Over the weekend the Journal reported that Nvidia is in talks to guarantee roughly $250B of financing so OpenAI can lease a 10-gigawatt data center a SoftBank subsidiary is building in Ohio. Bloomberg has now mapped a web of AI financing arrangements totaling north of $750B, the BIS and IMF are using the phrase “systemic risk,” and Michael Burry, who has puts against Nvidia and Palantir, resurfaced within hours. On Reddit, someone asked whether anyone has actually mapped the scale of the circular financing. That’s the strange thing about this debate., it is the loudest argument in markets, and it is running almost entirely on vibes and totals. The bears wave the capex number, the bulls wave demand anecdotes, and both sides treat AI revenue as a single undifferentiated bucket. But enough companies now disclose AI-specific revenue, segment margins, and depreciation schedules that you can just do the accounting. So let’s grade the bear case the way it deserves to be graded, claim by claim, against the disclosed numbers. Some of it survives and some of it is already refuted by data the bears don’t seem to have looked at. And the parts that survive point at very different names than the ones the famous shorts are aimed at.The depreciation schedules are a lieThis is Burry’s core argument and the most serious one on the board. Last November he called extending asset lives one of the more common frauds of the modern era: hyperscalers depreciate Nvidia-based hardware over five or six years while the real economic life, he argues, is closer to two or three, because each chip generation resets the performance-per-watt math. By his estimate the group will understate depreciation by about $176B across 2026-2028, overstating Oracle’s 2028 earnings by 26.9% and Meta’s by 20.8%.Score the concession first, because the record here is worse than most bulls admit, and none of it is hidden. It’s in the PP&E footnotes. Every hyperscaler spent 2022 and 2023 extending server lives, in the middle of the fastest hardware cycle in computing history, and each disclosed what the pencil stroke was worth:The three quantified moves alone sum to nearly $10B of first-year operating income created by estimate changes rather than operations, booked in exactly the years the AI fleets were supposedly aging fastest. Meta made its most recent extension in 2025, with the AI cycle in full sprint. Burry didn’t uncover a conspiracy; he read the footnotes and did the multiplication. And the leverage on the estimate is enormous where it matters most: depreciation consumes 14% of revenue at Oracle and rising, 48% at CoreWeave, 88% at Nebius. Rerun Oracle’s cloud numbers on a three-year life and the business goes from marginally profitable to outright negative before a dollar of interest expense. On the mechanism, Burry is right, and the bulls who wave this away are the ones ignoring data. But the conclusion has now been tested three ways, and so far it fails all three.The first test is the one company that already ran his experiment. Effective January 2025, Amazon cut the useful life of a subset of its servers and networking equipment from six years back to five, citing “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.” It also retired some equipment outright, taking $920M of accelerated depreciation in Q4 2024, and guided the two moves at roughly $1.3B of 2025 operating-income drag. That disclosure cuts both ways, and both cuts matter. It proves the schedules do move down when fleet reality demands it, with auditors attached; useful life is a live estimate, not a fixed fiction. And it prices the sensitivity in public: one year of life, on the affected fleet, at the largest capex spender on the planet, costs about $0.7B a year. The market absorbed it without a flinch. Burry’s thesis needs not one year at one company but three years at every company, all in the same direction, all at once. The single live datapoint we have shows the machinery correcting incrementally and cheaply instead.The second test is that GPUs rent by the hour in a public market, which means economic life here is not a matter of opinion. You can read it off the price curve, and the curve tells a two-act story. Act one belongs to the bears: H100 rental rates collapsed from around $8/hr at the 2023 peak to roughly $1.70 by October 2025 as supply caught demand. That crash is the honest core of the useful-life worry, proof that the revenue under a GPU can fall 75% in two years. But act two is happening now, and it runs the other way. One-year H100 contract rates rose about 40% off that October low to roughly $2.35/hr by March 2026, including a 10% jump in the four weeks around New Year’s, and on-demand rates across fifteen-plus providers now cluster near $3/hr. This is a 2023-vintage chip whose market rent is rising in year four. Run the payback math: at $2-3/hr and realistic utilization, a chip that cost $25-30k returns its purchase price in roughly two to three years of rented operation. A