Volatility isn't a bug in the AI trade. It is the admission fee. But the volatility that cracked AI equities this past quarter had nothing to do with benchmark scores, chip yields, or the next frontier model. It had everything to do with something far more boring and far more lethal: the cost of money.
Here is the signal most people scrolled past. A crypto-native outlet — not the Journal, not Bloomberg, a crypto outlet — ran a headline claiming AI stocks tumbled because rising borrowing costs were testing industry financing. Sit with that for a second. When the crypto press starts narrating the AI trade's debt problem, the story has already left the professional layer and entered the retail bloodstream. That is a sentiment marker, not a data point. And I treat sentiment markers the way I treat a 4am Telegram pump: with suspicion and a hard stop.
The headline was directionally right and almost entirely empty. No named companies. No basis points. No maturities. No spread levels. No durations. Just a vibe dressed up as news. And yet, buried under the filler, there is a real proposition, and it is a big one: the AI capital-expenditure supercycle is increasingly financed with debt, and debt gets repriced by interest rates. That single sentence is worth more than the article that failed to explain it.
So let me explain it. Because I have watched this exact movie before — in DeFi, in 2020, when subsidized yield met its cost of capital and the farmers left.
To understand why a crypto outlet is writing about AI borrowing costs, you have to understand what the AI buildout actually is. It is not a software story. It is an infrastructure story. It is steel, silicon, power, and cooling — the four most capital-hungry inputs in modern industry, assembled into buildings that cost anywhere from several hundred million to several billion dollars apiece.
Through 2023 and 2024, the hyperscalers — Microsoft, Google, Amazon, Meta — funded this buildout largely from operating cash flow. That was the comfortable phase. Combined AI-related capex across the major cloud players crossed roughly two hundred billion dollars in 2024 and was guided toward three hundred billion-plus for 2025. For a while, the market treated those numbers as pure upside: more capex meant more compute, more compute meant more capability, more capability meant more revenue, and more revenue justified more capex. The circular logic felt airtight because everyone inside the loop was winning.
Then the financing mix changed, and it changed quietly. In late 2024, Meta issued its first-ever corporate bond — roughly ten and a half billion dollars, with a meaningful chunk earmarked for AI infrastructure. Oracle, chasing enormous AI cloud contracts, leaned hard on debt markets to fund the buildout. Data center developers began flooding the ABS and CMBS markets. Private credit funds — the shadow banks of the post-2008 world — piled into data center construction loans. Sale-leaseback structures proliferated, letting operators convert owned assets into operating leases to free up capital for the next round.
Read that list again. Bonds. Asset-backed securities. Commercial mortgage-backed securities. Private credit. Sale-leaseback. These are not equity instruments. They are not cash-flow instruments. They are duration instruments, and duration instruments live and die by the discount rate. The moment AI capex migrated from funded-by-cash-flow to funded-by-credit, the entire trade acquired a new sensitivity it did not have before: it became rate-sensitive.
That is the context the empty headline was gesturing at. And it matters, because it means the AI trade is no longer just a bet on intelligence. It is a bet on the term structure of interest rates. Those are very different bets, and most people holding AI exposure have never repriced their thesis to reflect the switch.
I know what happens when a narrative funded by cheap credit meets a repricing, because I have lived it. In 2020, I ran yield farms across Uniswap, SushiSwap, and Compound with real money — fifty thousand USDC, sixteen-hour days, manually rebalancing to chase APY and arbitrage slippage between DEXs. The yields looked insane. Then I ran the math on what happened when the subsidy stopped. The answer was not pretty, and it is the answer AI is about to discover at a thousand times the scale.
The transmission chain nobody prices until it breaks
Start with the mechanism, because mechanisms are where narratives go to die.
The chain runs like this. The risk-free rate rises. Corporate bond yields rise with it, and for lower-rated issuers they rise faster — the spread widens on top of the base. That lifts the cost of debt for anyone financing a data center. The cost of capital for a compute project goes up. The internal rate of return on that project goes down. And at some point — a point that arrives faster for leveraged, marginal, or speculative projects — the IRR drops below the hurdle rate and the project gets shelved.
This is not speculation. This is arithmetic. A data center is a long-duration asset: heavy upfront capital, multi-year construction, decades of useful life, revenue that ramps slowly against a cost base that starts day one. When you finance a long-duration asset with floating-rate or refinancing-dependent debt, you have created a structure exquisitely sensitive to the path of rates. The asset does not care about your model's benchmark score. It cares about its cost of capital.
