The Jitters Nobody Priced Correctly
On a Tuesday in late November, two of the largest beneficiaries of the artificial-intelligence buildout repriced on credit concerns, and the third did not. Broadcom's spreads widened on the leverage it absorbed with VMware and on the capital intensity of a custom-accelerator backlog it must fund before it can bill; Oracle's bonds sagged further below par as investors re-examined a remaining performance obligation that now dwarfs its annual cash generation; Nvidia โ the firm at the gravitational center of the entire complex โ barely moved at all. The tape said something the headlines did not. This was never a story about AI demand. It was a story about funding structure, and the market had begun, quietly and then not so quietly, to discriminate between firms that can finance themselves and firms that must borrow to participate.
The data hides what the eyes refuse to see.
I have watched this movie before, from a different seat. In 2020, at the height of DeFi Summer, I spent twelve hours a day building Python models to track stablecoin velocity across Ethereum mainnet, trying to reconcile the yields protocols advertised with the capital that actually arrived. What I found was uncomfortable and formative: roughly seventy percent of the TVL growth I was measuring was illusory leverage โ the same dollar rehypothecated through recursive lending loops, each hop manufacturing a new claim on the same underlying collateral. That disillusionment pushed me away from chasing yields and toward a single discipline I have applied to every market since. Before you analyze demand, analyze who is paying for it, with whose balance sheet, and at what tenor.
Which is why the AI credit jitters are not, for a crypto analyst, a peripheral story. They are the main story.
Context: The Architecture of AI's Credit Boom
For three years, the AI buildout was funded by the least fragile money in the world: hyperscaler operating cash flow. When Microsoft, Alphabet, Amazon or Meta funds a datacenter out of retained earnings, there is no credit event to price, no refinancing wall to climb, no counterparty to question. That phase is over, and its ending is the single most important macro development of the past eighteen months โ more important, for crypto, than any protocol upgrade or ETF approval.
The second phase is vendor financing and circular capital. Nvidia's letter of intent to invest as much as one hundred billion dollars in OpenAI; OpenAI's multi-year compute commitment to Oracle reported in the neighborhood of three hundred billion dollars; Broadcom's engagement to co-develop custom accelerators measured in gigawatts of deployment. None of this is fraudulent, and treating it as fraud is a category error. This is simply how capital-intensive ecosystems bootstrap themselves โ the same way railroads financed towns and towns financed railroads in the nineteenth century. But it converts a technology bet into a duration bet, and duration is where credit cycles are born.
The third phase is private credit, and it is the phase that should command every macro watcher's attention. Meta's Hyperion campus, financed through a special-purpose vehicle capitalized at roughly twenty-seven to thirty billion dollars with Blue Owl. CoreWeave's GPU-backed debt facilities, which have grown past eleven billion dollars. Oracle's bond issuance of approximately eighteen billion dollars, layered on a debt stack now approaching one hundred billion. The marginal dollar funding AI infrastructure is no longer a corporate treasurer's dollar. It is a credit fund's dollar, and credit funds have different clocks, different covenants, and different tolerance for being wrong.
Why that distinction is structural rather than cosmetic: private credit is not a worse lender than a bank, but it is a less transparent one. It marks to model rather than to market, which is a feature in calm conditions and a bug in stressed ones. When a public bond falls, everyone sees the same price at the same time. When a private loan is marked at par because the model says par, the price discovery is deferred, not avoided โ and deferred price discovery is precisely the mechanism by which a repricing that should take weeks takes quarters, and then takes everything at once.
