Selling Shovels to Ghosts: The AI Capex Loop and the Crypto Credit Loop

Last week, a sell-side research desk at one of Asia's largest banks told the market that AI stocks are not in a bubble. The claim travelled through three hands before it reached a crypto audience: an earnings headline, a bank's reassurance, a repost on a digital-asset news wire. By the time it arrived, it had shed its date, its fiscal quarter, the name of the analyst who wrote it, and every number that might have made it falsifiable. What remained was a single sentence — not a bubble — with no anchor and no expiry. I have learned to read such sentences the way I once read volatility-clustering models: not for what they say, but for what their shape reveals about the people saying them. A text that carries no date does not want to be tested. A text that carries no valuation multiple does not want to be priced. And a crypto outlet that amplifies the 'not a bubble' view while ignoring the 'bubble' view is not reporting a debate. It is curating a mood. Why should a fund manager whose mandate is digital assets care about a bank's opinion on Nvidia? Because in 2026, crypto is no longer a parallel system. It is the levered end of the same curve. Let me lay the map. Over the past eighteen months, the marginal dollar of global liquidity has stopped being a monetary-policy variable and become a capital-expenditure variable. The hyperscalers — Microsoft, Google, Amazon, Meta — have guided to combined AI capital spending on the order of three hundred billion dollars, a figure that now functions as the single largest discretionary investment programme in the private economy. Nvidia's quarterly earnings are not, in this frame, a company report. They are the reading of that programme. When the market says Nvidia beat, it means the capex cycle is still expanding. When it says Nvidia missed, it means something much larger is turning. That is the first reason the Singapore reassurance matters. It is not a claim about a chipmaker. It is a claim about the durability of the entire liquidity impulse that has, for two years, lifted every risk asset — equities, credit, gold, and, yes, crypto. The second reason is structural. The text I am describing is a three-layer transmission: an earnings event, a bank's interpretation, a crypto wire's repost. Each layer can attenuate or distort the one beneath it. I have spent sixteen years watching this attenuation. In early 2017, as a junior quant in Stockholm, I submitted an anonymous report to three crypto newsletters predicting the liquidity traps ahead of the ICO boom. I learned then that the distance between a primary fact and its popular retelling is where most investors lose their money. The retelling is smoother than the fact. It is designed to be. Here is what the smooth retelling omits. The bank in question is a seller. Its research division sits inside an institution whose wealth-management arm holds substantial AI equity exposure. A 'not a bubble' call is not neutral; it protects the client base that generates the fees. I am not accusing anyone of bad faith. I am noting that the incentive gradient of a sell-side desk points, by default, toward reassurance. The question is never whether the analyst believes it. The question is whether the belief has been stress-tested. It has not. The claim contains no valuation multiple, no discount rate, no terminal-growth assumption, no invalidation condition. It does not tell you what would have to happen for it to be wrong. And a macro judgement that cannot be falsified is not a judgement. It is a mood with a Bloomberg terminal. That is the context. Now the analysis. Start with the logic. The argument runs: Nvidia's earnings were strong, therefore AI demand is real, therefore AI stocks are not expensive. Read it twice. The first clause describes the past. The second describes the present. The third describes the future, and it does not follow. This is the oldest error in the market, and it has a name most people have forgotten because the company survived long enough to become boring. Cisco. In March 2000, at the exact peak of the dot-com bubble, Cisco reported revenue growth of nearly sixty percent year over year. Its routers were not a fad; they were the literal plumbing of the internet. The demand was real. The technology was real. The company was excellent. And the stock still fell more than eighty percent, because the question was never whether the demand existed. The question was whether the price had already paid for a decade of it. A bubble is not defined by the absence of a business. It is defined by a price that has run ahead of the fundamentals that must eventually justify it. The Singapore desk collapsed that distinction. Nvidia's strength proves the demand of the last four quarters. It says nothing about whether the current market capitalisation — which crossed three trillion dollars in mid-2024, four trillion in July 2025, and brushed five trillion by October of that year — has already discounted the demand of the next sixteen. I have seen this exact reasoning before, and I have paid for it. In the summer of 2020, I audited the first liquidity-pool mechanisms of Uniswap v2 and Yearn Finance. I found that the yield-farming rewards were structurally unsound — the impermanent-loss math broke down in high-volatility pairs. I wrote a forty-page internal memo arguing for a hedged, stable-asset strategy instead of chasing APY. The firm ignored it and lost fifteen percent in two months. The lesson was not that the technology was bad. The technology was excellent. The lesson was that a real product and a sound price are two different questions, and the market is only ever confused about one of them at a time. Now the sharpest point, and the one the reassuring text will never touch. Nvidia does not merely sell to its customers. It invests in them. The company has taken positions in a constellation of AI-native firms — neoclouds, model labs, compute brokers — and those firms then use the capital to buy Nvidia's own GPUs. One of these arrangements, involving a large language-model developer, has been reported to carry forward compute commitments in the hundreds of billions of dollars against a revenue run-rate still measured in the low tens of billions. Sit with that structure. A supplier funds a customer. The customer returns the capital as revenue. The revenue is recognised, the margin is booked, and