On a Tuesday morning in Singapore, the chief executive of the country's largest bank used eleven words that carried more information than any AI earnings call this quarter. Tan Su Shan said the market is repricing AI financing risk. That was the whole signal. No spread levels. No default rates. No counterparty names. Three information points, one source, zero numbers.
I have spent eighteen years reading disclosures for a living. When a systemically important bank's CEO volunteers a sentence like that, you do not file it under commentary. You file it under mark. Banks do not forecast credit events in public. They describe the ones already moving through the tape.
The interesting question is not whether the repricing is real. It is where the repriced risk actually sits. And that answer is uncomfortable, because a meaningful slice of AI infrastructure debt has spent the last two years migrating onto crypto rails β tokenized, wrapped, and posted as collateral in DeFi pools β while the pricing models behind those wrappers still update on a weekly cadence.
Ledgers do not lie, only their auditors do. The on-chain ledger says the collateral is worth par. The off-chain loan book says otherwise. Both statements are being made at the same time, by the same counterparties, and only one of them is being marked.
Understand the structure before you judge the signal. The AI capital boom is not financed by equity. It is financed by debt, structured through special purpose vehicles, and increasingly intermediated by a small number of large banks and private credit funds. A hyperscaler signs a long-term compute contract. A developer builds the datacenter against that contract. The debt sits in an SPV. The SPV note is sold to a private credit fund. The fund holds it to maturity and marks it quarterly.
That structure worked while compute demand rose monotonically and rates fell. It works less well when the marginal return on frontier training flattens and a single training run costs hundreds of millions. The technical root of the repricing is arithmetic, not sentiment: capability gains per dollar of compute are decelerating while capital intensity per capability unit keeps climbing. When the input-output ratio stops improving, the debt that financed the inputs gets repriced.
Into that structure walks a second layer most credit analysts do not model. Over 2024 and 2025, the same asset class began appearing on-chain. Tokenized treasury products came first β short duration, clean, fine. Then the riskier end followed. Tokenized private credit. Yield-bearing RWA vaults. On-chain SPV notes promising 9 to 14 percent. DeFi lending markets began accepting these instruments as collateral, and the pitch was always identical: real-world yield, on-chain liquidity, institutional grade.
I have audited that pitch before. In 2017, at twenty-five, I spent three months manually tracing the ERC-20 transfer logic of a token offering that had raised $15 million, and I found an integer overflow in the vesting contract that would have cost 12 percent of the fund's assets. I cited line numbers in the bytecode. In 2026, I spent three months on a decentralized compute network that promised 60 percent GPU cost reduction through a novel sharding algorithm, and found the consensus layer added 40 percent to finality time β twelve critical inefficiencies, one formal report, project rejected. The lesson from both engagements is the same, and it applies here directly: a token wrapper does not change the duration, the credit quality, or the liquidity of the asset underneath it. It only changes who can redeem and how fast the price feed updates.
Now the mechanics, because the mechanics are the whole story.
A datacenter SPV note has no continuous market. There is no order book, no last trade, no depth. Its value is estimated by the issuer or a third-party valuation agent, typically monthly or quarterly. When that note is wrapped into a token and posted as collateral, the wrapper needs a price. That price comes from an oracle. The oracle has nothing better to point at than the same stale, issuer-reported NAV.
So the on-chain price of the collateral is, by construction, a lagged copy of an estimate. This is not a bug in any single protocol. It is the only honest way to price an illiquid asset on a 24/7 ledger. But it produces a specific failure mode: when the off-chain credit market reprices, the on-chain collateral does not move until the next NAV print. In the interval, the protocol is lending against a number that no longer reflects reality, and the borrower is free to draw against it.
Add the second variable: redemption. Most of these RWA vaults offer daily liquidity or a short stated window. The underlying asset has a five-to-seven-year duration and effectively no secondary market. That is textbook liquidity transformation β the same maturity mismatch that broke money market funds in 2008 and broke on-chain lenders in 2022. The wrapper did not invent new risk. It moved an old risk onto a faster clock, which is worse, because fast clocks amplify runs.
Here is where the DBS signal becomes operational. If the market is repricing AI financing risk, the discount rate applied to datacenter cash flows is rising. A higher discount rate lowers the present value of the SPV note. That repricing happens in the loan book first β in the private credit fund's quarterly mark. The token wrapper does not see it until the NAV feed refreshes. If the NAV feed is issuer-controlled, the refresh is discretionary. The lag is not a technicality. The lag is the product.
