The press release ran to five sentences. It named a rating agency that most equity investors have never quoted in an investment memo, assigned a grade composed of two letters and a symbol, and used the word "transparency" twice. It did not name a bond size. It did not name a coupon, a maturity, a rating outlook, or a single line item of use of proceeds.
And yet, when it crossed my desk, I read it three times. Not for what it said, but for the shape of the hole in the middle of it.
Anthropic — a company roughly four years old, still unprofitable, and consuming compute at a rate that would make a telecom CFO flinch — had obtained a BB+ credit rating from KBRA. One notch below investment grade. The financial wires treated it as a curiosity, a footnote beneath the model releases and the funding rounds. I treated it as the moment the AI industry stopped asking venture capitalists for permission.
Read the docs. Question the whisper. The loudest part of that release was the silence around the number.
To understand why a credit rating matters more than another funding round, you have to hold three eras of AI capital in your head at once, because they are layered on top of each other and most people are still arguing about the first one.
The first era, roughly 2019 through 2023, was pure equity. Capital arrived as SAFEs, priced rounds, and strategic checks from cloud providers who understood something the venture funds did not: that the investment was really a forward contract on compute consumption. Nobody asked for a rating because nobody needed one. Dilution was the only cost of capital anyone modeled.
The second era, 2024 into 2025, moved down the stack. CoreWeave built a business on GPU-backed debt facilities and then listed. Oracle issued debt against compute contracts. Meta raised tens of billions in corporate bonds to build data centers. The infrastructure layer discovered that a GPU cluster is a fixed asset with residual value, and fixed assets have been financed with debt since the first railroad. This was not exotic. It was the most ordinary thing in finance.
The third era is the one that just opened, and it is the one that should worry you. The model layer itself is now being rated. Not the data center holding company, not the power provider, not the real estate trust. The lab. The entity whose primary assets are researchers, weights, and a training roadmap.
There is a reason this did not happen earlier, and it is not that the labs were shy. Equity was abundant and cheap. When equity is cheap, debt looks like a constraint. When equity becomes expensive — and for frontier labs it has been getting more expensive in real terms since 2024 — debt starts to look like freedom.
I have watched this movie before, in a different theater. In 2021 and 2022 the crypto lending market ran exactly this experiment on far worse collateral, and we all know how it ended. Celsius, BlockFi, Genesis. The lesson was never that leverage is evil. The lesson was that nobody had tested the collateral. Everyone had tested the narrative.
Start with the mechanical fact that almost every commentary on this story skipped: BB+ is not "almost investment grade" in the way a B+ exam is almost an A. It sits one notch below BBB-, which means it lives on the other side of a cliff. Investment grade is not a description; it is a market segment with its own buyers, its own mandates, and its own index inclusion rules. A bond rated BB+ is invisible to a large fraction of the world's fixed income capital by charter, not by judgment.
The spread differential between the top of high yield and the bottom of investment grade has historically run between 200 and 400 basis points depending on the cycle. Run the arithmetic. A one billion dollar seven-year issue priced at roughly 7.5 percent instead of 4.75 percent costs about 27.5 million dollars a year in additional interest. Over the life of the bond, before fees and covenants, that is somewhere north of 190 million dollars paid for the privilege of being one notch away from respectable.
That number is not abstract. It is approximately the annual cost of a serious safety and interpretability research organization. The price of the rating is the research that would justify a better one. I have no evidence that anyone at Anthropic framed the trade that way, and I would be surprised if they did. Capital structure decisions are rarely made in the same room as research budget decisions, which is precisely why the second-order effects are so often missed.
The choice of agency is the second thing worth noticing. Anthropic did not go to S&P, Moody's, or Fitch. It went to KBRA, the fourth agency, best known for structured finance and municipal credit. There are legitimate reasons for this — faster timelines, more accessible analysts, a willingness to engage with unusual capital structures. There are also less flattering reasons, and the honest answer is that both are usually true. When a company picks the agency rather than the agency picking the company, the market should read speed as urgency.
