AI News

AI Bubble Monitor Part 3: Nvidia's Gross Margin Option and the Five-Layer Bubble Accumulation Chart

TMTPost · AGIUpdated Nov 16 20:39

AI summary

TMTPost published the third installment of Qiongqi Guancha's AI Bubble Monitor, breaking the U.S. AI industrial chain into five layers—debt, votes, companies, revenue, and global—and pointing out that Nvidia's data center business gross margin ranged between 71% and 75% across quarters of fiscal year 2026, with a midpoint of about 73%, far above the 15% to 25% range of ordinary hardware suppliers. The article argues that this difference is the rent the industrial chain pays to Nvidia each year, and deduces two dissolution paths—gross margin reversion and Nvidia proactively adjusting long-term contracts—while also listing five countervailing facts and observation nodes such as the November 17 earnings report.

Why read this

AI score 60

The article breaks down Nvidia's gross margin of over 70% into industrial chain rents, and presents two deduction paths—gross margin reversion and contract loosening—to help understand the pressure transmission along the AI compute chain.

Topics

Tags

Body · AI translation

By Qiongqi Observation

Bubble Monitor Part One dissected the debt layer, Part Two dissected the ballot layer; this bullet has scrambled the order—the exam paper has changed its own course, and this piece overlays five tables into one chart. The detailed accounts of the revenue layer will be calculated in a separate article.

NVIDIA's data center business gross margin is above 70%—fluctuating between 71% and 75% across quarters in fiscal 2026, with a center around 73%. A "normal hardware supplier" has a normal gross margin of 15% to 25%, and the contract manufacturing empire Foxconn has only single digits. The fifty percentage points of difference in between is the rent the entire industry chain pays it every year.

How much longer can this rent be collected? Spread out the books of the US AI industry chain over the past two years layer by layer, and the answer surfaces on its own: this is a stratified bubble chart, each layer pressing down on the next, and at the very top, the fattest bubble of all, is NVIDIA. Prick it, and everyone downstream can breathe a sigh of relief—not for the stock price, but for the rent. This piece completes the whole chart, then calculates the account of "how to prick it."

Layers one through five: how the bubble was blown up layer by layer

Debt layer. Compute leasing companies use long-term customer contracts to finance, borrow money to buy cards, and then lease out the compute—the economic lifespan of GPU hardware is about 2.5 years, the project loan cycle is 6 to 7 years, and data center property leases stretch to 15 to 20 years: short assets against long liabilities, with all three maturities mismatched by one layer each. This layer began leaking in autumn 2026: the yield on CoreWeave's unsecured debt once touched 15%, and its 9.625% bond due 2032 fell from par to 76 cents; Fermi's stock price fell from an IPO price of $21 (SEC shareholder letter: 32.5 million shares, priced at the top of the marketing range) to around $2, with zero revenue in the third quarter and net debt of $958 million. Private credit channels are also shrinking: two funds under Blue Owl received about $5.4 billion in redemption requests in the first quarter, with the backlog rolling to $4.7 billion in the second quarter and $4.2 billion in the third—the pool has not dried up, but everyone knows others are bailing water; according to industry media reports, when it sought project financing of $6 billion for a Pennsylvania project anchored to a CoreWeave lease, an executive at a large specialized lending institution replied in four words: we looked, didn't take it. At the center of the chain is Oracle: remaining performance obligations of $664 billion; according to market institution estimates, about 50% will be recognized as revenue within the next 36 months, while a year ago the market estimated this proportion at only 35%—the bill's due date is accelerating. Under third-party calculation standards, its debt plus leases account for 71.6% of capital; under the same standard, Google is 15.5%.

Ballot layer. In the US midterm elections on November 3, among about 50 data center-related ballot measures, the vast majority of voters supported stricter restrictions (Ballotpedia standard, as relayed by media); in August, the Texas governor first paused approval of new data center grid connections and ordered an audit. According to media reports, White House Economic Council Director Hassett publicly admitted: voters are "overwhelmingly inclined to restrict tech infrastructure construction." After regulation moved from Brussels to Washington, it then moved into county governments—data centers began to be voted on as local nuisances. The election results have landed: the Democratic Party regained both chambers of Congress, and Texas, Ohio, and Georgia, three major data center states, replaced their governors with ones advocating a pause on new projects. On November 8, when NVIDIA's board launched an independent investigation, it still said Jensen Huang "will continue to lead the company without interruption"; on November 11, the board accepted his resignation, effective immediately, announced the next day—the 8-K text did not disclose the specific reason for the resignation (contemporaneous media reports pointed to personal conduct controversy; he himself had admitted an extramarital affair in a public letter on November 8). The ballot pricer needs an outlet it can name—this is the interpretive structure of market observers; the board documents themselves do not state this logic; but the 8-K of November 11 objectively handed over that name.