This is the territory Europe's regulatory architecture was designed for, and it is instructive that the architecture does not yet reach here. MiCA brought genuine clarity to crypto-asset issuance and stablecoin reserves across the union; the American stablecoin legislation did something similar on the other side of the Atlantic. But neither framework was written with the AI credit channel in mind, and the gap is now visible: AI infrastructure debt is being issued into a regime whose disclosure standards are lower than those applied to listed equity, even as its systemic footprint becomes comparable. When I analyzed the legal fragmentation of MiCA across twenty-seven member states in 2025 and identified a five-billion-euro arbitrage in cross-border stablecoin settlement, the lesson was not that fragmentation is dangerous in itself. It was that capital will route around whatever structure is least legible to regulators, and that the least legible structure is usually the most levered.
There is a tempting analogy here, and I want to use it precisely rather than loosely. In May 2022, after Terra and Luna collapsed, I retreated to a cabin in Dalarna for three weeks of deliberate digital silence. I did not write reactive commentary. I built contagion models, because I wanted to understand the shape of the failure rather than the noise of it, and what those models told me is the same thing the AI credit complex is telling me now: a crash is almost never a failure of technology. It is a failure of funding structure. Terra's algorithmic mechanism was elegant. Its backing was not. The elegance was the marketing; the backing was the risk.
The difference โ and it is a real one, not a rhetorical concession โ is that AI's collateral is tangible. GPUs exist. Power contracts exist. Land and interconnection queues exist. Crypto's 2021 collateral was largely reflexive, its value derived from the willingness of the next participant to accept it. But tangible collateral financed at the wrong tenor is still a duration mismatch, and a six-year GPU depreciation schedule financed with three-year money is a duration mismatch wearing an infrastructure costume. That is the invisible architecture of this boom, and it is the architecture the jitters are beginning to price.
Core: Crypto Is Now the Marginal Collateral for AI
Here is the part the equity-market commentary keeps missing, because equity-market commentary is not paid to look for it. Crypto is not adjacent to the AI credit cycle. It has become one of its financing channels, and it has become so through mechanisms that most crypto analysts describe as bullish and most credit analysts do not describe at all. There are four transmission paths, and they interlock.
The Balance-Sheet Moat
Start with what the market actually did when the jitters arrived, because the reaction is more informative than the cause. Nvidia's balance sheet is a fortress: net cash, gross margins north of seventy percent, no meaningful leverage, and the ability to fund its own ecosystem out of pocket if it chooses. Oracle's is the opposite profile โ a backlog concentrated in a small number of counterparties, free cash flow turning negative as capital expenditure outruns it, and two of the major rating agencies revising its outlook to negative within weeks of each other. Broadcom sits in between, carrying the sixty-to-seventy-billion-dollar debt load it absorbed with VMware while simultaneously funding a custom-silicon roadmap.
The market's discrimination is not irrational. It is the oldest pricing mechanism in finance, applied to a new asset class: it is pricing who can afford the entry ticket. This is the same dynamic I have argued governs the exchange landscape, where a multi-billion-dollar penalty did not weaken the largest venue so much as entrench it, because the penalty functioned as an entry fee and the license it preserved became the moat. Scale and a fortress balance sheet are the only durable advantages in a capital-intensive industry, and a regulatory or financial barrier that costs billions to cross is not a barrier to the largest players. It is a gift to them.
The market is no longer pricing AI demand; it is pricing the funding structure behind that demand, and the spread between self-funded and debt-funded participants has become the single most informative signal in the complex.
The Miner-to-Neocloud Conversion
The most direct coupling between crypto and AI credit runs through the bitcoin miners, and it is far more advanced than the market's vocabulary admits. Core Scientific negotiated a merger with CoreWeave that its own shareholders rejected in October 2025 โ a decision that revealed something important: miner equity holders preferred the optionality of hashprice to the certainty of a contracted lease, because the optionality still carried upside and the lease carried only the counterparty's credit. TeraWulf leased capacity to Fluidstack with Google taking a stake and backstopping billions in lease obligations. Galaxy Digital signed a fifteen-year lease worth roughly four and a half billion dollars to CoreWeave for its Helios campus. IREN contracted with Microsoft for a deal reported near ten billion dollars. Cipher, Hut 8, and a lengthening list of others are converting megawatts into machine-learning capacity as fast as interconnection queues allow.
