Fink Says AI's Real Risk Is Rates, Not Bubbles. The On-Chain Curve Agreed Weeks Ago.
Hook
On the day the 10-year Treasury yield touched 5.34% — the highest print in more than two decades — the loudest argument in the market was still about whether AI is a bubble. That was the wrong argument, and the bond curve had already told anyone listening.
Larry Fink, the BlackRock CEO who spent the 1980s trading mortgage-backed securities before he spent the 2000s building the largest asset manager on earth, published a Wall Street Journal op-ed with a thesis that cuts against his own book: the danger is not that AI is overvalued. The danger is that there is not enough capital in the world to fund it. Ray Dalio, in a separate note, called the same market "a classic bubble approaching rupture." Joachim Klement of Panmure Liberum said the hyperscalers' free cash flow is essentially spent.
Three men. One asset class. Three mutually exclusive theses. And a 10-year yield that moved 120 basis points in a single year while the AI complex added a trillion dollars of market cap. The chart didn't lie — the plumbing was repricing faster than the narrative, and the plumbing is always where I look first.
I trade options for a living and I run an automated agent against on-chain rate markets. I do not care about the AI story. I care about the discount rate applied to it. And for the last several weeks, the discount rate has been screaming something the equity market refuses to price. This piece is about what Fink got right, what he got catastrophically wrong, and the specific number that will decide who is holding the bag when the financing window closes.
Context: Three Men, One Balance Sheet
Fink's argument is short and worth restating precisely, because the coverage of it has been lazy. He does not deny that AI valuations are rich. He reframes the problem. In his telling, there are three bottlenecks, and none of them are chip design.
First, the capital. He proposes securitizing compute infrastructure — the data centers, the GPU clusters, the power contracts — so that a much wider pool of investors can fund the buildout rather than a handful of hyperscalers and their bankers. This is, almost word for word, the logic that created the mortgage-backed security.
Second, the power. Fink points at China's expansion of generating capacity as the real competitive threat, not semiconductor export controls. The argument is that AI is electricity wearing a costume, and the United States is losing the electricity race on permitting timelines, not on engineering.
Third, the investor base. He cites the roughly 40% of Americans who do not own equities and argues they should be brought into the growth. That is a political statement dressed as a financial one, and it is the part of his op-ed that will get quoted in Washington.
Against him stands Dalio, who sees the signature of a late-stage bubble — leverage, concentration, retail enthusiasm, and a compelling story about a new paradigm. And against him stands Klement, who makes the sharpest empirical claim of the three: that the hyperscalers' free cash flow is largely exhausted, meaning the next phase of AI capex must be debt-financed. That single claim, if true, changes everything about how the AI complex behaves when rates rise. If it is false, Fink's optimism is structurally sound.
Before I go further, one flag, because I verify numbers before I build a thesis on them. The reporting around Fink cites a United States federal debt figure of 40 trillion dollars. Treasury Direct, the official source, has been printing a number closer to the mid-30s of trillions. The discrepancy matters because every crowding-out argument scales with the debt stock. If the real number is meaningfully lower, the capital-scarcity thesis weakens at the margin. If it is higher — because the broader measure includes intragovernmental holdings — then Fink is understating his own case. Either way, do not take a round number from a newspaper at face value. Pull the tape yourself.
Strip the personalities away and the article is not about AI at all. It is about fiscal dominance — the condition in which government borrowing needs and private investment demand compete for the same pool of savings, and the central bank loses effective control of the long end of the curve. Everything below flows from that.
Core: The Only Variable That Matters
The numerator is a distraction; the denominator is the trade
A stock is worth its future cash flows discounted back at some rate. The entire AI bubble debate is a fight about the numerator — will the earnings show up? That is a genuinely hard question and nobody has an edge on it. The denominator is a much easier question, and almost nobody is trading it.
Here is the arithmetic. If the risk-free rate moves 120 basis points, the fair value of a long-duration growth asset — one whose cash flows sit far in the future — compresses by roughly 20 to 25 percent, all else equal. You do not need the earnings to disappoint. You just need the discount rate to keep grinding higher, and the same stock that looked reasonable at a 4% risk-free rate looks expensive at 5.34%.
This is what Fink is actually saying, whether he intends it or not. The AI buildout is a capital-consuming machine, and capital is now priced higher than it has been in a generation. The financing cost is the story.
The dangerous setup is not a bubble popping. It is a double-kill: earnings that miss because capex is being depreciated against a shorter useful life than anyone modeled, colliding with a discount rate that stays elevated. That combination does not produce a correction. It produces a repricing.
BUIDL is Fink's thesis, already live on-chain
Here is the part that the macro commentators missed entirely, and the reason I am writing this for a crypto audience rather than a rates desk.
Fink's first and third proposals are not proposals. They are descriptions of what BlackRock is already doing, and they are happening on public blockchains.