Company layer. NVIDIA's own crack is not on the license list, but on its customers' books: in the official CFO explanation, days sales outstanding (DSO, the number of days from revenue recognition to cash collection) rose from 45 days to 60 days, with the reason stated—"large multi-quarter agreements with some investment-grade customers extended payment terms." An extra 15 days in a quarter is a direct reading of it fronting cash for customers. The market has calculated the loss from export controls three times over the past year, but has hardly calculated anything for DSO.

Revenue layer. Model companies that purely buy compute (the Anthropic and OpenAI layer) sign 7-to-10-year cloud contracts that are 80% non-cancelable and "pay even if you don't use it." Anthropic's details are: roughly $450 billion in cloud computing commitments over the next 7 to 10 years—over $100 billion with AWS over ten years, Google Cloud commitments rising to about $121.3 billion (revised prospectus), Microsoft commitments of $47.2 billion to $49.2 billion in spending through June 2033, about $86 billion with Nscale and Lambda (signed through 2033/2034), plus a $160 billion chip commitment with Broadcom under negotiation; on the prospectus's aggregate basis, purchase and lease commitments total $895.4 billion, with $69.8 billion due within 12 months—an average of $5.8 billion a month. Yet its operating profit wobbles around zero: last quarter it turned profitable for the first time (+$974 million), and this quarter it swung back to a loss (-$567 million, a Non-GAAP operating margin of 15%, down from 19% the previous quarter); third-quarter revenue was $16.96 billion, up 46.7% quarter over quarter—the previous quarter was still 144%; compute and infrastructure costs were 58% of revenue. For every $1.72 of revenue it burns $1 of compute; last quarter that figure was still 2.08. It has also left itself an escape hatch: its contract to rent compute from xAI is cancelable on 90 days' notice—the only one among the third parties that has not welded its retreat shut. OpenAI is even more extreme: about $70 billion a year on the paying side, of which $60 billion a year goes to Oracle alone starting in 2027, against annual revenue of about $50 billion—a structural gap, filled by funding rounds, and the $1.25 trillion round has not yet closed. In the third quarter its revenue was about $10 billion, up 50% quarter over quarter, and its adjusted operating loss widened from $10.1 billion to $10.9 billion.

Global layer. The bottom layer has changed: NVIDIA's China story can no longer be told. On IDC's basis, Huawei's Ascend shipped 572,000 units in the first half, accounting for 28% of China's data center AI chips, while NVIDIA shipped 678,000 units for 34%—the gap has narrowed from a generational one to a neck-and-neck one; Morgan Stanley estimates China's AI chip self-sufficiency rate at 35% in 2025 and 68% in 2027; Cambricon's inventory rose by nearly 70%, and Hygon's prepayments sit on the balance sheet—domestic substitution is not a narrative, it is an accounting line item. One calibration: in the most advanced process nodes, HBM supply and software ecosystems, China is still a generation behind. But "good enough" is just as lethal to monopoly rent as "most advanced"—rent is collected on scarcity, not on a technology score. Of course, if the technology gap reopens with the next-generation architecture, the rent will rebound—this is already written in the second counterpoint; read the two side by side.

Stack the five layers together and you get an inverted pyramid: the base is loosening, while the tip still hangs at the high-water mark of five trillion.

The biggest bubble is the option on gross margin

NVIDIA's market value at its peak was about five trillion dollars. Unpack that number: what it buys is not today's profit, but the assumption that "a gross margin center above 70% can be sustained for another fifteen years"—what the market is pricing is not a P/E ratio, but an option on gross margin. A "normal hardware supplier" has a normal gross margin of 15% to 25%. The fifty percentage points in between are the rent the entire industry chain pays it every year.

Who pays the rent? Divide the downstream by contractual relationship into three types, each with different books:

Party B, the chip buyers: they hold GPU assets themselves and produce compute. Pure Party Bs are the sublessors like CoreWeave, Fermi and Nebius—the mismatch is all on their own books, making them the most fragile; sovereign clouds like G42 and HUMAIN are the same type, only with different parents; hybrid Party Bs are Google, Microsoft, Amazon and Meta—the cards they buy first feed their own businesses, and their self-used capacity is their anchor, so no matter how far things fall, they will not be the first to die.

