SKYWAY SUMMIT
Investments · Topic Brief
Anatomy of the AI Bezzle
Research conducted by investor Michael Burry, published 22 May, makes a figure-dense argument: not that AI is fake, but that a large share of today's demand is temporary, financed, and counted as permanent. The findings span a dangerously concentrated customer base, a supply chain that cannot bend, a manufactured demand phase termed "tokenmaxxing," and a funding chain that ends, literally, in Bermuda.
Two-thirds of receivables sit in three customers
The starting point is structural, not sentimental. Michael Burry's research shows NVIDIA's single largest customer represents roughly 30% of accounts receivable, and the top three together account for 64%, a figure that jumped eight points in a single quarter. The trajectory is the story: a balance sheet that was reasonably diversified in 2020 is now acutely dependent on a handful of buyers.
The zig where there were only zags
Over recent years NVIDIA's revenue grew 4.9× and its total receivables grew 4.9×, in lock-step, as you'd expect. But the largest customer's receivable balance grew 13.4×. In the latest quarter, for the first time in thirteen quarters, that customer rose as a share of receivables (to ~30%) while falling as a share of revenue (to ~21%). Receivables and revenue moved opposite directions for the one buyer that matters.
$182 billion in commitments that can't be cancelled
Against a $5 trillion market cap, NVIDIA carries $182 billion of forward purchase commitments, $119 billion of it tied to a single customer, a sum that exceeds its annual operating cash flow. These are custom, non-cancellable lines at TSMC: non-fungible. That is the crucial contrast with Cisco, which wrote down roughly half its commitments after a ~15% revenue dip in 2001 but whose components were fungible and sold to many other buyers. When demand wobbles in a concentrated, non-fungible system, the move is not gentle.
operating cash flow
"Tokenmaxxing": a training binge mistaken for a trend
A token is the unit a language model bills on, roughly three-quarters of a word; about 1,500 words costs ~2,048 tokens. The bull case says insatiable token consumption proves durable demand. The counter is that much of it is tokenmaxxing, quota- and leaderboard-driven overconsumption during a one-time training and benchmarking phase. The cited anecdotes are specific: one investor logged 925 million tokens in two weeks for ~$320 and mapped a path to 250M/day; METR finds models can now run autonomously ~12 hours, up from ~1 hour a year earlier; Meta engineers competed on token leaderboards (since removed); a Meta executive described a top engineer "spending the equivalent of his salary in tokens"; NVIDIA's CEO floated annual token budgets worth half an engineer's base pay.
The reason it must reverse is structural. Enterprises sit in a pyramid, and the pressure at every level is to migrate up, harvest your own usage data, build in-house, and cut third-party token spend. The reversal is termed compression.
Does Jevons Paradox rescue it, cheaper compute unlocking a wave of new users? Not in the near term, the letter argues: there are no new users left to unlock. Adoption is already near-universal and saturated; the analogy offered is Excel, on nearly every PC for four decades yet rarely used past a fraction of its capability. And the binding constraint underneath all of it is power.
S&P Global projects a ~19 GW shortfall, 40% of need, for U.S. data centers by 2028. Whether AI demand proves permanent or temporary, the shortage of electrons to feed it is the one fact bulls and bears agree on, and energy scarcity is terrain Skyway Summit already maps.
Whose money funds the build-out?
This is where the analysis earns its length. Near-term build-out commitments approach $3 trillion, far beyond what even the four largest hyperscalers can fund from cash flow. So the marginal dollar must come from debt or off the balance sheet. First, the accounting softens the optics: useful-life assumptions stretch depreciation across hardware that may be obsolete in under a decade, which the letter argues overstates earnings by ~20% at Oracle and Meta.
| Company | Useful-life assumption (2025-26) | Move |
|---|---|---|
| Amazon | 5.0 yr | shortened |
| Microsoft | 6.0 yr | , |
| Alphabet | 6.0 yr | , |
| Oracle | 6.0 yr | earnings ~20% high |
| Meta | 5.5 yr | extended |
GPU clusters go obsolete in a few years; the data centers housing them are financed over 15-19. That duration mismatch is the seam the whole structure lives in.
Mechanism one, NVIDIA helps fund its own demand
Mechanism two, the dollar takes an offshore detour
Long-dated data-center debt needs a long-dated buyer. Life insurers fit perfectly: they sell long-duration annuities and need long-duration assets to match. So the chain runs from American savers, through a U.S. insurer, to a Bermuda captive reinsurer, and into the AI build-out.
Many private life insurers doing this are owned by private-equity firms, Apollo's Athene is the archetype, which use the captive as a home for illiquid credit. Michael Burry's research shows the concentration plainly: a count of Bermuda Monetary Authority filings finds eight alternative-asset managers account for 36 of the ~200 captives created since 2023.
Real demand, but dangerously concentrated, temporarily inflated by a training phase, expressed through a non-fungible supply chain, and financed through an increasingly fragile, off-balance-sheet stack. The thesis is structure, not story. The timing is deliberately undefined, described as "a finger on the trigger," resolved by filings, not a date.
- NVIDIA's next 10-Q, does the largest customer's receivables-vs-revenue divergence persist or widen?
- The OpenAI S-1, an expected IPO filing should expose demand quality at the most leveraged link in the funding chain.
- The Microsoft Copilot cutoff (30 June), and whether other large enterprises follow the same adopt-then-compress path.
- Bermuda BMA filings, captive-reinsurer leverage and related-party activity as the off-balance-sheet stress gauge.
A repeatable process, not a hunch
- Start in the primary documents. The whole thesis is built from filing footnotes others skim, receivables by customer, depreciation assumptions, forward-commitment lines, Bermuda condition reports.
- Reason by mechanism, not vibe. Not "it's a bubble" but "here is the exact mechanism that broke Cisco and Sun, does this case carry it?"
- Invert the strongest bull argument. It engages "insatiable demand" and "Jevons" directly and turns each into evidence against the position.
- Blame incentives, not conspiracy. Every actor is rational; fragility is emergent, all of them underwriting one temporary signal.
- Be honest about timing. Define the observables that confirm or kill the thesis; refuse to pretend you know the date.
This brief summarises and interprets a published investor letter by Michael Burry for research and educational purposes. All figures are Burry's own calculations and claims, drawn from his letter and the sources he cites (S&P Global, the Federal Reserve, the Bermuda Monetary Authority, and press reporting); they have not been independently verified by Skyway Summit. Nothing here is investment advice or a recommendation to buy or sell any security. Source material ages and markets move, verify current data before acting.