Three weeks ago I pulled the 10-Q of a hyperscaler and found something that stopped me cold. A $2.1 billion "customer contract" booked as committed revenue traced back through two SPVs and a convertible note to a fund the same hyperscaler had seeded nine months earlier. Money left the left pocket, promised to buy compute from the right pocket, and returned as revenue. I have seen this movie. In 2022 I watched a $500 million algorithmic peg do the same thing with UST, and I shorted it three weeks before the death spiral. Code doesn't lie. Balance sheets, when they are structured like this, do.
Circular financing is not new. Crypto built it, scaled it, and burned it down twice before AI borrowed the blueprint. In 2021, Alameda funded projects that funded FTX-adjacent ventures that parked tokens back on Alameda's books as collateral. In 2020, DeFi yield farms paid depositors in tokens minted by the same protocol, which depositors sold into liquidity the protocol itself had seeded. The APY was real. Yield is just delayed volatility.
The AI version runs on the same rails with cleaner branding. A chipmaker invests in a lab. The lab commits to buying that chipmaker's silicon on multi-year contracts. The chipmaker books the contract as backlog, the stock re-rates, the lab's paper valuation climbs, and a private credit fund lends against that valuation to fund the next round. Every node looks solvent in isolation. The system is solvent only if the loop never slows.
The difference from crypto is scale and opacity. Crypto's loops ran on-chain, where I could watch wallets settle in real time. AI's loops run through special purpose vehicles, private credit facilities, and vendor financing agreements that never touch a public tape. That is not less dangerous. It is more. The surveillance infrastructure that let me front-run Terra does not exist here, and the absence of a visible order book is exactly what makes the risk unpriceable.
Let me break the loop into its three moving parts. Each is a lever, and each amplifies the others.
First, revenue recognition. When a vendor is also an investor, the contract it books is a related-party transaction. Under GAAP, related-party revenue must be disclosed, but the disclosure often sits in a footnote, buried beneath a headline number. The metric that matters is not total backlog; it is the ratio of related-party backlog to third-party backlog. When that ratio climbs, the growth is internal, not market-verified. I watched the same distortion in 2021, when NFT volume on a marketplace I was arbitraging included wash trades between wallets controlled by the same collector. Volume metrics are deceptive without holder distribution analysis. Backlog metrics are deceptive without counterparty analysis.
Second, the depreciation mismatch. A GPU has an economic life of three to five years under heavy load. A multi-year compute contract is priced assuming the hardware stays productive for its full term. If utilization drops below roughly 60%, the fixed cost of the contract exceeds the revenue it generates, and the lab is underwater on paper before it is underwater in cash. That is the exact structure that killed Celsius: long-dated liabilities, short-dated assets, and a duration gap that only closed one way. Yield is just delayed volatility, and so is backlog. Both are promises that resolve on a schedule the seller does not control.
Third, the leverage channel. The public market sees the equity. It does not see the private credit that funded it. Structured products and SPVs let the same underlying compute contract be pledged as collateral in three different facilities, each of which believes it holds senior claim. I saw this in 2017, auditing the vesting schedule of an ICO that let early whales extract 20% of supply through an integer overflow. The exploit was invisible in the whitepaper and obvious in the Solidity. The modern equivalent is a warehouse facility nobody outside the arranger can inspect. Smart contracts are brittle, but at least they are auditable. A private credit covenant is neither.
Here is where my applied mathematics background earns its keep. Model the loop as a feedback system. Let capital flow C into a lab, which converts to compute demand D, which converts to vendor revenue R, which converts to valuation V, which converts back to available capital C'. The loop is stable while the gain on each edge exceeds 1. The moment terminal demand fails to grow faster than the loop compounds, the gain drops below 1, and the system flips from self-reinforcing to self-destroying. I ran this model on UST in early 2022 and calculated that a $500 million outflow would break the peg. The peg broke in four days.
The AI loop has a longer fuse because the capital is stickier and the assets are physical. But the fuse is finite. Datacenters take eighteen to thirty-six months to build, and the demand signal that justifies them is verified quarterly. If growth slows even one quarter during construction, the industry builds into a glut, and GPUs booked at premium rental rates find no bid on the secondary market. Arbitrage hides in plain sight: the spread between contracted compute prices and spot rental prices is the market's only real-time tell on whether the loop is still inflating. When spot prices fall below contract prices, the loop is contracting, and the paper revenue is already stale.
The stock market rewards the loop. Every contract signed becomes backlog, every backlog becomes a growth narrative, and every growth narrative becomes a lower cost of capital. The lower cost of capital funds the next investment, which signs the next contract. This is a reflexive system, and reflexivity runs until the underlying reality can no longer support the perception. The reality here is inference revenue, the actual dollars enterprises and consumers pay for AI outputs. That number is growing, but not at the rate the loop's capital deployment implies.
The physical layer is where this gets genuinely ugly. Compute demand propagates up a supply chain, through GPUs, HBM memory, optical transceivers, transformers, and grid interconnects, and each link has its own bullwhip. A 20% demand miss at the lab becomes a 40% order cancellation at the chipmaker and a 60% collapse in the electrical equipment backlog two tiers up. Power and land are asset-specific: a datacenter built for dense GPU workloads cannot be repurposed into a warehouse, and a substation built for a hyperscaler cannot serve a suburb. When the loop contracts, the write-down lands on physical assets with decade-long depreciation schedules. That is a slower, deeper loss than any token crash, because you cannot liquidate a substation in a weekend.
Concentration is the final amplifier. A handful of entities sit at the center of the network: the chipmakers, the hyperscalers, the largest labs. Counterparty risk is not diversified; it is stacked. In 2008, AIG sat at the center of the CDO web and its failure threatened the whole system. Here, the center nodes are the same entities providing the capital, booking the revenue, and supplying the compute. Measures what matters, not what feels good, and what matters is the correlation between these nodes, which rises to nearly one under stress. Diversification across AI names is an illusion when they all clear through the same few balance sheets.
The consensus now is that AI circular financing is a bubble about to pop. I disagree with the timing, and I think the consensus is looking at the wrong number. Bubbles do not pop when the structure is fragile; they pop when the funding stops. Crypto taught me this. Terra was structurally broken for months before it died. The peg held because new capital kept arriving. The trigger was not the fragility. The trigger was a withdrawal.
So the question is not whether the loop is fragile. It is what stops the funding. For AI, that is not a sentiment shift; it is a credit event. Watch the private credit spreads on AI-adjacent structured products. When those spreads widen even 100 basis points, the marginal facility gets repriced, the marginal lab cannot fund the next round, and the loop contracts on schedule. Equity investors will see it last, because equity is the residual claim on a structure they cannot fully inspect. Survival beats speculation, and survival here means tracking the debt, not the narrative.
There is a second blind spot. The bears assume the loop is uniformly distributed. It is not. Some labs are funded by genuine terminal demand; others are pure financial engineering. The distinction is invisible in the headline valuation and visible only in the related-party disclosure. Most analysts will not do that work, so the mispricing persists.
Here is what I am watching, with levels. Contract-to-cash ratios above three to one on any lab are a red flag. Spot GPU rental prices below contracted prices for two consecutive quarters is a confirmed contraction. And a 100 basis point widening in AI private credit spreads is the starting gun. The loop is not a mystery. It is a duration mismatch wearing a growth story, and duration mismatches always resolve. The only question is whether you are the one holding the paper when it does.