Party C, the compute buyers: they do not touch chips, only sign long-term contracts—Anthropic, OpenAI, and a host of Chinese model companies. The common structure: the income statement is the story, the cash flow statement is the oxygen tank. Anthropic's oxygen is enough to burn for three years (provided the IPO goes smoothly); OpenAI's oxygen is supplied "by funding round."

Those who are both Party B and Party C: the four major cloud players. When buying chips they are Party B; when feeding compute to their own businesses they are Party C—this dual identity gives them a composure others lack: if NVIDIA dares to cut off their supply, their own businesses hurt first.

How does the pressure transmit? The direction is reverse: weak end demand → Party C revenue falls short of assumptions → Party C stretches Party B's payment terms (that is where NVIDIA's 15 days of DSO comes from) → Party B bond spreads widen → banks pull loans and Party B cuts new orders → NVIDIA's guidance comes under pressure → the equity layer collapses first. Each link pushes the pressure toward the thicker, less mobile link in the chain—and at the end of that push is NVIDIA: it is the thickest, and it cannot run.

So the mood across the industry is a precise arithmetic problem: do they want it dead? No. Do they want it weak? Yes. Weak to what degree? Weak enough to give the rent back and loosen the rigidity—but the goods must still be delivered on schedule.

The two moves that would pop it: return the rent, void the contracts

What follows is a mechanism exercise: if this rent is to be dissolved, in theory there are only two moves. This is logical deduction, not a prediction of the future—both moves have verifiable checkpoints at their dividing lines.

Step one: gross margin reverts. Push the gross-margin center of gravity above 70% back down to the low teens, demoting the compute pope to an ordinary hardware supplier. The essence of this move is price discovery—tens of billions of dollars in rent each year flows back downstream, cloud providers' procurement prices are repriced, third-party compute costs are repriced, and the entire industry chain's profit-margin table is rewritten. Note: cutting gross margin is not crashing the stock price—market-cap evaporation is the shareholders' ledger; a gross-margin decline is what actually means the rent is truly going back. What everyone is breathing is gross margin, not the candlestick chart. There is no shortcut on this road; it can only be worn down by four things: customer multi-sourcing (AMD, TPU, and Trainium shares climbing from single digits toward 30%), competitor supply (the inference-chip technology gap is the smallest, and inference is the future's largest volume—inference-market gross margins are naturally lower than training's), antitrust (the old CUDA-bundling case can be reopened at any time), and circular-financing blowback (customers who are also shareholders will change their attitude toward gross margin). The timetable is three to five years, grinding down one notch each quarter. And the company itself has already drawn the first notch: the Q3 gross-margin guidance given in the August earnings report is about 74% (plus or minus 50 basis points), bottoming at 71% to 72% in Q4, and stabilizing at 72% to 73% in fiscal 2028—the main downward driver being extreme memory-chip pricing. Distinguish clearly: this is cost squeeze, not rent receding; a squeeze can be passed on through price increases (the company explicitly says it will raise prices); but it punctures the market's first assumption—above 70% is not a constant, and even the company itself is pointing down from 73%. In the same quarter there is a symmetrical ledger entry: NVIDIA's own total future purchase commitments rose to $279 billion, mainly related to memory chips. The pope is also paying rent upstream—the rent chain is connected, and no one stands outside the chain.

Step two: it opens its mouth and voids the contracts. This step is more poisonous, and it requires even more finesse. First, the constraint: from the standpoint of NVIDIA's commercial interests, voluntarily voiding contracts is an extremely costly second-worst option that would never happen in a normal year—what is being deduced here is how that corner gets formed. And the guard of that corner has just changed: interim CEO Colette Kress comes from a finance background, and according to The Information, she has repeatedly opposed high-risk deals pushed by Jensen Huang, most typically opposing up to $250 billion in credit support for OpenAI, a deal ultimately split to about $100 billion; the market sees her as a more prudent helmsman than her predecessor, and she has begun diversifying the company's financial exposure to a small number of customers. An interim leader naturally wary of circular financing is precisely the one most likely to say "adjust long-term contracts" out loud—the initiator of contract voiding has shifted from the founder to the CFO. The root of pressure on everyone in the chain is that pile of rigid 7-to-10-year "pay even if unused" long-term contracts—Anthropic's $450 billion, Oracle's $664 billion RPO, all piled up from such long-term contracts. There is only one low-cost way to unlock them: the supplier moves first. The initiator of contract termination decides everything—if the second and third parties tear up the contracts first, it is a buyer-default narrative, and the credit layer explodes on the spot; if NVIDIA moves first to adjust contracts, it is a supplier-adjustment narrative, and the entire chain's right to claim damages is instantly activated, the second party's procurement obligations loosen, and the third party gets all the chips for renegotiation. The same contract-voiding action, in different directions: one is a hard landing for the whole chain, the other a textbook soft landing.