six-year schedule therefore doesn’t require year-six heroics to be honest accounting; it requires years four through six to earn anything at all, and the market is currently paying 2020-vintage A100s a rising rent in year six. CoreWeave disclosed exactly that last quarter, price increases across A100s, H100s, H200s, and L40s, and DigitalOcean independently reported the same on its own fleet. Two operators, different customer bases, one signal: inference demand is bidding up old silicon, not stranding it. And a substantial share of the capex, the shells and the power infrastructure on fifteen-to-forty-year lives, was never in dispute at all; in a market where the binding constraint is energized capacity, an old GPU in a powered rack keeps its job because the scarce asset is the rack.The third test is scale, and whether the market already paid the trade. $176B over three years is roughly $59B a year, against a group spending $700B-plus on 2026 capex and generating well over $500B of combined operating cash flow. If Burry is entirely right, group earnings are overstated by high single digits, concentrated in the levered names. Real, but it re-times earnings; it does not reveal the demand to be imaginary. And his most-overstated name has already been repriced as if he won: Oracle sits nearly 60% off its high, had its worst week since the 2001 bust, got cut to BBB- explicitly on customer concentration, and saw its CDS blow out to the widest since 2008. The insight was right; the trade has largely been paid.What would flip us to his side has a mechanism and a date, not just a mood. The 2023-to-2025 rental crash proved the curve can fall fast when supply catches demand, and Nvidia is now moving to an annual product cadence, with Vera Rubin shipping in the second half, which ages fleets faster than the two-year rhythm the six-year schedules were calibrated against. If Rubin volume lands into a demand plateau, the 2024 repricing repeats, the year-four rents currently defending the schedules roll over, and Burry’s math stops being a timing argument and becomes an impairment forecast. The H100 rental indexes through the Rubin ramp are the single most information-dense series in this entire debate.Bottom line the mechanism confirmed, and better documented than the bulls like to admit; the conclusion fails its three empirical tests so far. Amazon’s reversal shows the estimates self-correct at single-digit-billions cost, the rental curve shows year-four assets earning rising rents, and the name with the most alleged overstatement already trades like it. The flip signal is used-GPU pricing rolling over into the Rubin ramp, and if that happens, reported margins across the levered tier really are fiction.The demand is circular, and Nvidia is CiscoThe financing web is real, so let’s state it at full strength rather than strawman it. Nvidia holds around $50B of equity spread across its own customers, including a $30B stake in OpenAI. It is now discussing a $250B debt backstop so that same customer can lease capacity it will fill with roughly $350B of, eventually, Nvidia silicon. Amazon put $13B into Anthropic with up to $20B more committed, then committed $50B to OpenAI’s latest round, bankrolling both sides of the frontier race at once. Money goes out as investment, comes back as revenue, and the revenue justifies the valuation that funds the next investment. Bloomberg’s $750B tally is the sum of arrangements like these. That is the geometry of vendor financing, and vendor financing at the end of a cycle is how you pull forward demand that wouldn’t otherwise exist. In 2000, Cisco did exactly this, and the demand turned out to be borrowed from a future that never arrived.Here is what the fabricated-demand version of this claim has to explain away, though. This year, Nvidia effectively deleted the world’s second-largest market from its own income statement. The H20 is dead in China, next-quarter guidance assumes zero China data-center revenue, and something like $20B a year of demand went to zero. If demand were substantially a financing loop, losing a fifth of it should show. Instead Nvidia grew revenue 85% to $81.6B, data center up 92%, gross margin at 75%, and guided the next quarter to $91B. Meanwhile TSMC, which has no stake in anyone’s financing web, posted record revenue and says it is sold out of leading-edge capacity. And the composition of Nvidia’s demand is broadening away from the names in the loop: under its new segmentation, hyperscalers did $38B last quarter while AI clouds, enterprise, industrial, and sovereign customers did $37B, growing 31% sequentially versus 12%. The half of demand growing fastest is the half furthest from the circular diagrams.The token data says the same thing from the consumption side. After DeepSeek in January 2025, the last time “the demand is fake” was the consensus short, routed token volume quadrupled in eighteen months. Total inference grew faster than efficiency improved. That experiment