The major credit rating agencies are facing a credibility crunch point. They are increasingly being asked to rate borrowers intimately bound up with the dizzying artificial investment boom.
That includes a wide array of secured data centre financings, including those classified in the US into commercial mortgage-backed securities and asset-backed securities.
Some of these deals are getting triple-A ratings — such as a transaction in the market now from GI Partners’ vehicle TechCore, backed by a single redeveloped data centre in Illinois.
It also includes unsecured corporate bonds. In June SpaceX, the satellites-to-AI company controlled by Elon Musk, raised $86bn in its IPO and 10 days later came cap in hand to the bond market for $20bn.
Loss-making SpaceX was adorned with Baa1/BBB+/BBB ratings from Moody’s, S&P and Fitch. At least it has $94bn of cash to cover its $66bn of liabilities.
Anthropic and OpenAI — expected to float in mega-IPOs in October and next year, respectively — may get market capitalisations as stratospheric as SpaceX’s.
But although a high equity valuation gives huge comfort to lenders, you can’t use it to repay debt. You need free cashflow for that.
So far, there is not much of that about in AI land.
Anthropic and OpenAI are much more concentrated plays than SpaceX. Having invested massively in computing power, their hopes of profitability rest solely on being able to monetise their frontier models by persuading more customers to pay more for them.
Their revenues are growing very fast — in Anthropic’s case, to an annualised $65bn in July, up from $9bn at the end of last year.
Sceptics, however, doubt that breakeven is within reach, as compute costs are rising at roughly the same pace as revenue. Both companies have taken on large financial obligations, though these are not necessarily accounted for as classical debt.
Meanwhile, Chinese competitors such as Moonshot, DeepSeek and Z.AI are challenging the US firms’ technological leadership with cheaper alternatives.
Coveting respectability
Today, the Financial Times reported that Morgan Stanley and Goldman Sachs, two of the banks advising Anthropic and OpenAI most closely on their IPOs, were “lobbying” the rating agencies to give the companies investment grade ratings once they list.
The AI industry’s insatiable thirst for capital is putting the rating agencies in a tight spot.
For many observers, it is glaringly reminiscent of the years before the 2008 financial crisis, when, in the heat of a securitization gold rush, the agencies failed — innocently or culpably — to spot the dreadful credit quality of many US subprime mortgages and the fatal optimism of many precarious pieces of financial engineering.
That time, the crisis built up slowly. Securitization structures and ratings evolved gradually from well-tested forms. Mortgages’ safe track record meant no one raised the alarm when originators cut corners on underwriting.
New structures were graded highly because they made sense on paper, based on historic performance of their constituent parts.
As GlobalCapital observed in January 2007 of one ill-fated concoction, the constant proportion debt obligation, “Sceptics might argue that this is probably not the same kind of triple-A as the Swiss government”. The same might be said of data centre CMBS today.
GlobalCapital is not claiming to have been Cassandra: that 2007 article was otherwise glowing about CDOs — in line with the structured finance market at the time, which believed almost unanimously that it had discovered new ways to make gold.
This time is different. The investment craze is far huger, faster and more central to the economy. That means its supporters are even more powerful and vocal.
But it also means the risks are not hidden. Thousands of commentators have pointed out hundreds of cracks and dangers in AI’s Tower of Babel.
The rating agencies cannot plead ignorance, or that they are applying robust methods to incrementally more stretched transactions.
When it comes to rating new companies like Anthropic and OpenAI, with unprecedented business models, there will be huge pressure on the agencies to use new metrics and forms of analysis, more “appropriate” for these supposed titans than boring old free cashflow and leverage.
Against this, they will have to pit their traditional, human credit skills, based on long study of corporate credit under different conditions.
It will be interesting to see how they apply their environmental, social and governance standards to companies whose businesses rely on burning vast amounts of fossil fuels to put people out of work.
The agencies will have to make difficult decisions, in the face of demands for lenient treatment. They should remember: if they say no, no one will thank them. But if they say yes and get it wrong, everyone will curse them.