Yield is the interest paid for ignorance. The 12 percent on that tokenized credit vault was never free. It was compensation for duration, illiquidity, and valuation opacity β three risks that the on-chain wrapper made easier to sell and harder to see.
There is a third layer, and it is the one that worries me most. Circular financing. Through 2024 and 2025, a large share of AI demand was manufactured internally: a chip vendor invests in a customer, the customer commits to buy chips, the vendor books revenue against the commitment, and the revenue supports the valuation that justifies the next investment. Vendor financing is an old practice. In a high-rate, repricing environment it becomes a valuation problem, because revenue quality is only as good as the counterparty's ability to keep paying β and the counterparty is financed by the same credit market that is now tightening. When the loop runs forward, it looks like growth. When it runs backward, it looks like impairment, and it runs backward fast.
Now add the stablecoin link, which almost nobody prices correctly. The largest stablecoin issuers hold reserves in short-dated Treasuries and repo. That is clean, liquid, senior collateral. But the demand for stablecoins is itself a function of the on-chain credit complex β of tokenized yield products, of RWA vaults, of the DeFi lending markets that accept them. If the tokenized credit layer experiences a redemption event, the demand for stablecoins shifts, and the reserve management assumptions that looked safe at the aggregate level get tested at the margin. I am not forecasting a stablecoin break. I am noting that the reserve composition is clean while the demand function is not, and those two facts are being conflated in every risk model I have seen.
Let me quantify what I can, and be honest about what I cannot. I do not have DBS's AI-related credit exposure; nobody outside the bank does. I do not have the aggregate size of tokenized AI-infrastructure debt, because most of it sits in permissioned pools that do not publish holdings. What I can observe is the structure: tokenized credit referencing real-world cash flows, posted as collateral in DeFi markets, priced by lagged oracles, redeemable on a faster clock than the underlying. That structure has a name in every risk manual I have written. It is a hidden leverage channel.
My 2020 stress test taught me its shape. I ran 1,000 scenarios on Aave v1 and Compound v1 with $50 million at risk. The finding was not that the protocols were unsafe. It was that the risk parameters adjusted slower than volatility. Reserve factors moved on governance timescales; liquidity crunches moved on block times. The gap between the two is where losses live. That gap is now wider, because the asset underneath is no longer a liquid token with a real market price. It is a datacenter note that trades once a quarter, if at all.
Code is law, but human greed is the bug. The code will faithfully liquidate at the oracle price. The oracle will faithfully report the last NAV. The NAV will faithfully reflect a valuation that predates the repricing. Every component behaves exactly as specified, and the system still marks the book wrong. No audit catches this, because nothing is broken. The specification is the vulnerability.
Here is the counter-intuitive part. The consensus reading of the DBS statement is benign. Repricing equals discipline. Eliminating weaker participants equals a healthier industry. Both framings are comfortable, and both are incomplete.
The blind spot is that the repricing does not stay in the loan book. When AI credit spreads widen, the first institutions to feel it hold the debt on a mark-to-market basis. The second hold it as collateral. And the elimination of weaker participants is not a bloodless culling. It is a sequence of forced sales, and forced sales need buyers. If the collateral being sold is a tokenized note whose oracle still prints par, the pool buying it believes it is acquiring par risk. It is not. It is acquiring repriced risk at yesterday's price.
This is what the positive framing hides. The repricing has already started off-chain. The on-chain marks have not caught up. The lag is not a rounding error. It is the exact window in which informed sellers exit and passive pools absorb. I have watched this before. In 2022, the collateral looked liquid right up until everyone tried to redeem at once, and then liquidity vanished faster than the price feed could update. We build bridges in the storm, not after the rain. The protocols that survive the next repricing will be the ones that built redemption gates, duration-matched liabilities, and conservative oracles before the spreads moved β not the ones that add them after the first NAV cut.
The asymmetry is stark. Upside from the tokenized credit trade was capped at the coupon. Downside is uncapped, because it is tied to the credit quality of a leveraged SPV in a sector whose capital intensity is rising while its return per dollar falls. The investor who bought 12 percent thought they were being paid for yield. They were being paid for ignorance, and the bill is now arriving.
So watch the right signals. Not the headline. Watch the NAV feeds on tokenized credit vaults and how often they refresh. Watch the redemption gates and whether they shorten or lengthen. Watch the first datacenter SPV that misses a coupon, and whether its tokenized wrapper marks down the same week or the next quarter. The distance between those two dates is the size of the problem.
DBS told us the repricing is underway. The on-chain collateral is still marking yesterday's price. The question is not whether the two numbers reconcile. It is who is holding the position when they finally do.