The BUIDL fund — BlackRock's tokenized money market product — launched in 2024 and has been expanding across chains ever since. It is a tokenized wrapper around Treasury bills, sitting on Ethereum and later on other networks. It is, functionally, step one of Fink's plan: take a traditional asset, wrap it in a programmable instrument, and open it to a broader set of holders who can settle it 24/7 without a prime broker.
Tokenized Treasuries as a category have grown from a curiosity to a multi-billion-dollar market in under two years. I watched this happen from inside it. In early 2024, after the spot Bitcoin ETF approval, I ran a script against the premium and discount between ETF shares and spot BTC on Coinbase and executed more than fifty round trips across venues during the initial volatility, netting roughly eight thousand dollars in what was, for a two-week window, genuinely close to risk-free. The lesson I took from that was not that arbitrage is easy. It was that institutions were arriving with better plumbing than retail had ever had, and the same plumbing would eventually wrap everything — including the assets funding AI.
So when Fink says "securitize compute," I hear something specific. I hear: we are going to take GPU cash flows, put them in a legal wrapper, tokenize the wrapper, and sell it to a global base of yield-seekers who can buy it from a phone. That is not a metaphor. It is a product roadmap, and the rails already exist.
The duration mismatch nobody wants to price
This is where the analysis gets uncomfortable, and it is the single most important technical point in this article.
A mortgage-backed security worked — until it didn't — because of a structural alignment. A 30-year mortgage is backed by a 30-year asset: a house that, historically, holds or appreciates in value over the life of the loan. The cash flows amortize. There is a legal claim on a physical thing. The duration of the liability roughly matched the duration of the collateral.
Now replace the house with a GPU.
A high-end data center accelerator has an economic useful life measured in three to five years before it is obsolete, and its residual value collapses hard. Its cash flows are not amortizing and predictable — they are spot-priced compute rentals, and spot compute prices are among the most cyclical numbers in technology. When demand softens, rental rates do not glide down. They gap.
So when you securitize compute, you are wrapping a short-duration, high-beta cash flow in a long-duration wrapper. If you issue a ten-year note against a four-year depreciating chip, you have manufactured a structural maturity mismatch and an embedded terminal-value bet. The credit event is not a missed coupon. The credit event is the depreciation schedule.
I have seen this movie in adjacent markets. Aircraft leasing ABS and container shipping ABS both securitize depreciating, cyclical, mobile assets, and both price at several hundred basis points wide of mortgage paper because investors understand that the collateral can lose value faster than the loan amortizes. Compute paper will not trade at mortgage spreads. It will trade wide, and the spread between compute-ABS and Treasury yields will become one of the most-watched new numbers in credit markets within two years. If someone tries to price it tight by arguing that demand is structural, they are repricing a cyclical asset as an annuity, and that is the exact error that built 2008.
One more forensic detail. The depreciation assumption is a management choice. Extending the assumed useful life of a GPU flatters current earnings and makes the securitization look better collateralized. Every quarter, read the useful-life footnote before you read the revenue line. The accounting is the alpha.
The buyer of last resort is now a stablecoin
Now the part that ties fiscal dominance directly to crypto.
Stablecoin issuers are among the largest holders of short-dated United States government debt on the planet. The largest issuer's reserves are overwhelmingly T-bills. Collectively, dollar-token issuers have become a structural, price-insensitive bid for the front end of the Treasury curve. This is, literally, Fink's third step happening without him: a broader investor base funding the government, transacted in dollars, settled on-chain, and completely outside the traditional banking system.
The irony is thick. The same crypto industry that was supposed to be an escape hatch from the dollar has become one of the dollar's most reliable buyers of its debt.
But here is the technical nuance that almost everyone gets wrong, and it is the reason stablecoin growth is not the solution to the problem Fink describes. Stablecoins are solving the wrong maturity. They buy bills — the very short end. The deficit, and therefore the pressure, is at the long end. Term premium is a function of duration risk, and no amount of T-bill demand from token issuers compresses the term premium on 10- and 30-year paper. The stablecoin bid lands exactly where the Federal Reserve already has a lever, and does nothing where the Fed has lost one.
So the crypto industry is providing a marginal bid for the least problematic part of the curve while the genuinely dangerous part — the long end, where the crowding-out happens — remains unanchored. If you are holding a thesis that stablecoin growth will somehow backstop Treasury demand and keep rates contained, you are watching the wrong tenor. Pull up the 30-year, not the 3-month.
What my own book did
I run an open-source AI trading agent against a personal dashboard. I backtested it across 2020 through 2024, got a Sharpe ratio north of 1.3, and deployed ten thousand dollars of live capital in early 2025. It found a recurring arbitrage in cross-chain bridge pricing and generated roughly three thousand dollars a month for a stretch.