What a rating buys is scale. Global fixed income is measured in the hundreds of trillions; the high yield segment alone is measured in the low trillions. That is a different order of magnitude from the venture pool, which has been contracting in real terms and concentrating into fewer, larger checks. For a company at a valuation in the hundreds of billions, continuing to raise equity is not merely dilutive — it is structurally awkward. The list of institutions capable of writing a check that size is short, and every name on it now wants governance concessions.
Debt does not ask for a board seat. That is the entire appeal, and it is also the entire trap.
What the rating takes is subtler than interest expense. It takes a calendar. A rated issuer lives on someone else's reporting rhythm: quarterly disclosure, covenant tests, rating agency meetings where you explain your own business to people who will never run a training job. In my years on the governance side — I spent much of 2020 organizing two hundred small MakerDAO holders to block a collateral expansion I believed was underpriced — I learned that risk almost never hides in the code. It hides in the meeting where nobody wants to be the person who says the number is wrong.
A covenant package is that meeting, made contractual and permanent. It is a board member who never speaks at the offsite but can veto the budget in writing. And unlike a token holder, a covenant does not care about your mission statement.
There is a pricing consequence that equity holders have not internalized. Once a company carries a public rating, it acquires a second valuation lens running on entirely different inputs. Equity values the terminal state; credit values the path. A company can be worth more on the growth story and simultaneously be a deteriorating credit, and the two views do not reconcile — they simply coexist, and the credit view constrains the growth view over time, because it governs access to the capital that funds the growth.
For the token market, this is a more useful lesson than any specific number. Most crypto projects have spent a decade optimizing for a single audience: the marginal buyer of a narrative. Debt introduces a second audience with veto power and no interest in the narrative at all. Projects that cannot survive being read by that second audience should not be borrowing.
Here is where the structure gets genuinely interesting, and where I think the credit analysts earned their fee.
Amazon has invested roughly eight billion dollars into Anthropic. Google has invested roughly three billion. Both are also the primary suppliers of the compute that Anthropic burns to train and serve its models. Both are, in various configurations, customers or distribution partners as well. Investor, supplier, and counterparty, in the same entity, on the same cap table.
Rating agencies have a term for this: related-party concentration. It shows up in every structured finance model as a haircut, because the analyst's job is to ask what happens to the cash flows if the relationship changes. The uncomfortable question is not whether Amazon and Google are creditworthy. They obviously are. The uncomfortable question is whether Anthropic's revenue growth and its cost base are independent variables, or whether they are the same variable viewed from two directions.
Money raised in the bond market, paid to a shareholder as cloud revenue, which supports the revenue figure used to justify the bond. That is a loop, and loops are only visible when you draw them. I say this with some feeling. In the winter of 2022 I spent three months running free counseling sessions in Rome for a hundred and fifty retail investors who had lost money in the FTX collapse, helping them work through tax treatment and the grim mechanics of claims. Almost none of them had been fooled by a bad model. They had been fooled by a structure nobody had drawn for them. The equity story and the revenue story were the same counterparty, and the diagram fit on a napkin if anyone had bothered to make it.
So here is my test for any rated structure, AI or otherwise: can you draw the money flow on one page without lifting your pen? If you cannot, the rating is a description of the past dressed as a forecast.
Step back from Anthropic specifically and look at why the debt exists at all. It exists because the cost of staying at the frontier has become a capital expenditure rather than a research expense. A training cluster is not a salary. It is a building full of depreciating metal that produces a measurable stream of inference revenue, which is exactly the profile lenders have financed for a century.
This is why CoreWeave could pledge GPUs and borrow billions. It is why Oracle could issue against compute contracts. And it is why the model labs are following. The asset is real, it has resale value, and it generates cash.
Except for one assumption, buried in a footnote, that nobody outside a credit committee ever reads: useful life.
Most GPU depreciation schedules assume somewhere between five and six years. That assumption is defensible for inference workloads running stable models on stable hardware. It becomes much harder to defend when the same silicon is asked to run frontier training jobs that push memory bandwidth and interconnect to their thermal limits for months at a time. If the economically useful life under those conditions is closer to three years, the collateral is overstated by a very large fraction, and every ratio built on top of it inherits the error.