What is actually happening beneath the press releases is a swap. Miners are trading a volatile commodity margin for a contracted, credit-linked one, and in doing so they are importing AI credit risk directly onto crypto-native balance sheets. The equity market rewarded the swap generously, awarding an HPC multiple to businesses that had previously been valued on hashprice and power cost. But when the credit jitters arrive, the same convertible notes and secured facilities that funded the conversions reprice first, and the HPC premium compresses fastest โ because it was always a credit premium wearing an infrastructure label.
The miner-to-neocloud trade is not a diversification away from crypto risk; it is a leveraged long position in AI credit, dressed as a pivot to contracted revenue.
Underneath the leases sits the collateral question nobody wants to answer precisely: what is a GPU worth in year four? The depreciation debate that flared in late 2025 โ six-year schedules defended by the industry against two-to-three-year schedules advanced by skeptics โ is not an accounting footnote. It is the load-bearing assumption of an enormous private credit complex. If residual values are overstated, the collateral behind a large slice of that debt is overstated, and the loans that looked overcollateralized at origination look undercollateralized exactly when refinancing is needed. In 2024, working with three analysts on a forty-page study of Bitcoin's correlation with Swedish government bond yields through the ETF approval window, we demonstrated that institutional adoption had decoupled crypto from tech-sector beta. What we did not claim, and what the market has since forgotten, is that decoupling from equity beta is not decoupling from credit beta. We are waiting for the market to reveal its true cost, and the reveal is not a price. It is a mark.
Stablecoins as the Settlement Rail
The second transmission path is quieter, and I believe it is the one that will matter most by the end of the decade. AI compute is metered, billed, and increasingly paid programmatically. Machine-to-machine transactions do not want a settlement asset that takes two days to clear, that closes on weekends, or that requires a correspondent bank in the loop; they want an asset that is instant, programmable, and denominated in the unit in which compute contracts are written. That asset is a stablecoin, and its adoption as an operational rail rather than a speculative instrument is the part of the story that sentiment has not yet caught up to.
The Helsinki pilot I documented earlier this year โ automating utility settlement through smart contracts, with no human in the loop and no bank in the middle โ was a small proof of a large claim. As autonomous agents begin to transact on behalf of institutions, the demand for programmable money stops being a crypto-native curiosity and becomes operational infrastructure. The payment-rail experiments from the card networks and the agentic payment standards emerging from the largest exchanges are all pointing at the same destination. The stablecoin float, now measured in the hundreds of billions and growing, is quietly becoming the settlement layer for machine-to-machine compute.
That has a consequence the market has not priced. Stablecoin reserves are held predominantly in short-duration sovereign debt, which means the float's backing and the AI buildout's financing are now leaning on the same collateral base โ short-dated government paper โ from two directions at once. Stablecoin float is no longer merely crypto's internal liquidity gauge; it is becoming the settlement rail for an AI economy, which couples its collateral base to the same duration the AI capex cycle depends on.
Tokenized Private Credit and the Reflexive Loop
The fourth path is the least discussed and the most reflexive, because it closes the loop between the two systems. The AI credit itself is coming onchain. Tokenized private credit vehicles are onboarding precisely the asset class now under scrutiny, and the tokenized T-bill complex has already demonstrated that institutional buyers will accept a blockchain wrapper when the underlying is familiar. The implication is uncomfortable and, I think, inevitable: if a meaningful share of AI infrastructure debt is tokenized, then crypto's search for real yield terminates, in part, in AI capex credit risk.
