The number that travelled was $60 billion. The number that mattered was "double-digit yield."
Somewhere between those two figures sits the entire credit story of the AI buildout, and almost nobody reading the headline stopped at the second one. A double-digit yield is not a compliment. In fixed income, yield is a price for risk accepted, not a prize for a promising story. When a lender is offered 10%-plus on a secured claim against depreciating hardware, the lender is not saying "we believe in your future." The lender is saying "we have priced the probability that your future does not arrive, and we want to be paid for it."
Read the sentence again, carefully. Anthropic offers Blackstone a double-digit yield. Anthropic is the party raising capital. Blackstone is the party supplying it. This is not an equity cheque, not a strategic partnership, not a vote of confidence in a foundation model roadmap. It is a financing, structured around chips, priced at a level that in 2025 credit markets belongs to the unrated, cash-flow-negative, fast-depreciating end of the spectrum.
To understand why that pricing matters, you have to understand what is actually being financed. Not "compute." Chips.
That distinction is not semantic. A data center is immovable, long-lived, and financeable like real estate. A chip is movable, has a three-to-five-year economic life, and trades in a thin secondary market that depends entirely on which architecture wins the next cycle. Lenders treat those two assets very differently. The phrase "chip financing" tells you the collateral is the silicon itself — the part of the AI stack that depreciates fastest and prices worst in a forced sale.
The structure that makes this work is not new. In 2025, Meta's Hyperion data center was financed through a special purpose vehicle with preferred equity, anchored by Blue Owl at a scale north of $27 billion. Blackstone has been assembling AI infrastructure exposure of its own, including a reported ~$40 billion joint acquisition of Aligned Data Centers alongside MGX. Oracle has been borrowing against GPU fleets. CoreWeave built its capital structure on GPU-backed floating-rate debt. xAI priced high-yield paper at a 12.5% coupon. Anthropic's plan is simply this pattern arriving at the model layer, where the borrower owns no data center and has no investment-grade balance sheet to lean on.
That is the context the headline compressed into four words. Now the forensics.
Start with the pricing ladder, because in credit, price is the rating.
Investment-grade AI-adjacent issuers — Oracle, Meta — have printed debt at spreads near 100 basis points, all-in around 5%. GPU-collateralized data center operators with floating-rate facilities have been paying roughly 9%-11%. xAI's high-yield coupon came in near 12.5%. A "double-digit yield" therefore does not sit in the investment-grade band. It does not sit in the crossover band. It sits squarely in the unrated, negative-cash-flow, rapidly-depreciating-collateral bucket.
That is the credit file Anthropic's financing describes. Unlisted. No public rating. Large net loss. Revenue run-rate climbing fast, cash burn climbing alongside it. In the loan market, that profile does not get a 5% coupon. It gets a double-digit one, and the lender asks for covenants before it gets that.
Now run the arithmetic, because the arithmetic is where the narrative breaks.
If the double-digit rate applied to the full $60 billion, annual interest would land between $6 billion and $7.2 billion. Anthropic's 2025 revenue run-rate has been reported in the several-billion range, with definitions loose and figures disputed. Even at the optimistic end, a full-stack double-digit coupon would consume an amount of cash roughly equal to the company's entire top line. That is not a financing. That is a foreclosure with extra steps.
So the double-digit number almost certainly does not apply to the whole stack. The only internally consistent reading is a tranched structure: senior debt in the mid-to-high single digits, mezzanine at 9%-11%, and an equity or residual tranche targeting 15%+ IRR. Blackstone taking "double-digit" points to the junior layers. The headline took the highest number in the capital stack and presented it as the cost of the whole.
That single inference changes every downstream conclusion. The real question is not "how expensive is Anthropic's capital" but "how thick is the subordination, and who is standing behind it." That question is not answered anywhere in the reporting.
Then there is the duration problem, which is the one I would flag first in a credit committee.
AI accelerators have an effective economic life of roughly three to five years — shorter if the inference-optimized silicon curve steepens. Infrastructure debt typically runs seven to fifteen years. If the debt outlives the asset, the borrower keeps paying rent on hardware already written down to scrap. This is a classic equipment-finance mismatch, amplified here because the underlying technology is iterating faster than any asset class ever securitized at this scale. The lender knows this. It is precisely why the lender wants double digits.
And the collateral is not homogeneous. A general-purpose GPU has a functioning secondary market, however illiquid. A custom ASIC — a TPU, a Broadcom design — has almost no resale value outside its originating ecosystem. If the collateral mix skews toward ASICs, the recovery assumption in the lender's model is not "residual value," it is "scrap value plus a call option on the sponsor's goodwill." The sponsor in that scenario is also the supplier, also an investor, and also a competitor. That is not a normal credit relationship.