And here is the part the headline missed entirely: the pain is not distributed evenly. It is distributed by balance sheet. The hyperscalers can fund from cash flow and absorb a rate shock. The leveraged compute lessors — the so-called Neoclouds — cannot. The data center developers dependent on ABS and private credit cannot. The difference between those two groups is the entire story, and the headline flattened it into two words: AI stocks.
I have audited enough protocols to recognize what happens when a structure is only solvent under favorable conditions. The tell is always the same. The operator talks about growth, never about the refinancing calendar. The pitch deck has a technology slide and no maturity wall. That is not a business. That is a duration bet wearing a business suit.
The circular financing trick
Now the part that genuinely worries me, because I have read enough tokenomics to recognize a reflexive loop when I see one.
A meaningful share of AI revenue is not clean, arms-length, end-user revenue. It is circular. Chip designers invest in AI startups. Those startups spend the money on chips. The chip designer books the revenue, the startup books the capability, and the market books both as independent validation. The cleanest example is a chip company investing in AI labs that then buy its GPUs — but the pattern is broader. Cloud providers invest in model labs. Model labs commit to multi-year cloud spend. The cloud provider books committed revenue, and the commitment becomes a comp in the next round.
I am not accusing anyone of fraud. Circularity is not the same as fabrication. But circularity does something specific and dangerous to valuation: it inflates the demand side of the equation while hiding how much of that demand is self-referential. When credit is cheap, nobody interrogates the loop, because the loop is generating exactly the growth everyone wants to see. When credit gets expensive, the loop tightens — because the startup can no longer raise the debt to fund the chip purchase that generated the revenue that justified the valuation that raised the next round.
Code is law, but human greed writes the loopholes. And the biggest loophole in AI accounting right now is that revenue growth and revenue quality are being treated as the same variable. They are not. A dollar of arms-length enterprise revenue and a dollar of intra-ecosystem revenue look identical on an income statement and behave completely differently when the funding environment flips.
This is why the rate shock matters more than the headline suggested. It is not just raising the cost of capital. It is stress-testing whether the revenue was ever as organic as the narrative claimed.
AI revenue is the new TVL
Here is the insight I want you to sit with, because it is the reason I am writing this at all, and it is the reason a DeFi strategist is the right person to write it.
In 2020, DeFi TVL — total value locked — was the number everyone quoted. It went parabolic. It was cited as proof that DeFi was winning, that capital was migrating on-chain, that the future had arrived. And then, in 2022, when token emissions slowed and the incentive programs wound down, what emerged was real TVL — and it was a fraction of the headline number. The gap between headline TVL and organic TVL was the gap between subsidized and real capital. The subsidies had been masking how much of the locked value was there only to farm emissions and would leave the instant the subsidy stopped.
AI revenue in 2025 is the new TVL.
When you finance growth with cheap credit and circular deals, your headline revenue number includes a large component of subsidized, reflexive, incentive-driven activity. That is exactly what emissions did to TVL. And just as a rate shock in DeFi revealed the organic base, a rate shock in AI reveals the organic revenue base. The rising cost of borrowing is not causing AI revenue to fall. It is revealing how much of it was never organic to begin with.
This is the most important reframe available right now. The bears say AI revenue is fake. The bulls say AI revenue is real. Both are wrong. AI revenue is real and subsidized, in the same way DeFi TVL was real and subsidized. The headline number was never a lie. It was a gross figure that needed a net figure to be meaningful — and the net figure only becomes visible when the cheapest source of funding is withdrawn.
I learned this the hard way. In 2020 I farmed tokens I did not understand because the APY was north of three hundred percent. I treated gross yield as real yield. It was not. The moment emissions fell, the yield collapsed to single digits and impermanent loss ate the difference. I don't make that mistake twice. When I see a headline growth number now, my first question is no longer how fast is it growing. It is: what happens to this number when the cheapest source of funding disappears?
Apply that question to AI. The answer is the entire thesis of this piece.
GPUs as collateral: the Neocloud time bomb
Let me get specific, because abstraction is where analysts hide.
The most fragile node in this entire structure is the GPU-backed lender — the Neocloud. CoreWeave is the canonical example, but it is a category, not a company. The model is straightforward: buy massive quantities of GPUs, use them as collateral for debt, lease the compute out to AI labs and enterprises, and use the lease revenue to service the debt. On paper it works beautifully. GPUs are productive assets. They generate cash flow. The debt is secured.
Now apply the discount rate. The Neocloud's cost of debt rises. Its refinancing risk rises. And here is the ugly part: GPUs are depreciating collateral. The asset backing the loan loses value on a schedule of roughly eighteen to thirty-six months, driven by each new chip generation. So you have a borrower whose collateral is depreciating fast, whose debt cost is rising, and whose lease revenue depends on the same AI demand cycle that funds everyone else in the chain.
