Gotta Dance

THE NUMBER: $500 billion — and it’s all priced on one number Nvidia itself gets to set. Six of the biggest firms on Wall Street signed up to move over half a trillion dollars into AI compute, and the entire structure rests on a single assumption: how long a GPU stays productive. The hyperscalers book five-to-six years (Amazon 5, Meta 5.5, Google 6). Michael Burry says two-to-three, and if he’s right the industry is overstating profits by something like $176 billion between now and 2028. Jensen Huang writes the chip roadmap that decides who’s right. And he’s offering to backstop 25% of the bet. Read that twice. The man who controls the depreciation schedule is also insuring the collateral. Everything below is about what that means, why it’s a genuinely smart move, and why the lender is still the one who eats it.

The trade Jensen actually made

Larry Fink stood on CNBC Monday, flanked by Jensen Huang and the CEOs of Goldman Sachs, Blackstone, Apollo, KKR and Brookfield, and compared what they’d just built to the day he started the mortgage-backed securities market in the 1970s. He meant it as a boast. “The next future for financial engineering,” he called it. He’s right that it’s the next future. He just told on himself about which future.

Here’s the deal, stripped of the press release. The six firms signed memos of understanding (not contracts, memos) to mobilize more than $500 billion of third-party capital to finance AI “factories.” The pitch, in Huang’s own words, is that a rack of Nvidia GPUs is no longer a fast-rotting piece of electronics. It’s “an investable asset, an infrastructure asset” — productive, revenue-generating, fungible, financeable like a toll road or an apartment building. Compute is revenue, and revenue can be borrowed against.

The money isn’t Nvidia’s. That’s the whole point. The first three years of this buildout got paid for out of hyperscaler balance sheets, and some of those balance sheets are now cash-flow negative from the spending. Nvidia’s own $100 billion pledge to OpenAI last year never actually materialized — it turned into a $30 billion check into the funding round. So the balance-sheet era is tapped out, and Jensen went looking for someone else’s checkbook. He found the biggest pools of long-duration capital on earth: pensions, insurers, private credit, the roughly $9 trillion sitting in U.S. money-market funds. At $50 to $60 billion a gigawatt and something like 70 gigawatts of demand in the U.S. alone, the math only closes if you can reach that money. This is how you reach it.

One problem, and it’s the problem the whole thing is built to paper over. A toll road doesn’t get obsoleted by a better toll road every three years. A GPU does. Bankers have never really lent against chips, because chips fall off a cliff the moment the next generation ships. So Jensen solved it the only way a chip vendor can. He offered to backstop up to 25% of the depreciation himself — a residual-value guarantee on his own product. And he can offer it, because he’s the one who decides when the successor chip ships. He isn’t guessing at the depreciation curve the way a bank would. He’s drawing it.

That’s not a footnote. That’s the trade. Nvidia is underwriting a risk it partly controls, on collateral it manufactures, to move product it can no longer sell fast enough on customers’ own credit. Brilliant, honestly. Also the exact moment the underwriting stopped being independent.

The one number the house gets to set

Everything about whether this is a smart loan or a landmine comes down to one variable: useful life. Kyle Reidhead laid the math out cleanly. On every $100 billion of GPU capex, a three-year life means $33 billion a year of depreciation. A six-year life cuts that to $17 billion. An eight-year life with resale value gets you to $10 billion. So the gap between Burry’s world and Jensen’s world is north of $20 billion of profit a year on every $100 billion spent — and the industry now spends well over $500 billion a year. That’s not a rounding error. That’s the difference between a boom and a fraud, hinging on a single number.

Michael Burry — yes, the guy Christian Bale played, the one who shorted the last housing bubble into the ground — has been saying the real life is two-to-three years, because each new model makes the last one worthless. On his numbers the hyperscalers are overstating profits by roughly $176 billion through 2028. And on a ten-year horizon, I think he’s probably right. Silicon doesn’t care about your depreciation schedule.

But here’s why Jensen can take the other side and mean it. He controls the roadmap. He decides the cadence of the successor chips, and outside of Google quietly building its own TPUs — which aren’t obviously as good, and whose talent has been walking out the door — there is nobody else on the planet who can design a frontier chip, build the fab, market it, and get it into production at sufficient volume inside a five-year window. Burry might be right for ten years. Jensen only has to be right for five. And as the old line goes, the market can stay irrational a lot longer than the short seller can stay solvent. He controls the clock, he controls the only competition, and he’s insuring the gap. That’s not a gamble. That’s a house edge.