Which brings me to the part that should concern anyone holding this risk indirectly: the residual value risk is being transferred, not eliminated. In a high-yield equipment financing, the lender absorbs the technology-obsolescence loss. That is what the coupon buys. So when the AI silicon cycle turns — and it will turn, because every silicon cycle turns — the loss lands on the credit investor, not on the model company. If those credit investors are insurance balance sheets and pension allocators reached through private credit funds, the loss lands on savers who never chose AI exposure.
This is where my own work becomes relevant, and I want to be precise about the limits of the analogy.
I spent 2017 auditing ICO tokenomics equations by hand and found that two of the top ten had inflation schedules that made their own whitepapers mathematically impossible. In 2020 I mapped Uniswap V2 liquidity depth and found oracle-manipulation arbitrage that three funds later paid my firm to keep mapping. In 2022 I modelled algorithmic stablecoin contagion and published the arithmetic of the Terra collapse before it finished collapsing. In 2026 I led a project running AI models over ten million on-chain transactions to detect wash trading, and we found bot networks inflating volume by roughly 15% on specific DEXs.
The through-line across all four is identical. When an asset's price depends on a number nobody independently verifies, the number eventually becomes fiction. AI compute financing has exactly that property right now. The "value" of a chip fleet is whatever the financing model says the residual is. There is no third-party mark. There is no liquid index. There is no attestation of utilization. The collateral is priced by the party that benefits from a high price.
Compare that to what on-chain credit actually gives you. Tokenized Treasury products report daily NAV against a benchmark anyone can check. On-chain lending markets publish utilization, collateral factors, and liquidation thresholds in real time. You can watch a position move from 150% collateralized to 110% and know exactly when the auction triggers. That is not because DeFi is safer — it frequently is not — but because the ledger is public and the narrative is not load-bearing. Ledgers do not lie, only the narrative does.
The uncomfortable observation for my own industry is that the same private credit complex now financing AI chips is also a dominant issuer in tokenized credit. The funds buying AI infrastructure paper are, in several cases, the same vehicles whose yield products get wrapped and sold on-chain as "real-world yield." Which means the AI capex cycle is quietly becoming a component of on-chain RWA yield — and almost nobody buying that yield has mapped the underlying collateral to a depreciating ASIC.
Here is the angle the coverage missed entirely, and it is not the one you expect.
The consensus read is that double-digit pricing signals confidence in Anthropic. The contrarian read is that the equity market and the credit market are pricing two different companies. Equity reportedly values Anthropic in the $350 billion range. Credit is charging double digits. Those two numbers cannot both be right about the same capital structure unless the equity holder believes in a terminal state the lender does not need to believe in. When equity and credit disagree about the same issuer, credit has historically been correct earlier — not because lenders are smarter, but because lenders get paid first and therefore have to be right about the downside, while equity only has to be right about the upside.
The second blind spot is narrative transfer. This story has nothing to do with blockchain. Not one component of it touches a chain. Yet it surfaced in crypto media because AI infrastructure has become the universal narrative magnet — any story involving large numbers, structured yield, and scarcity gets pulled into the crypto feed regardless of whether a single byte touches a ledger. That is a signal about attention, not technology, and it is worth noticing how easily a private credit transaction is reframed as a crypto story simply because the yield is high.
The third blind spot is my own industry's RWA thesis, and I will say it plainly. Tokenizing this credit does not democratize it. It socializes the residual risk while concentrating the fees. The three-year RWA storytelling exercise has largely been traditional institutions discovering they do not need a public chain to issue a note — and this transaction proves it. Blackstone did not need a token to move $60 billion of chip exposure. It needed a legal SPV, a depreciation schedule, and a covenant package. The chain was never the constraint.
The number to watch is not $60 billion. It is the tranche disclosure that will never be filed, because Anthropic is unlisted and owes nobody a capital stack diagram.
Watch three things instead. First, whether any of this paper appears inside tokenized private credit vehicles within twelve months — if it does, the yield will look attractive and the collateral note will be one line long. Second, whether a third-party residual value index for AI accelerators emerges; its absence is the single largest unhedged assumption in the entire structure. Third, whether the next model company prices its compute financing above or below this one. If the next coupon comes in lower, the market learned something. If it comes in higher, the market is repricing the entire asset class.
Resilience is built in the red, not the green — and nobody has yet seen this structure in the red. Trust the math, ignore the hype. The math here says a double-digit yield is a confession, not a compliment.