economist Ara Kharazian notes that the top 1% skews heavily toward the tech sector and AI products and services, and that this was a level of concentration risk unseen in any other software category they tracked.

As I pointed out in Small Is Beautiful, and Toby Nangle does in How big is the open-model threat to AI hyperscalers?, these 1% of the customers who provide 80% of the revenue that is supposed to deliver $1.1T to the four hyperscalers now have the means, motive and opportunity to run open-weight models on in-house infrastructure.

Cris Tolomia's Thomson Reuters built its own AI model on Chinese open-source tech to slash AI costs is an example of this trend. Jeremy Hsu quotes data from Mozilla in Paying for frontier AI models buys 4-month head start at 5x the cost:

benchmarking company Vals AI has been evaluating different open and closed models by using its own neutral harness. When every model ran on the same harness in the Terminal-Bench 2.1 evaluation, the open-weights model GLM 5.2 from Chinese company Z.ai (Zhipu AI) scored within a point of Anthropic’s Claude Opus 4.7 and 4.8 while costing about five times less per completed task.

In other words, paying for closed frontier models buys about a four-month head start at about five times the per-task cost—but only when currently looking at tasks taking [a human] between eight and 12 hours.

Below 8 hours there is no advantage for closed-source, above 12 hours neither model can reliably succeed. This isn't a great basis for OpenAI and Anthropic to project hundreds of billions of dollars of future revenue.

Zitron notes that these two unprofitable companies haven't yet had to pay for their commitments:

On the low end, that means that Anthropic and OpenAI account for over $200 billion dollars worth of expected revenues for Microsoft, Google and Amazon in 2027, which is contingent on their ability to raise venture capital or debt, which is contingent on the continued growth of their businesses, which is contingent on growing AI spend from a small subset of customers, many of whom are funded by venture capital.

The reason this hasn’t been a problem yet is that when you sign these contracts, you tend to pay a small up front fee, and the capacity in question is yet to come online.
...
In other words, Anthropic and OpenAI are currently in the teaser rate period where all of that capacity — and all of the associated costs — are yet to hit.

Next year, at least $200 billion in compute costs are coming due.

The question is whether Anthropic and OpenAI, two unprofitable, unsustainable AI labs that lose tens of billions of dollars a year, will be able to afford to pay them.

If they can't pay, about half of the revenue backlog for Microsoft, Oracle, Google and Amazon goes away. That reduces their EBITDA and potentially triggers downgrades.

But the agencies' ratings are just input to the market for bonds; it is the market that sets the actual inerest rates. Nangle followed up with How much should lenders charge hyperscalers?. He takes off from a report by Ludovic Subran et al from Allianz entitled The (A)Iceberg beneath tech debt: Recognized calm, rising spreads below the waterline.

Source

Nangle starts with this chart, comparing the yields on the hyperscalers bonds with the average for similarly rated companies. It shows that:

it looks like the market is pricing Oracle and SpaceX as quality junk, and maybe Meta in line with BBB credit, and the rest of them as sort-of-in-line with the average AAA-AA-rated US corporate issuer. Is this cautious enough?

On recognized debt, the market’s calm is justified . . . [but] . . . [t]he real risk lurks below the waterline: off-balance-sheet debt lifts the debt burden by nearly 150% on average and pulls the model-implied credit quality down 1-2 notches.

There follows a complex description of Allianz' model that maps from stock prices to bond ratings, summarized thus:

Allianz calculates the ‘distance to default’ for the hyperscalers — a measure of credit riskiness calculated using equity inputs and balance sheet data — over time. And following work that Moody’s KMV has done mapping distance-to-default to default frequencies, they map the kind of ratings associated with these outputs.

Source

Nangle compiles this table of the results:

Working solely off recognised balance sheet data, Allianz finds that the ratings implied by equity punters’ collective histrionics are higher than ratings assigned by the agencies for Alphabet, Amazon, Microsoft and Nvidia, a touch lower for Meta and a disaster for Oracle and SpaceX.

I think that's not quite right. Moody's gives Microsoft their highest rating, Aaa, and so does Allianz' model. The model gives Alphabet, Amazon and Nvidia Aaa, which is higher than Aa2, A1 and Aa1 respectively. But overall equity investors are (naturally) more optimistic than bond investors.

But these ratings are a testimony to the effetivness of the hyperscalers' off-balance-sheet techniques for distracting both the equity markets and the rating agencies. Nangle writes:

Once they whisk debtlike uncommenced leases into the equation, they find that Microsoft and Amazon drop into mid-investment-grade territory and that Meta falls into the quality end of junk.

Nangle's first chart showed the market already rating Meta at "the quality end of junk". But once the bond market assimilates the off-balance-sheet debts the downgrades and interest rate increases will be quite dramatic.