has been run once at full scale, and the efficiency-kills-demand side lost.Now the Cisco comparison specifically, because it deserves a precise answer rather than a dismissal. Cisco’s 2000 problem wasn’t just vendor financing; it was who the customers were. The dot-com buildout was funded by unprofitable startups spending raised equity, and when the equity window shut, the demand evaporated because the customers did. Nvidia’s three largest customers generated, respectively, $98B, $62B, and $19B of EBITDA in excess of their capex last year. They are funding the buildout out of operating cash flow from search, software, ads, and retail. The customers that do look like 2000, OpenAI losing a reported $14B a year, the neoclouds spending two to eight times their revenue on capex with borrowed money, are real, and they are exactly where the backstop news should worry you. When your customer’s lenders require your guarantee before they’ll fund the lease, your balance sheet is absorbing risk the credit market declined. That is late-cycle behavior and it deserves to be called what it is.Circularity is a genuine concentration and credit risk located in an identifiable minority of demand, the levered edge. It is not evidence the demand complex is a loop, because the majority of the money is cash-flow-funded and the demand keeps showing up in places the loop doesn’t touch. Burry’s Cisco at the center line has the geometry right and the customers wrong.There’s no revenueThis one was true in 2024, arguable in 2025, and is now simply out of date, though the bears keep repeating it as if the disclosures stopped. So let’s list what is actually disclosed, because this is the revenue side of the ledger everyone says doesn’t exist. Microsoft’s AI business crossed a $37B annual run-rate. OpenAI is running around $25B of revenue. Anthropic just posted its first profitable quarter at north of 80% API gross margins, straight through a price war. One layer up, the application companies now report AI-specific lines: Salesforce’s Agentforce at a $1.2B run-rate growing more than 200%, ServiceNow’s AI contract value past $1B, Atlassian’s AI-led Service Collection past $1B in ARR, Workday past $500M, Adobe’s AI-first ARR past $500M and tripling. And the cloud segments where AI lands as consumption are reaccelerating in a way that is hard to fake: AWS grew 28% last quarter, its fastest in fifteen quarters, and Google Cloud has gone from 48% growth to 82% (includes TPU sales) and kept accelerating through the most recent print, with backlog at $514B. By one outside tally, AI infrastructure now produces about $1.19 of revenue per dollar of depreciation, having crossed 1.0 within the past year. That figure rests on a single dataset, so hold it loosely, but the direction matches every disclosed line above.Add it up honestly, though, and you get the concession, and it’s a big one: the income statement still doesn’t balance. Direct, disclosed AI revenue, counted generously, is in the low hundreds of billions against $700-725B of 2026 hyperscaler capex, guided higher again just last week. The gap is real, it is the number the whole market is arguing about without computing, and anyone telling you the revenue has already caught the spending is selling something. But a gap between investment and revenue during a buildout is not, by itself, a bear case. It’s what a buildout is. The question that actually separates the outcomes is whether utilization fills capacity before depreciation and funding costs catch up. And on that question we have one nearly complete, fully disclosed natural experiment: Google Cloud went from a 9% operating margin to roughly 30% in eight quarters as demand filled its installed base, earning a 56% margin on each incremental revenue dollar along the way. AWS at maturity runs 35% margins after depreciation, at $129B of revenue. That is what the far side of the gap looks like when it closes. The levered tier’s entire investment case is a wager that the same curve exists for them, on borrowed time. Nothing in their current numbers proves it yet, and nothing disproves it.No revenue is false and measurably so, the revenue is not yet arriving fast enough to cover the capital bill, and someone is holding the duration risk in between is true, and is a better bear case than the one the bears keep making.This is worse than 1999Apollo’s Torsten Slok put the institutional version on the record: the top ten companies in the S&P 500 are more overvalued today than the top ten were in the 1990s. Coming from the sell side’s favorite economist, it became the era-defining quote of the correction. As a P/E observation about index concentration it’s defensible. As an analogy it fails on the one dimension that made 1999 lethal: what you got underneath the multiple. Cisco at the peak traded north of 100x earnings that depended on customers burning venture money. Nvidia, after this year’s de-rate, trades around 23x current-year consensus and roughly 16.5x next