Trace the loop and it becomes difficult to find a clean seam. A crypto investor buys a tokenized credit fund for yield; the yield is serviced by datacenter leases; the leases are underwritten against offtake from firms whose credit is now being questioned; the offtake is denominated in dollars settled through stablecoins; the stablecoins are backed by short-duration sovereign paper that the same rate environment reprices. Every link is defensible in isolation. The chain is not. This is the structural truth that the decoupling narrative obscures: crypto did not sidestep the AI credit cycle. It became the venue where that cycle receives its most liquid, most reflexive, and most continuously priced expression โ which is a privilege in the expansion and a liability in the contraction.
The Contrarian Angle: Decoupling Is Structural, Not Tactical
The consensus in crypto right now has two halves, and both are wrong in the same way. The first half insists that AI credit jitters are a TradFi problem, that Bitcoin is a non-correlated reserve asset, and that capital rotating out of overvalued AI equities will find its way into hard, fixed-supply money. The second half insists that this is 2000 all over again, that the AI bubble is bursting, and that crypto will be dragged down as collateral damage. Both halves share an assumption that should be examined rather than absorbed: that crypto is a passive recipient of whatever AI's credit cycle decides to do.
The data hides what the eyes refuse to see. Crypto is neither insulated nor merely collateral damage. It is the marginal funding channel, and marginal funding channels do not get to watch from the sidelines; they are the first to be asked for more margin and the last to be told why.
Where the bulls are right โ and they are right about something important โ is the horizon. Over three to five years, the case for Bitcoin as a genuinely non-correlated reserve asset is strengthening, not weakening, and it is strengthening for a reason that has nothing to do with AI: the sovereign-debt dynamics that make a fixed-supply, bearer, settlement-neutral asset useful are deteriorating, not improving. The 2024 study I helped produce was a structural claim about institutional adoption, and that claim survives contact with the current cycle.
Where they are wrong is in mistaking a structural decoupling for a tactical one. In a genuine liquidity event, correlations do not remain politely low; they converge toward one, because everything is ultimately priced in the same dollar and funded through the same plumbing. Bitcoin's correlation to the Nasdaq is not the variable that matters in a stress scenario. Its correlation to dollar funding conditions is, and that correlation approaches one at precisely the moments when diversification is most needed and least available. Decoupling is a property of the regime, not of the week.
The blind spot, then, is not that the market is ignoring AI credit risk. It is that the market is pricing AI credit risk as though it were exogenous to crypto โ a storm on someone else's horizon โ when crypto has spent three years volunteering to be the place where that storm makes landfall. Waiting for the market to reveal its true cost is not passivity. It is the discipline of refusing to pay for clarity that has not yet arrived, and it is the only stance that respects both the structural case and the tactical reality without pretending they are the same thing.
Takeaway: Positioning for the Funding-Structure Repricing
The question is not whether AI is real. It is real, and the compute will be built one way or another, because the demand is genuine and the strategic stakes are too high for the buildout to stop. The question that matters for positioning is narrower and harder: who holds the duration when the funding structure is repriced โ and the honest answer is that crypto has volunteered, through converted miner balance sheets, through tokenized credit products sold on yield, and through a stablecoin float that now shares a collateral base with the very capex it settles.
Positioning follows from that recognition rather than from any forecast about AI demand. In a bull market, the temptation is to read every AI headline as a crypto tailwind; the discipline is to read it as a balance-sheet question instead. Watch the markers that precede a credit repricing rather than the price action that follows it: private-credit fund gates and the timing of their NAV marks, the GPU residual-value assumptions embedded in each new facility, the 2027 and 2028 maturity wall of miner convertibles, and above all the widening spread between self-funded and debt-funded AI participants โ the same spread that Nvidia's indifference made visible on a Tuesday when nobody was looking. In a cycle defined by the cost of capital, the balance sheet is not a footnote to the thesis; it is the thesis.
The forward-looking question is not whether the AI buildout survives the repricing of its own funding. It is this: when that buildout is finally refinanced in the open market, will crypto be the asset that absorbs the duration โ or the market that finally prices it? The data hides what the eyes refuse to see, and the answer will arrive as a mark, not a headline.