Reading the AI Bubble Through a Ledger: Duration, Circular Deals, and the Crypto Rotation Myth

Here is the reality of the story that landed in my feed last week: four sentences, zero numbers, and a source identified as "Castle Securities CEO Griffin." There is no Castle Securities I can verify against any registry I trust. There is Citadel Securities. There is Ken Griffin, its founder. The distance between those two names is the distance between a market-moving warning from the largest market maker on earth and a transcription error. I spent an hour trying to resolve which one it was. I could not. That is the first data point, and it has nothing to do with AI. It is a signal-integrity problem. The outlet that ran it was a Web3 news source. That matters. AI-bubble stories are the load-bearing wall of the "capital will rotate back into crypto" narrative, and outlets with that incentive frame the story toward collapse. So I read the piece twice — once for what it said, and once for what it was built to make me feel. What it said: enterprises are auditing their AI bills at Anthropic and at OpenAI. What it was built to make me feel: the AI trade is over, and the money is coming home. The first claim is worth taking seriously. The second is bad math. This article is about the gap between them. The AI bubble thesis has three load-bearing walls, and none of them are "AI doesn't work." Anyone selling you that version is selling a story, not an analysis. Wall one: the capex-revenue gap. Sequoia's David Cahn framed it as the "$600B question" — back out from AI infrastructure spend and you need roughly $600 billion in annualized end-market revenue to cover the capital cost. Hyperscaler capex alone crossed into the $400 billion range in 2025. The gap is not a rounding error. It is a structural mismatch between installed capacity and demonstrable throughput. Wall two: circular deals. A chipmaker invests in a model company. The model company commits to buying cloud compute. The cloud provider buys the chipmaker's silicon. The revenue closes a loop. This is legal. It is also, from a valuation standpoint, a machine for converting one dollar of capital into three dollars of "independent demand." Wall three: depreciation. GPUs get booked with a five-to-six-year useful life. The competitive life of AI silicon, driven by architecture iteration, is closer to two to three years. The spread between those two numbers is where profit quality goes to die. I have seen this exact geometry before. Not in AI. In DeFi, in the summer before the 2022 collapse, when I was mapping the on-chain ledgers of failed lending protocols and finding that $2 billion in locked assets had not failed because of smart contract bugs. They failed because of the disconnect between on-chain truth and off-chain data sources. Centralized oracle manipulation. The AI sector has the same defect. Its oracle is revenue recognition. And the revenue recognition is being fed by the same counterparties that book the capex. That's the frame. Now the engineering. The capex-revenue gap is a throughput problem, not a demand problem. I want to be precise about what the gap actually measures. It is not "nobody wants AI." It is that the installed base of compute is running ahead of the billable use cases that can pay for it. In DeFi terms: this is TVL without fee revenue. A protocol can show $2 billion in total value locked and generate $40,000 a day in fees. The TVL is real. The business is not. I have written before that "liquidity fragmentation" is a manufactured problem — a narrative VCs use to sell the next aggregator. The AI capex gap is the same move in reverse. The gap is real, but the story told about it is designed to sell you something: either the collapse trade or the rotation trade. Both are products. The mechanical question is throughput. How many tokens of inference can the installed base serve, at what price, to customers who can prove a return? That is a queueing problem. And queueing problems have answers. The current answer is uncomfortable: the queue of high-value, verifiable workloads is shorter than the queue of installed capacity. Circular deals are the AI equivalent of wash trading, and I have audited wash trading. When I mapped the Celsius and FTX ledgers, the thing that broke was not a function. It was an assumption about where the money came from. The protocols looked solvent because they were counting their own counterparties' promises as assets. The circular AI deal is the same structure with better paperwork. Chipmaker funds model company. Model company signs a compute commitment. Cloud provider books the commitment as backlog and orders more silicon. Every leg is a real contract. Every leg is also the same dollar, re-counted. I said it after the 2022 crash and I will say it again: we didn't have a smart contract bug. We had a data-integrity bug wearing a smart contract's clothes. The AI sector's circular deals are a revenue-integrity bug wearing a purchase order's clothes. The valuation consequence is specific. When you discount the AI revenue stack, you have to strip out the fraction that is internally financed. If a material share of the $600 billion needed revenue is being generated by counterparties inside the loop, the effective independent demand is lower, and the discount rate on the whole stack should be higher. Nobody publishing a bullish model is doing that subtraction. Auditing isn't about finding intent. It's about finding the assumption that breaks first. Depreciation is where the accounting becomes engineering fiction. The GPU useful-life assumption is the most under-discussed number in the entire sector. Five to six years is the book. Two to three years is the competitive reality, because the moment a new architecture lands, the old silicon does not stop working — it stops being worth running, because the cost per token on the new part undercuts it. That is not a bug in the model. It is the model. And when a company stretches the useful life to protect earnings, it is doing exactly what a protocol does when it marks an illiquid reserve at the last traded price. The mark is defensible until the day it isn't. The ledger doesn't lie about duration. Only the depreciation schedule does. This is the same discipline I applied in 2017, when I sat in a co-working space in Austin and manually audited the transfer logic of fifteen ERC-20 launches. I found integer overflows in three of them. The code was law, but human error was the bug. Here the human error is an accounting assumption, and the overflow is a profit number that will not survive contact with a two-year hardware cycle. The enterprise bill review is the