That is a duration mismatch with a depreciating-collateral kicker. In a low-rate world it works, because you can always refinance cheaply and roll the collateral into the next generation. In a rising-rate world, the refinance window closes at exactly the moment the collateral is worth least. That is not a stress scenario. That is the base case for the marginal Neocloud.
I saw this exact shape in 2022. Terra's UST was collateralized by a mechanism — the mint-and-burn arbitrage — that worked perfectly as long as confidence held and broke catastrophically when it did not. I held a small UST position into the de-peg because I over-trusted the mechanism and under-weighted the exit liquidity. I lost twelve thousand dollars in hours. The lesson was not that algorithmic stablecoins are bad. The lesson was: any structure whose solvency depends on continuously favorable conditions is not a structure, it is a bet. GPU-backed lending is that bet, repackaged for the AI era, with a chip roadmap standing in for an arbitrage curve.
The duration math: why 100bp eats 20% of the multiple
Now the equity side, because this is where the headline was technically correct and analytically useless.
AI equities are, almost without exception, long-duration assets. Their valuation is dominated by cash flows projected far into the future — 2027, 2030, beyond. When you discount distant cash flows, the discount rate does enormous work. Small changes in the rate produce large changes in present value. This is not opinion. It is the mechanics of discounted cash flow, and it is the reason growth stocks trade like long bonds.
The practical rule of thumb: for a high-duration growth asset, a one-hundred-basis-point rise in the risk-free rate can compress the valuation multiple by roughly fifteen to thirty percent, depending on how far out the cash flows sit and how much of the value is terminal. Now layer on a second effect: rising rates also tend to slow the economy, which trims the growth expectations sitting in the numerator. You get hit on the discount rate and the growth rate at the same time. That double hit has an old name. It is called a de-rating, and when it compounds it becomes a double kill.
So the headline's claim — AI stocks tumbled as borrowing costs rose — is mechanically coherent. Of course they did. If you build a trade on the implicit assumption of low discount rate plus infinite growth, and then the discount rate stops being low, the trade re-rates. The only questions are how much, and for whom. And the answer depends entirely on which AI stocks you actually mean.
Which brings me to the part I care about most.
The asymmetry: who bleeds, who feasts
Here is where the headline's homogeneity does real damage. AI stocks is not an asset class. It is a label slapped across at least four completely different business models with completely different rate sensitivities.
Tier one, the hyperscalers. Microsoft, Google, Amazon, Meta. Fortress balance sheets, massive operating cash flow, investment-grade credit. A rate shock raises their cost of capital marginally and strengthens their competitive position dramatically, because their leveraged competitors get squeezed. In a credit tightening, the strong get relatively stronger. That is not a bug in the system. It is the system.
Tier two, the Neoclouds and compute lessors. High leverage, GPU collateral, refinancing dependence. This is the bleeding edge. Rate shocks hit them first and hardest. Some will not survive a sustained high-rate environment, and the ones that do will survive by selling assets to tier one.
Tier three, the data center developers and REITs. Capital-intensive, reliant on ABS and private credit. A rate shock raises their capitalization rates, which lowers the value of every existing asset on their books, which impairs their ability to raise the next round. Slow bleed, high sensitivity, long tail.
Tier four, the semiconductor supply chain. Indirectly exposed. If data center orders get delayed, GPU, HBM, and networking demand guidance gets cut with a two-to-four quarter lag. This is where the expectation gap forms — the market prices the order cut before it shows up in the numbers.
Now map that onto a portfolio. If you are long AI indiscriminately, you own all four tiers, and you have no idea which part of your book is a rate bet and which part is a technology bet. That is not investing. That is buying a label and calling it a strategy.
What this means for on-chain capital
I am a DeFi yield strategist, so let me close the loop and tell you why a crypto audience should care about an AI financing story.
First, correlation. AI equities and crypto have become increasingly correlated as both have been absorbed into the same risk-on, liquidity-driven, long-duration bucket. When the discount rate rises, both get repriced. If you believe your BTC and ETH exposure is uncorrelated to the AI trade, look at what happened to both on the days the AI names sold off on rate fears. The correlation is not zero, and in a liquidity event it approaches one. Diversification across two expressions of the same macro factor is not diversification.