Nobody is lying. That is what makes it dangerous. The 2022 crypto credit crisis was not primarily a fraud story; it was an optimistic collateral curve that met a market that repriced faster than the curve. Rounded assumptions, compounded across a hundred balance sheets, become a systemic event with no single villain. The audit begins where the press release ends, and in this case the press release ended before the most important number was printed.
By the time this column publishes, someone will have written that Anthropic's rating is bullish for tokenized private credit. They will be partly right, for the wrong reason.
The right reason is that credit is where the next wave of real-world asset tokenization actually has teeth. Tokenized treasuries were a warm-up exercise: clean collateral, obvious marks, no credit judgment required. AI infrastructure debt is the opposite. It requires someone to underwrite a depreciation curve, a counterparty relationship, and a technology risk simultaneously. If that paper starts appearing in on-chain credit pools — and the demand for yield in a bull market all but guarantees it will — then the protocols holding it are no longer intermediaries. They are credit analysts, whether or not they have hired any.
The alpha is not in the wrapper. The alpha is in the collateral, and the collateral is a five-year assumption about a three-year asset. Anyone allocating to an on-chain pool that holds AI-adjacent paper should be able to answer, in one sentence, what the underlying hardware is worth in year four.
There is a second, quieter read-through. Decentralized compute networks — the DePIN GPU marketplaces that spent two years being dismissed as a narrative with no customers — are now, functionally, price discovery venues for the exact asset class being pledged as collateral. Their clearing prices are no longer a curiosity for token traders. They are an input to credit models. That is an enormous and underappreciated upgrade in their relevance, and almost nobody is pricing it.
And a third: the same wall is coming for DAO treasuries. A diversified stablecoin treasury is equity. The moment a protocol borrows against projected fee revenue to fund an acquisition or a liquidity program, it has issued debt, and it inherits every discipline that implies. Most DAOs have no covenants, no disclosure cadence, and no one whose job is to say no. They will learn this the way everyone learns it, which is expensively. Last year, designing the human-in-the-loop consensus framework for an AI-crypto protocol, I watched fifty developers and sociologists argue for three weeks about how an autonomous agent should behave. Not one of them asked who would underwrite the agent if it started borrowing. That question is now on the table.
The consensus reading of this event is already forming, and it is tidy: AI is 2000-era telecom, the debt is the fiber, and the crash is scheduled. I think that framing is comfortable and mostly wrong, not because the debt is safe, but because it locates the risk in the wrong place.
The first thing the consensus misses is that debt does not primarily add risk. It adds a clock. The interesting question is not what the default rate will be. It is what stops getting funded once the clock starts. Interpretability research has no revenue line. Long-horizon alignment work has no revenue line. A covenant package will never explicitly prohibit either, because no analyst writes a clause about research philosophy. But budgets are zero-sum inside a company the way attention is zero-sum inside a person, and a quarterly reporting cadence changes what a leadership team celebrates.
The measurable risk here is not insolvency. It is a slow reallocation of internal capital away from the work that made the brand trustworthy in the first place.
The second thing the consensus misses is the missing outlook. A rating without a published outlook is a level without a vector, and markets are extremely good at pricing levels and extremely bad at pricing vectors. The absence of a stated stable, positive, or negative view is itself information: it means the agency sees the trajectory as unresolved, which for a company this young and this unprofitable is the most honest thing in the whole affair.
And the third: the rating may pay for itself somewhere nobody is looking. Banks and governments buy AI from vendors they believe will exist in five years. A BB+ from a recognized agency is a procurement asset before it is a financing instrument. The marketing value sits in the sales conversation, not the bond covenant.
So watch the follow-on. Which lab gets rated next, and at what level. And when the first downgrade inside the AI complex arrives, mark my words, it will not be triggered by a model that failed a benchmark. It will be triggered by an analyst revising a depreciation schedule.
The question I would put to every allocator this quarter is not whether Anthropic can service its debt. It is this: which on-chain credit pool is holding AI-adjacent paper tonight, and does anyone in that room actually know what the collateral is worth in year four?
Alpha hides in the silence of the audit.