most credible thing in the whole story. Seat-based pricing at roughly $30 per seat per month is easy to approve and impossible to defend in an audit, because an idle seat produces no measurable output. Per-token pricing is the opposite: cheap to start, unpredictable at scale, and terrifying to a CFO the moment an agent makes twelve tool calls to answer one question. Both models lack defenses under audit. That does not mean spending stops. It means spending sorts. The workloads that survive are the ones with a verification mechanism — compilation, tests, CI, a measurable ticket-resolution rate. Software engineering survives because it has a compiler that tells you whether the output is correct. Customer support survives because you can measure handle time and first-contact resolution. The workloads that do not survive are the ones with no oracle for correctness: general-purpose copilots, creative generation, executive decision support. These are the seat-heavy products. These are the ones getting cut. This is the structural divergence the original article missed. It framed the audit as a general verdict on AI. It is not. It is a verdict on unverifiable AI. The distinction is the whole ballgame, and it is the same distinction I have been making about blockchain for eight years: a system that cannot prove its own output is a system that cannot be audited, and a system that cannot be audited cannot hold a premium. One more mechanical note the original piece got backwards by omission. It named two vendors as if that were an accusation. Enterprise API spend is highly concentrated — Anthropic and OpenAI together hold more than half of it. Concentration is not a risk signal in this case. It is a moat. In an audit, the buyer does not switch from a concentrated supplier to a fragmented one; it switches from many pilots to one or two primary vendors. The audit accelerates consolidation. The head two win. The thin-wrapper application layer — the SaaS shells with no proprietary data — is where the damage lands. The most interesting consequence of the ROI audit is that the surviving workloads are exactly the ones where cryptographic verification becomes an economic input rather than an ideology. I run a project called Verifiable Truth. We prototype zero-knowledge proofs for training-data provenance — proving that a model's output traces to authentic source material without exposing the source. For two years this read as philosophy. Under an ROI audit, it reads as a line item. If a CFO cannot verify that a model's output is grounded in real data, the model's output cannot be booked as a reliable business input. Provenance stops being a values statement and becomes a control. The open-source price anchor is the second force. Llama, Qwen, DeepSeek — the gap to closed flagship models on mid-complexity tasks keeps narrowing, while the cost gap stays at ten to thirty times. The natural response to a bill audit is not switching vendors. It is tiered routing: high-value tasks to the closed flagship, long-tail tasks to an open model or a fine-tuned small model. I have watched this movie in DeFi. The router always wins, because the router captures the spread between the best venue and the lazy venue. The model layer is about to become a routed market, and routed markets compress margin at the endpoints. The real price ceiling is not in the model layer at all. It is in the cloud layer. If enterprises move to self-hosting and fine-tuning, the biggest beneficiary is the hyperscaler — Bedrock, Vertex, Azure AI — not the open-source model vendor. The model layer gets squeezed from both sides: commoditized from below by open weights, taxed from above by compute rent. That is a structurally worse position than the market currently prices. The dominant story in my feed is that when AI reprices, capital rotates into crypto. I want to test that mechanically, because it is the single most expensive assumption retail is currently holding. First: the capital base is shared. The same funds, the same family offices, the same retail brokerage accounts that own AI infrastructure exposure also own crypto. Rotation is not creation. If AI infrastructure reprices 50 to 70 percent — and that is the scenario range if rates do not fall and growth decelerates from triple digits to sub-50 — the margin calls land on the same balance sheets. Crypto is not a lifeboat. It is the higher-beta asset in the same portfolio. Second: the duration mismatch is worse in crypto, not better. When credit tightens, the first thing sold is the asset with no cash flow and the longest duration. That is crypto before it is anything else. Third: flow follows fear, but only if the protocol holds. In a liquidity cascade, crypto's market structure — thin order books, reflexive liquidations, funding-rate spirals — does not hold. The 2022 comparison is not rhetorical. When Celsius and FTX broke, the capital did not rotate into "quality crypto." It left the asset class. The honest version of the rotation thesis is narrower and less exciting: a small share of speculative capital may rotate, and the beneficiaries will be the assets with verifiable cash flow and real fee revenue, not the long tail. The rest is a story sold to people who need the bubble to be someone else's problem. And the part nobody wants to hear: the AI bubble being real does not make the technology fake. Carlota Perez's installation phase is defined by overinvestment and financial-capital dominance, and it is followed by a turning point and a deployment phase, not by the technology disappearing. Railways overbuilt. Telecom overbuilt. The internet overbuilt. All three deployed. The distinction between "this is overvalued" and "this is worthless" is the difference between a repricing and a funeral. The market is currently pricing the funeral into the long tail and the repricing into the leaders, and most people are holding the wrong side of that trade. The node to watch is the IPO window. If a leading model company lists in 2026 or 2027, that is the first time public markets get to mark private valuations. Everything upstream — the circular deals, the depreciation schedules, the capex guidance — gets tested against a live bid. That single event will tell you more about the real discount rate on AI revenue than every analyst note published this year. My position is unchanged and it is not a prediction. The gap is real, the duration is mispriced, and the rotation trade is bad math. What survives the audit is what can prove its own output. Code is the only law that doesn't negotiate with a CFO. Silence is the loudest audit trail in the market. Right now the AI sector is very loud. That is the tell.