But the interesting output was not the P&L. It was what the agent flagged and what I overrode.
Repeatedly, the agent identified the spread between on-chain stablecoin lending yields and the tokenized Treasury yield as a "risk-free" carry trade. It is not risk-free. It embeds the credit of the token issuer, the smart contract, the bridge, and the chain. The agent was pricing a spread it had no way to underwrite. That is a lesson in humility for anyone deploying automated capital: the model prices what it can measure, and it cannot measure the thing that kills you.
The second thing the agent taught me is more relevant here. It does not care about the AI narrative. It prices funding spreads. And the on-chain funding spreads — the cost of borrowing dollars against crypto collateral, the premium for locked yield — have been moving in the same direction as the 10-year Treasury all year. The decentralized rate market and the sovereign rate market are telling the same story, and that story is: capital is getting expensive, and it is going to keep getting more expensive.
The options trade nobody is positioned for
I am an options strategist, so let me end the core section where I actually live.
The market is pricing equity volatility as though the only risk is an earnings disappointment. It is pricing rate volatility as though the central bank still controls the curve. Both of those are wrong in the same direction.
Here is the asymmetry. If you own long-duration AI equities, you are structurally short a rate put. You have sold insurance on rates staying contained, and you may not know it. Meanwhile, rate hedges — swaptions, long-duration Treasury puts, steepeners — are cheap relative to the equity protection everyone is crowding into. The crowd is buying the hedge for the risk they can see (a bubble popping) and ignoring the hedge for the risk that is actually moving (the discount rate).
Risk isn't a feeling. It is a number, and the number is 5.34% and climbing. When the cost of the denominator rises, you do not need a single bad earnings report to lose money. You just need to be holding duration.
Contrarian: Fink Is Right About the Diagnosis and Wrong About the Cure
Everyone is watching the bubble. The bubble is the wrong frame, because a bubble is a claim about price, and price is a symptom. The disease is the financing structure. Fink identifies the disease correctly and then prescribes the pathogen.
His cure — securitize the compute, slice it, sell it to a broader base — is the reflex of a man who spent his formative years in the mortgage market. When you cannot fund something through the traditional channel, you do not stop building. You repackage the cash flows and find a new buyer. That instinct built the modern asset management industry and it also built 2008. The difference between those two outcomes was never the sophistication of the structure. It was the quality and duration of the collateral. And compute collateral is shorter-lived and more cyclical than a house. Fink is proposing to run the same play with worse collateral.
Then there is the dispute that actually matters, and it is resolvable with public data. Fink says the hyperscalers are the best-capitalized players in the market. Klement says their free cash flow is spent. Both cannot be right. The public filings lean Fink's way today: aggregate hyperscaler operating cash flow still exceeds aggregate capital expenditure, though the gap has been narrowing sharply as capex guidance climbs. So Fink is correct on the level and Klement is correct on the trajectory, and the trajectory is what kills you. When the coverage ratio crosses one, the buildout flips from self-funded to debt-funded, and a debt-funded buildout is a rate-sensitive buildout. That crossover is the event to watch, and nobody is watching it.
Here is the blind spot that no one on either side of the bubble debate is pricing: the term premium. Everyone is arguing about the Fed's next move. The Fed controls the short end. The long end is being set by supply — by the volume of duration the Treasury must place — and by the shrinking pool of price-insensitive buyers. The Fed can cut and the 10-year can rise at the same time, and if that happens, every model that assumes monetary policy drives the cost of capital is broken.
I bought the pixel, not the promise. The market bought the promise — the AGI narrative, the productivity miracle, the new paradigm. The pixel is the balance sheet, the depreciation schedule, and the funding cost. When the promise and the pixel disagree, the pixel wins. It always does.
Takeaway
Three numbers, and I am watching all three daily.
The first is the 10-year Treasury yield. A sustained move above 5.5% is the tripwire. That is the level at which the discount-rate compression stops being a headwind and becomes a repricing event across every long-duration asset on earth, AI and crypto included.
The second is the hyperscaler free cash flow to capital expenditure coverage ratio. Above one, Fink is right and the buildout is self-funded. Below one, Klement is right and the buildout is debt-funded and rate-sensitive. That single ratio resolves the entire debate, and it updates four times a year.
The third is the growth of tokenized Treasury products and any first-of-kind compute securitization. If BUIDL-style wrappers keep expanding and a compute-ABS eventually prints, Fink's roadmap is executing in real time and you can track it on-chain before it hits a headline.
The forward question is simple and nobody wants to answer it. If the largest capital buildout in modern history gets funded by the same securitization machine that funded subprime, and the collateral depreciates in four years instead of thirty, who holds the paper when demand for compute goes soft? The answer will not be the hyperscalers. It will be whoever bought the yield.
Liquidity vanishes when the music stops. It just does it faster on-chain.