The near-term tape backs him, too. H100 rental rates rose from about $1.70 a GPU-hour in late 2025 to $2.35 this year. Blackwell B200s command $5.30 to $7.05. Six-year-old A100s are still throwing off revenue. As long as demand outruns supply, the collateral holds. The question was never today. It’s what these things are worth in year four, when the next architecture is shipping and the neocloud that borrowed against a warehouse of aging silicon has to make a payment.

Same tell, five rooms

Now step back from the chip and look at the week, because the Nvidia deal isn’t an outlier. It’s the loudest instance of the pattern.

Anthropic said it will watermark every word Claude writes — an invisible mark baked into the text, to satisfy an EU transparency rule that AI output be labeled as such. Fine. Except the watermark is legible only to Anthropic. The regulator can’t read it. You can’t read it. The company being regulated is the only party who can verify its own compliance. That’s not transparency. That’s a company grading its own homework and handing the teacher a sealed envelope.

OpenAI shipped GPT-5.6-Cyber, a model that completes 95% of exploit-development requests, versus 1.5% for the consumer version with its guardrails on. Ninety-five percent is OpenAI’s own measurement, on OpenAI’s own benchmark, with no system card published and no outside lab having reproduced it. Access runs through a vetted “Daybreak Red” list that OpenAI writes. So the most capable offensive-cyber tool of the year ships with a number its maker graded and a guest list its maker controls.

OpenAI also bought back roughly $7 billion of employee stock this month — with its own balance sheet, not new investors — and held the valuation flat at $852 billion, unchanged from the last round, right ahead of a reported IPO. It marked its own book.

And one floor down, the plumbing rhymes. Anthropic’s compute is financed through $71 billion of off-balance-sheet chip-lease debt in special-purpose vehicles arranged by Apollo and Blackstone, with Broadcom’s investment-grade credit rating (not Anthropic’s) backstopping the senior notes. Its European deal leans on a $1.3 billion standby letter of credit from JPMorgan so a seven-month-old startup’s risk sits on a bank’s book instead. Everywhere you look this week, the party that’s supposed to sit between the story and the capital — the regulator, the independent evaluator, the outside underwriter, the public market — has been quietly replaced by the interested party. Nvidia insures its own collateral. Anthropic reads its own watermark. OpenAI grades its own model and marks its own price. The referee is gone, and the players are calling their own fouls.

We wrote about this exact structure yesterday, in a different room. Anthropic is making Claude Code’s auto mode the default on August 14: the AI reviews the AI, the human steps out of the loop, because on the numbers the human was only rubber-stamping anyway. That’s the same move. What was a safety-loop story on Monday is a capital-markets story today. Same tell, bigger stakes.

Why it isn’t 2008 — and why that’s not comforting

The lazy read is “this is subprime again.” It isn’t, and the difference matters.

Subprime worked as a catastrophe because the risk was hidden. Nobody thought they owned it. The correlation — that all those mortgages would default together — was the surprise that blew the system up. AAA tranches were AAA right up until they weren’t, and the whole point was that the people holding the paper didn’t understand what they held.

This is the opposite. Everybody can see it. Prakash spelled the machine out on X, step by step: Nvidia advises the banks on “reference designs” that make the data centers fungible; fungibility lets the debt get repackaged into asset-backed securities, CLOs, and CDOs; tranching earns investment-grade ratings; the paper gets resold to pension funds and insurers; and the banks trade idiosyncratic, project-specific risk for sector-wide risk. His words, not mine: “CDOs from 2008 haha.” He didn’t even get to the credit-default swaps that will inevitably get written on top, or the straight equity and debt that already trade around it. Everyone in this trade knows it’s one giant correlated bet on AI demand. They’re doing it anyway.

Why? Because Wall Street does not care what the underlying asset is, as long as it can slice it, rate it, and trade it. Give the Street something fungible with a plausible cash flow and it will build a derivatives complex on top of it by Friday. Jensen’s real genius wasn’t financial daring. It was handing Wall Street a standardized, insurable, tradable thing where there used to be a bespoke pile of depreciating metal. If the ducks are quacking, you feed them. The ducks are quacking.

So the right historical rhyme isn’t the mortgage meltdown. It’s Chuck Prince, Citigroup, July 2007, explaining why his bank kept dancing while smart people were already edging toward the exits: “As long as the music is playing, you’ve got to get up and dance.” The kid in Singin’ in the Rain sang the same two words, “Gotta dance!”, as pure, uncontainable joy. Prince meant them as a confession. That’s the whole trade in one phrase: it looks like exuberance and it’s actually compulsion, and both crowds end up on the same floor. Or go back further, to the 1870s railroads Michael Parekh points to — enormous capital thrown at real infrastructure that genuinely got built and genuinely mattered, wrapped around financing structures that produced spectacular booms and spectacular busts. The rails were real. The tracks got laid. Plenty of the financiers still got wiped out. We’ve seen this movie. The technology being real is not the same thing as the financing being safe. Those are two different bets, and this week Wall Street took the second one.