Second, the RWA story. For three years the pitch has been that tokenizing real-world assets — including, lately, data center exposure — would bring institutional capital on-chain. Here is the uncomfortable truth I have held since the beginning of that narrative: the institutions do not need your public chain to finance a data center. They have the ABS market, the CMBS market, the private credit market, and a dozen investment banks. The RWA pitch assumed tokenization solved a problem institutions had. It did not. The problem was never distribution. It was credit assessment, and a blockchain does not assess credit. When rates rise and credit quality becomes the binding constraint, capital flows to whoever can underwrite risk — and that is not a permissionless pool. The RWA narrative is a three-year storytelling exercise, and the rate shock is about to expose how little of it was ever about real capital formation.
Third, and most important: the AI trade is running a DeFi playbook, and DeFi already knows how it ends. Subsidized growth. Reflexive revenue. Depreciating collateral. Duration mismatch. Leverage stacked on a narrative. I have traded through every one of those failure modes with my own money. The AI industry is discovering, in real time and at a scale of hundreds of billions of dollars, the lessons that cost me my first sixty percent in 2017 and my UST position in 2022.
In 2017 I put half a million RMB into three low-cap ERC-20 tokens during the ICO frenzy, trusting hype velocity and social volume instead of whitepapers. Two of them rugged within weeks. The third spiked four hundred percent and then gave it all back. Net loss. The lesson was not that crypto is a scam. The lesson was that I had financed a narrative with capital I could not afford to lose, on a timeline I did not control. That is precisely the structure of a leveraged data center build in a rising-rate environment.
The bear-market frame: survival, not upside
We are in a bear market. I want to be explicit about what that means for how you read all of this, because the temptation in a piece like this is to hunt for the short.
In a bear market, the question is not how do I make ten-x. The question is which of my positions is a rate bet in disguise, and can I survive a repricing. That reframe changes everything. If you hold AI-adjacent crypto exposure — compute tokens, GPU marketplaces, decentralized AI networks — you need to ask a different question than you asked in the bull market. Not is the technology good, but is the demand for this technology funded by cheap credit, and what happens when the credit is no longer cheap.
The decentralized compute tokens are the most exposed. Their entire thesis is that they can undercut centralized compute on price. But if centralized compute gets repriced — if Neoclouds fail and their GPU capacity gets dumped on the market at distressed prices — then the price advantage of decentralized compute narrows or inverts. The bear case for decentralized compute is not that the technology does not work. It is that the centralized incumbents are about to flood the market with cheap distressed capacity. That is a second-order effect almost nobody is pricing, and it is the kind of thing that turns a good narrative into a bad position.
I watched the same dynamic in the 2022 lending markets. When the big centralized lenders froze, the distressed assets did not disappear. They got auctioned into the hands of whoever had cash and a mandate. Decentralized alternatives did not win by being better. They won, if they won at all, by being solvent when everyone else was not.
Here is the angle almost everyone gets wrong, and it is the one I would bet on.
The consensus read of AI stocks tumbling on rising borrowing costs is bearish: the AI bubble is popping, the capex supercycle is over, get out. That read is half right and strategically wrong.
The correct read is that this is not the end of the AI buildout. It is the beginning of its consolidation. Rising rates do not kill the buildout. They kill the leveraged participants and hand their assets to the cash-rich ones. The hyperscalers come out of a rate shock with more market share, cheaper acquisitions, and a more defensible moat. The rate shock is a transfer mechanism, not a termination mechanism. It moves capital from the fragile to the funded.
And there is a sentiment signal buried in the headline itself. When a crypto outlet starts narrating an AI financing crisis, the narrative has gone mainstream. Narratives that have gone mainstream are either at their emotional peak or in their early panic phase. Both are dangerous to trade naively. Retail sees AI stocks tumble and either panic-sells everything or buys the dip blindly. Smart money sees it and starts separating the balance sheets. The move is not to be long or short AI. The move is to be long the strong balance sheets and short — or simply avoid — the leveraged ones.
That is the trade. Not a direction. A spread. Retail trades direction. Precision trades structure.
So watch the right signals, not the headline. Ten-year yields and high-yield spreads. Quarterly capex guidance from the hyperscalers — the first downward revision is the tell, not the tenth. Data center lease signings and cancellations, the quiet data that precedes the headline by a quarter. Refinancing calendars for the leveraged compute names. And the gap between headline AI revenue and the revenue that survives without cheap credit — the organic TVL of the AI trade.
The AI supercycle is not dead. It is being repriced, and repricing separates the funded from the fragile. The same thing happened to DeFi in 2022. The protocols that survived were the ones whose revenue did not depend on the subsidy. Ask the same question of your AI exposure, and of your crypto exposure that rhymes with it: when the cheap money stops, what is actually left standing — and will you still be holding it?