year’s, with 75% gross margins and consensus free cash flow north of $200B this fiscal year on its way toward $280B. Microsoft, Alphabet, and Meta fund the buildout from operating cash flow and trade at 20-30x. You can argue those multiples embed optimism about AI returns. You cannot argue they resemble 1999’s arithmetic, and the profitability difference isn’t a detail, it’s the difference between a repricing and an extinction event.Here’s what’s frustrating about the dot-com comparison: pointed at the right tier, it’s genuinely useful. There is a layer of this market where 1999 is happening on schedule. Oracle is spending 83% of revenue on capex, running $40B-plus of consensus-negative free cash flow, carrying $167B of debt, one notch above junk, with half its $638B backlog tied to a single money-losing customer. CoreWeave’s P&L is a pass-through from customer to depreciation schedule to lender. Nebius spent nearly eight times its revenue on capex last year. OpenAI is losing $14B a year while discussing a trillion-dollar IPO. That tier is levered, concentrated, and priced for a conversion curve it hasn’t demonstrated. The dot-com analogy belongs there, not on the ten largest companies in the index, which are, almost embarrassingly, the most cash-generative entities in economic history. The comparison is a category error at the top of the index and a fair warning one layer down.The exhibit the bears earned: Alphabet’s quarterGive the bear case its due, because last week produced the single best piece of evidence it has ever had, and it came from the strongest balance sheet in the group. Alphabet grew revenue 24% to $119.8B, capex came in at $44.9B for the quarter, roughly double a year ago, and free cash flow printed negative $5.9B, the first negative quarter since 2004. Full-year capex guidance went up again, to $195-205B, the second raise this year, with management saying it scales further in 2027. And the funding mix told you what the income statement didn’t: roughly $100B of debt raised this year, an $85B share sale in June, reportedly its first in more than two decades, and the buyback stopped. Read that as revealed preference. A company that spent a decade retiring shares is now issuing them to fund data centers. Management is telling you, in the language of capital allocation rather than earnings-call adjectives, that the AI bill now competes directly with shareholders and the bill is winning. The seven percent drop was the market repricing exactly that.So the bears are right that the cash math has turned, and turned hard. Where the popular reading goes wrong is in treating a funding-model transition as a demand falsification. Nothing in the quarter said demand is weak; revenue accelerated, the cloud segment surged, and the backlog grew. What the quarter said is that even the best-positioned spender can no longer fund this buildout invisibly out of operating cash flow, which moves every hyperscaler one step closer to the discipline the bears have been demanding. The irony is that this is bullish for the thing bears say doesn’t exist, capital discipline, and bearish mostly for anyone who assumed the buildout was free.The statement doesn’t balance in 2026. On the disclosed trajectory, spend up ~77% and disclosed AI revenue lines growing 100-200%+ from a much smaller base, it doesn’t balance in 2027 either. Whether it balances at all gets decided by whether the Google Cloud fill curve generalizes, and the earliest honest readings on that are dated: hyperscaler capex guides and cloud margins this week (Microsoft, Meta, Amazon all report), Nvidia’s August 26 print with the first Vera Rubin guide, Google Cloud’s margin in late 2026 when the big capex vintage hits depreciation in force, Oracle’s FY27 depreciation line against the roughly $12B path, and secondhand GPU pricing every single quarter.With eyes wide open the soft spots in the trade deserve acknowledgment. The Amazon 2027 free-cash-flow consensus we lean on spans negative $40B to positive $100B across analysts, which means the Street collectively has no idea and Microsoft’s headline demand metrics now bend around a single partner: commercial RPO grew 99% including OpenAI and 26% without it, which is neocloud-grade counterparty concentration hiding inside the bluest chip in software. If the backstop deal is the moment vendor financing tips from supporting demand to manufacturing it, the place it would show first is in disclosures like that one. The bear case, as argued by its most famous advocates, gets the mechanism right and the target wrong. Depreciation timing, funding stress, and counterparty concentration are the real fault lines, and they run through Oracle, the neoclouds, and the loss-making lab layer, names and layers that have mostly already been repriced for it. They do not run through the demand data, which keeps refusing to cooperate with the fabrication story, and they do not make the ten most profitable companies in history into 1999 dot-coms.