So the most precise formulation is: want it to breach, not to cut supply; want it to partially perform, keeping the contracts loosely performed. The downstream optimum is for it to "breach just right"—enough to pay compensation, not enough to cut the lifeline. This equilibrium point has three brakes, which make it as narrow as a knife's edge: the breach-compensation dose is too small (late-payment penalties are usually only 1% to 5% of the contract value) to fill the $895.4 billion hole; once compensation is large enough to hurt its own credit, the pile of "claims against NVIDIA" in the second party's hands turns into exposure to a weakened counterparty—pressure comes back under a different name; and lock-in without substitutes instead gives customers the most valuable thing—a legitimate excuse for multi-sourcing, with its own market share dismantled one breach at a time. When the day truly comes that it cannot deliver, it is not the buyer who wins; everyone hard-lands together.

This script's premise is gradualism—gross margin grinds down by the year, not chopped by the quarter. If the rent recedes too fast, the sublandlord's GPU collateral collapses first, and the debt layer explodes before the rent layer. Dosage decides soft versus hard; rhythm decides life versus death.

Why must these two moves be made by its own hand? Because its income statement is the only one in the whole chain thick enough to absorb both bills at once—"rent returned + breach compensation." Pressure never disappears; it is eaten into its own statement one bill at a time—the equity layer evaporating cushions the first layer, gross-margin reversion cushions the second, breach compensation cushions the third. After three layers of cushioning, it goes from pope to ordinary supplier, and everyone in the chain gets a window to breathe.

Why now: two bottoms are leaking at the same time

The first bottom: monopoly rent. The original design of US AI hegemony was a three-step game—blockade China to lock in scarcity (in January H200 switched to case-by-case review; in November NVIDIA added a 51% non-China customer floor to new cloud-customer contracts), monopoly pricing to hold above 70%, global harvesting to prop up valuations, and then the game of passing the parcel could continue. The first link of this design has already failed: China has caught up at the "good enough" level—good enough for inference, usable for training, and no longer dying from supply cuts. Ascend's 28% versus 34% is a close-quarters gap; the self-sufficiency ramp curve can be seen directly in the earnings reports.

The second bottom: the financing side. The 10-year Treasury yield hit 5.34% in early October, a new high since 2002—the era of cheap money is over—and AI infrastructure is the newest, largest industry in the global economy that is still leveraging the cheap-money logic. That same autumn, chipmakers began investing in their customers: TSMC's board approved three deals each capped at $500 million (buying Anthropic common stock at the IPO price, buying OpenAI preferred stock, and setting up a semiconductor investment fund); the same IPO also included a concurrent private placement totaling $26 billion—reports list NVIDIA and Google at $10 billion each, some of which are non-binding subscription intentions. Suppliers are lining up to subscribe to their customer's stock—the loop of "invest in the customer, the customer buys cards, you recognize revenue" is a flywheel when times are loose and a noose when they tighten.

There is one more gauge that is easiest to overlook: the changing of the moat. Over the past two years, America's toolbox has shifted from the CUDA ecosystem and process-node advantages all the way to RASA, the 51% rule, location audits, and ballots. Every inch of tightened controls is circumstantial evidence that the technological moat has receded an inch—if the rent could still be collected technologically, who would be willing to pay the political cost to collect it?

With both bottoms leaking at once, the valuation at the top is left hanging. The bullseye of this Jensen Huang bullet was never the scandal—the scandal is only the trigger. The real target is that arithmetic question: **How many years are left for a gross margin center above 70%?** And China's compute is the heaviest weight pressing on that question.

The opposing side

Here are five opposing facts.

One, pricking NVIDIA's bubble may kill the counterparty first: if the rent retreats too fast, the sublandlord's collateral (secondhand GPU prices) collapses along with it, and the debt layer blows up before the rent layer—the crux of dosiology lies in the tempo, not the direction.

Two, China's "catching up" is still one generation behind: advanced process nodes, HBM supply, and the software ecosystem remain hard constraints. "Good enough" can break a monopoly, but it cannot break a technological generational gap—if the next-generation architecture opens up another generational lead, the rent story gets a new lease on life.