And the overshoot is baked in. As one sharp anonymous account (SightBringer) put it, the hyperscalers are trapped in a prisoner’s dilemma: underbuilding compute could permanently cripple you, so everyone keeps spending past the point where the marginal economics make sense. The overshoot isn’t a mistake at the level of strategy. It’s the equilibrium. Once credit finances the buildout directly, it gets reflexive — capex becomes chip revenue, chip revenue validates AI demand, validated demand supports valuations, high valuations justify more capex. Future expected intelligence demand gets converted into present-day concrete. The system builds far ahead of realized demand, on purpose, because the incentives point that way for every single player. That’s the same governing-dynamics trap we invoked Monday — everyone acting rationally, marching the group off the cliff together.

AI is the whole casino now

Zoom all the way out, because this is bigger than one chip company’s clever balance-sheet trick.

AI is the venture business at this point. Full stop. That’s where the dollars go, that’s where the deals happen, and good luck funding a non-AI startup right now — good luck even walking a non-AI deal into your partnership meeting without getting looked at like you brought a rotary phone. It takes real conviction, or a willingness to be alone and wrong for a while, to put a single dollar into something that isn’t AI. Only a handful of players, a Bezos or a Sequoia, have the standing to look at the other 90% of the economy while everyone else piles into the same trade. They’ll probably be better for it. Contrarians usually are, eventually.

Go up the cap stack and it’s the same story. Private equity has been trading companies back and forth among themselves for a couple of years now, because there’s no real IPO window for anything that isn’t tech or AI and the debt markets are backed up. But the PE firms are sitting on a mountain of dry powder they have to deploy. So — AI. The public markets are being carried by a handful of names, all of them tech and AI. The whole financial system has quietly concentrated itself onto one theme, and the theme now has its own purpose-built financing architecture, courtesy of Jensen and six firms with more capital than most countries.

That’s the context that makes this deal a smart move rather than a reckless one. Jensen looked at a market where every allocator on earth is desperate for AI exposure and the momentum is overwhelming, and he built the on-ramp that lets all that money flow into his customers’ data centers, which means into his chips. Kudos to him and his team. It’s the right move for Nvidia, cleanly executed, and it extends this cycle further than it would have gone otherwise.

The honest verdict has three parts, and only the first one is comfortable. Jensen wins — smart, and probably right for five years. Wall Street wins — it’ll make a fortune structuring and trading all this illiquid, esoteric paper regardless of how the underlying performs, because the fees are in the motion, not the outcome. And the lender? Good luck. Substandard returns from here, most likely, on a credit instrument wrapped around an asset that might halve in value the day the next architecture ships.

We’ve been marking this to market

None of this is a swerve for us. On August 5, in Not For Credit, we argued that credit was the wrong instrument for a market this fat-tailed — that the smart money wanted equity’s upside or nothing, and that lending into a hits business gets you the downside without the payoff. On August 7, in Know Thyself, we flagged that the physical floor everyone’s fleeing to was already cracking, with depreciation running ahead of the marks. On August 10, Memo to the Governors made the case that the money accrues to whoever owns a productive asset, not a dormant one. And yesterday, A Tragedy of the Commons, we showed the machine reviewing the machine once the human steps out of the loop.

Today braids those threads into the next turn. The instrument is credit, exactly the wrong one. The collateral is depreciating, exactly as feared. And the referee who’s supposed to price all of it has been replaced, in room after room, by the party that profits from the answer. This isn’t a sixth lap around the same track. It’s the moment the bespoke financial engineering of a few labs became a standardized, tradable asset class with the manufacturer’s name on the guarantee — and the moment everyone decided to dance anyway.

What this means for you

You’re probably not underwriting a gigawatt of GPUs. Doesn’t matter. This trade is being built under your feet, and you’re likelier to be long it than you think.

Re-run your own compute math on a three-year clock. If you’re signing multi-year AI-infrastructure commitments, price them the way Burry would, not the way Jensen would. If the deal only works assuming the chips stay productive for six years, you’re the one absorbing the depreciation risk Wall Street just tranched away. Make the vendor wear it, or walk.