the cycle repeats. On the income statement, it looks like demand. In the cash-flow statement, it looks like a loop. I know this loop. I lived through its crypto twin. In May 2022, as TerraUSD unwound, I had to liquidate ten million dollars of algorithmic-stablecoin exposure to save what remained of my fund. I spent the following three months reading the governance files of Anchor Protocol and Terraform Labs. The mechanism was elegant. Anchor promised a twenty-percent yield on a stablecoin. The yield was funded, in part, by the very token the system was built to defend — a self-referential engine that looked like a yield farm and behaved like a pyramid. The protocol held, but the consensus fractured. It was never a technology failure. It was a governance failure dressed as a yield. The AI capex loop is not identical. It is larger, it is legal, and it is run by some of the most competent operators in the world. But the shape rhymes, and the shape is what matters. When a supplier finances its own demand, the revenue is real in the accounting sense and fragile in the economic sense. The moment the financing stops, the demand stops with it. There is no external verification. There is only the loop, turning, until it does not. The second thing the text omits is more technical and, to my eye, more damning. It is the question of depreciation. A GPU is a capital asset. Its cost is spread across its useful life. If a hyperscaler assumes that life is six years, the annual depreciation charge is smaller, reported profit is larger, and the AI business looks more profitable than it is. If the true useful life is three years — because the chips are obsolete, because the workloads have moved, because the next architecture is faster — then the profits are an artefact of an accounting assumption. Through the second half of 2025, this became the technical core of the bubble debate. Michael Burry — the investor who read the mortgage market before 2008 — took public positions and published arguments that the hyperscalers were extending depreciation schedules to flatter earnings. Whether he is right is an empirical question. But notice what the reassuring text does with it: nothing. A claim that AI is 'not a bubble' which never mentions the depreciation assumption is a claim built on sand. The entire profitability of the buildout hinges on a number that is chosen, not measured. Here is where my crypto background sharpens the reading rather than softening it. In DeFi, we spent years learning that the integrity of a system depends on the integrity of its least verifiable input: the oracle. A lending protocol is only as honest as its price feed; if the feed lags, the liquidations fire wrong, and the whole edifice clears at a fiction. The depreciation schedule is the AI trade's oracle. It is a single, unverifiable assumption on which an enormous structure of reported earnings depends. And, like Chainlink's federated nodes — decentralisation solved by centralisation — the assumption is dressed as a technical given when it is actually a human choice. Let me widen the lens to the supply chain, because the text's blindness here is instructive. The comforting narrative frames AI as a chip story. It is not. The real constraints, in order, are advanced packaging, largely at TSMC; high-bandwidth memory, sold out across SK Hynix, Samsung, and Micron; and — the binding one — electrical power. The grid connection, not the GPU, has become the rate-limiter on American data-centre deployment. Hyperscalers have begun signing nuclear power-purchase agreements precisely because the electron, not the transistor, is the scarce input. Why does this matter to a bubble call? Because a supply chain bottlenecked on electricity behaves differently from one bottlenecked on chips. It is slower to overbuild and slower to correct. It stretches the cycle. And it means that the true unit economics of AI — the cost per inference — are increasingly a function of energy prices, not silicon prices. The reassuring text, still narrating a chip story, is two bottlenecks behind the reality. And the third omission, the one that touches my mandate most directly: China. Export controls have cut Nvidia's China data-centre revenue from more than twenty percent of the total to a low-single-digit share, with multi-billion-dollar write-downs along the way. The current earnings growth contains no Chinese contribution. That is, in one sense, a de-risking — the geopolitical tail is already priced. But it is also a permanent amputation of total addressable market, a structural discount the valuation may not fully reflect. The text mentions none of it. The text's final omission is competition, and it is the one that should worry a long-term holder most. A healthy, non-bubble market is one in which excess profit attracts entry and margins mean-revert. Nvidia's gross margin has run above seventy percent for years — an anomaly in any manufacturing business. That margin is a standing invitation to its own customers to design around it. Google's TPU, Amazon's Trainium, Meta's MTIA, Microsoft's Maia: the four largest buyers are also the four most capable competitors, and their in-house silicon is already carrying a growing share of inference workloads. Broadcom, which supplies much of that custom silicon, has become the quiet hedge against Nvidia's own dominance. A genuine 'not a bubble' argument would have to explain why this margin survives the competition it is financing. The text does not try. So where does this leave crypto? This is the question I actually care about, and the answer is uncomfortable for anyone who still believes in decoupling. Over the past two years, crypto has been quietly reclassified. Bitcoin's spot-ETF approval in January 2024 did not merely open a new pool of capital; it changed the asset's identity. I led the integration of a fifty-million-dollar initial tranche into traditional portfolio allocations that year, working with a three-person team to navigate the SEC and MiCA frameworks. I saw the machinery from the inside. What the ETF did was make Bitcoin legible to the same allocators who own the Mag Seven — and legibility has a cost. It means Bitcoin is now traded, in the marginal flow, as a high-beta expression of the same risk appetite that funds the AI trade. The consequence is precise. Bitcoin's correlation to the Nasdaq is no longer a curiosity; it is the asset's dominant pricing input. When