Three, the gross margin cannot be annualized away in one go: it requires diversification of sources, competitor supply, and antitrust action to advance on three fronts simultaneously, on a timetable of three to five years. If any link loosens in the interim—for example, if inference demand explodes until supply is once again short—the rent will rebound.

Four, the most realistic opposing point: the initiative to scrap contracts lies with NVIDIA, not with the downstream players. What the downstream can do is corner it into "having only one road left: scrap the contracts"—and whether that corner can be reached will get its first readings from the November 17 earnings report and Anthropic's November 18 pricing.

Five, and the most fundamental: this distribution analysis rests on "the demand side really is cracking." If the November 17 earnings report shows DSO falling and guidance raised, the entire premise that "the crack is on the buyer's side" must be rewritten—this is not a flaw, it is precisely the discipline of the ledger.

Watch list

Three tiers by horizon.

Near term: NVIDIA's November 17 earnings report (early morning Beijing time on November 18)—compare against the company's August guidance: revenue around 108 billion, plus or minus 2%, gross margin about 74% (the fourth-quarter guidance has already explicitly indicated a bottom of 71% to 72%); the next quarter's DSO figure (after 45 to 60 days, see whether it is a step or a plateau). Anthropic's November 18 pricing—which end of the $70 to $78 range it lands on (underwriters have already considered raising it above 78, per B-grade reports), and how it performs on the first day; in the third quarter it posted revenue of 16.96 billion, up 46.7% quarter on quarter, and the offering has already been fully subscribed—the first public weighing of the third party's oxygen.

Within a year: the quarterly trajectory of NVIDIA's gross margin—above 70% is the first notch, and which direction it moves each quarter thereafter; whether there is official wording about "adjusting long-term contracts" (the first signal of contract-scrapping action); the quarterly share of multi-source procurement by the four major cloud players; and verification of China's actual self-sufficiency ramp against Morgan Stanley's 35% to 68% curve.

Within two years: whether gross margin grinds below 50%; whether NVIDIA slides from selling cards to underwriting compute and running operations itself (that is the day the "bagholder" turns from narrative into a line item in the financial statements); the conclusions of the independent investigation into Jensen Huang and the three conditions for his comeback (the stock price retraces by more than half and holds there for two quarters, the successor cannot handle the renegotiation table, and the investigation finds no company-level violations).

Conclusion

Back to the question at the start: how much longer can this rent be collected?

Everyone benefits from pricking NVIDIA's bubble, and everyone cannot be the first to say it. Cloud players want to use its weakness to renegotiate procurement prices, model companies want to use it to loosen supply and change terms, politicians want to use one name to answer to the ballots, and even its own employees are asking why the board does not give them a procedure first—but everyone still needs it to keep delivering.

This is the division of labor of this era: it serves as the cushion while everyone catches their breath; it kneels to fulfill contracts while the rigidities across the whole chain unlock one by one; once the rent has fully retreated and the contracts have fully loosened, it is downgraded from compute pope to ordinary supplier—the title is lowered, the duties are not, goods still have to be delivered on schedule, only the tax is no longer collected.

Bubble monitoring continues to be recorded by the table. November 17 is the first reconciliation.

Sources: SEC filings (NVIDIA 8-K, Fermi/CoreWeave 10-Q), NVIDIA Q2 CFO Commentary, CoreWeave Q3 earnings, TSMC board resolutions, Anthropic prospectus and roadshow documents, IDC China AI chip share, Morgan Stanley self-sufficiency estimates, Ballotpedia referendum data, CNBC, WIRED, Reuters and other public reports; stock prices are Nasdaq historical data, and market cap conversions are estimates; gross margin guidance see Caixin's August 27, 2026 earnings report; Anthropic revised prospectus data see Jiemian and Eastmoney reports from November 2026; Kress-related reports originate from The Information (via secondary account).

Notes: ① Gross margin above 70% is on the data center business basis (fluctuating in the 71%–75% range across quarters of fiscal 2026, with a center around 73%); "a dozen or so points" is benchmarked against the 15–25% range of mature hardware suppliers; ② China's compute "catching up" is limited to the "good enough" level, and the gap in the most advanced process nodes is retained as is; ③ The two-step "pricking" (gross margin normalization, contract-scrapping unlock) is a mechanistic deduction and does not constitute action advice for any company; ④ The part on Jensen Huang's personal relationships cites only facts he himself acknowledged in his November 8 public letter, and the rest is not elaborated.

Source

TMTPost · AGI