If you’re bullish, own the equity, not the paper. This is the whole point. A substandard credit return on an asset that can crater is the sucker’s seat. The upside in this cycle lives in the equity — in the companies running the compute hot and turning it into revenue — not in a coupon on a loan the manufacturer had to backstop to get done.

Find out what you already hold. The paper in this machine ends up where structured credit always ends up: tranched, rated, and sold to pension funds and insurers. That’s your retirement account and your annuity. Next statement, next annual meeting, ask the plain question — how much of our “diversified private credit” is now AI-infrastructure debt? You may be dancing without having chosen the song.

The music’s playing. The buildout is real, the chips are real, the demand is probably real for a good while yet. Just don’t confuse any of that with the financing being safe, and don’t be the one still on the floor holding a coupon when it stops. Because it isn’t Jensen who’s holding the paper — he made sure of that.

Bullish on AI? Own the equity. Never be the lender in a game where the house sets the clock.

---

Sources

• NVIDIA AI Factory Compute Is Becoming an Investable Asset Class — NVIDIA, Aug 11, 2026 (the $500B platform; 25% residual-value support; “compute is revenue”; A100 longevity; H100/B200 rental rates)
• Wall Street just endorsed Jensen Huang’s ‘big concept’ for AI. What now? — CNBC, Aug 11, 2026 (Fink’s 1970s MBS quote; Solomon “these are real assets”; Zelter “securitize it”; Flatt; MOUs, no contracts; the $100B/$30B OpenAI history)
• Nvidia and Wall Street team up on $500 billion bet on AI infrastructure — CNN, Aug 11, 2026 (circular-financing concern; $5.3T Nvidia market cap; deVere depreciation caveat)
• Why Jensen Huang’s $500 billion AI financing plan faces a big risk from China — CNBC, Aug 11, 2026 (Ben Emons depreciation risk; 11–17% yield estimate; non-investment-grade borrowers / neoclouds)
• Nvidia Scores $500 Billion in Financing for Its Customers — Tae Kim, Aug 10, 2026 ($50–60B/GW; ~70 GW US demand; ~$9T money-market fuel; money flows to neoclouds/OpenAI/Anthropic/SpaceX, not the top-3 hyperscalers; fungibility vs. ASICs)
• Gaming the System for AI: Nvidia, OpenAI & Anthropic (ARD #138) — Michael Parekh, Aug 11, 2026 (“all three are about arranging capital”; “redesigning the board”; 1870s railroad parallel; $3.5T Morgan Stanley spend forecast)
• Prakash (@8teAPi) on the securitization chain — X, Aug 10, 2026 (depreciation insurance → reference designs → ABS/CLO/CDO → tranching → pensions/insurers; idiosyncratic → sector-wide risk)
• Kyle Reidhead on the GPU-useful-life bet — X, Aug 10, 2026 (Amazon 5 / Meta 5.5 / Google 6; Burry 2–3 yrs; ~$176B overstatement; $33B/$17B/$10B depreciation math)
• SightBringer (@_the_prophet__) on the buildout as equilibrium and the financial regime — X, Aug 10, 2026 (prisoner’s dilemma overshoot; reflexivity loop; “future expected intelligence demand converted into present construction”)
• Aravind Srinivas: “Compute is the currency” — X, Aug 10, 2026
• OpenAI’s cyber model answers 95% of exploit requests; Zuckerberg presses Washington — The Implicator (Marcus Schuler), Aug 11, 2026 (GPT-5.6-Cyber / Daybreak Red; 95.0% self-graded, no system card; $852B self-funded buyback)
• Anthropic’s $965B IPO Path Runs Through Seven Compute Corridors — Yahoo Finance, Aug 11, 2026 ($71B off-balance-sheet chip-lease SPVs; Broadcom residual backstop; JPMorgan $1.3B SBLC)
• Anthropic to start embedding invisible watermarks in Claude’s AI-generated text — Fortune, Aug 11, 2026 (EU compliance; marks readable only by Anthropic)
• Riot Platforms signs ~$9.1B AI data-center deal with Anthropic — Aug 11, 2026 (20-yr, 191 MW Texas, up to $16.1B)
• Prior CO/AI issues: Not For Credit (Aug 5), Know Thyself (Aug 7), Memo to the Governors (Aug 10), A Tragedy of the Commons (Aug 11)
• Chuck Prince, Citigroup CEO, Financial Times, July 2007 (“as long as the music is playing, you’ve got to get up and dance”)
• Michael Burry, AI-chip depreciation / useful-life thesis (2025–26); The Big Short (2015), Christian Bale as Burry