liquidity expands, both rise. When the AI capex narrative wobbles, both wobble. The 'digital gold' story survives in the marketing and dies in the correlation matrix. In the deep end, liquidity is the only oxygen — and crypto, at the levered end of the curve, suffocates first. There is a second-order effect that few are pricing. The AI trade and the crypto trade now compete for the same marginal dollar, because they are sold to the same investor with the same story: a technology that will reprice the world. When capital is abundant, both can win. When capital tightens, the allocator does not choose the better technology. The allocator chooses the better collateral. And for the institutional investor, Nvidia's earnings — audited, quarterly, legible — are better collateral than a token whose fair value no one can agree on. That asymmetry is the real risk to my book, and it has nothing to do with blockchain technology. It is a competition for attention and balance-sheet space, and crypto is currently losing it. One more thread, because it is the most honest signal in the entire debate and the text ignores it completely. In August 2025, an MIT research group published a widely cited finding that roughly ninety-five percent of enterprise generative-AI pilots had produced no measurable return on investment. The methodology is contestable, and the optimists contest it. But the number matters less than its function: it became the empirical battleground of the whole argument. If the application layer cannot generate cash, then the capex that funds the compute layer has no terminal value. The loop has no exit. This is where crypto's own history offers a warning the equity market has not internalised. In 2021, I managed a portfolio heavily weighted in NFTs — three pieces, a quarter of a million dollars, bought on the conviction that digital ownership was a new cultural paradigm. I watched the speculative frenzy detach from the artistic value, and then I watched the crash take sixty percent of the fund. Art was the asset, but attention was the currency. The lesson was not that digital ownership was fake. The lesson was that a market can be real and still be priced for a future that never arrives. Decentralised compute is the crypto-native version of the same question. Protocols like Render, Akash, and io.net promise to aggregate idle GPU capacity into a permissionless market. The thesis is elegant: if AI compute is scarce and centralised, then a distributed supply should command a premium. But the economics are brutal. The neoclouds and the decentralised networks compete on the same unit — cost per useful FLOP — against hyperscalers who own their power contracts and amortise their silicon at scale. And the decentralised networks, like the neoclouds, depend on the same financing conditions as the AI trade itself. When the loop tightens, they do not get a discount. They get cut first, because they are the least collateralisable borrowers in the chain. There is a temptation, in moments like this, to treat the crypto market as a refuge — the honest asset class, the one that never needed a depreciation schedule. That temptation is a trap. Crypto's own history is a graveyard of the same coordination failures, and its current correlation to the AI trade means it is not a refuge but a mirror. The difference is that crypto fails faster and more visibly, which is why we learned these lessons first. Pattern recognition is the only true hedge. Not the pattern of the chart — the pattern of the incentive. I have watched four cycles now, and the shape is always the same: a real technology, a genuine need, a price that runs ahead, and a structure of financing that hides the gap until it cannot. The ICO boom of 2017 had it. DeFi summer had it. The NFT market had it. Terra had it. The AI buildout has it now, and the only difference is the balance sheet it sits on. What makes this cycle harder to read is that the governance failure is distributed. In Terra, the failure had an address — a protocol, a foundation, a founder. In the AI capex loop, the failure has no single address. It is spread across a supplier's investment arm, a set of customer balance sheets, a depreciation assumption chosen by an accountant, and a power contract signed by a utility. No one is lying. Everyone is optimising locally. The system fails collectively. That is the modern shape of a bubble: not a fraud, but a coordination failure among rational actors. Here is the contrarian turn, and I hold it loosely because the evidence cuts both ways. The consensus, in every room I sit in, is that crypto follows AI. If the capex cycle breaks, the whole risk complex breaks, and Bitcoin leads the fall. That is the trade. But it assumes the correlation is fundamental — that the two assets are linked by shared cash flows. They are not. They are linked by shared liquidity and shared marketing. That is a much weaker bond, and weaker bonds break. When the AI trade unwinds — and it will, because every capex cycle does — the capital does not evaporate. It rotates. And the rotation has a destination the current market refuses to consider: the assets that were abandoned for AI. For two years, crypto has been the funding source for the AI trade, not its beneficiary. Retail attention, developer talent, and speculative capital all left for the shinier story. When that story stalls, the marginal dollar has to go somewhere, and the somewhere is the asset class that is cheapest, most liquid, and most hated. I am not predicting a crypto bull market from an AI bear market. I am refusing to accept the deterministic correlation. The two markets are coupled through liquidity and decoupled through fundamentals. The coupling is what everyone trades. The decoupling is what no one prices. Alpha is not found; it is harvested from chaos. So, positioning. In a sideways market, the chop is not noise; it is the mechanism by which the next cycle is priced. I am watching three signals and ignoring the rest: the hyperscalers' quarterly capex guidance, the depreciation disclosures buried in their filings, and the funding spreads of the neoclouds. When those three turn together, the loop is unwinding. Crypto will not lead that turn. It will amplify it — and then, if history is any guide, it will be the first thing to reprice the morning after. The question is not whether the bubble bursts. The question is what you own when the oxygen returns.