Archive. This is smartdisruptions.com as it looked from September 28 to October 1, 2026: the Notebook design. About this design · Go to the current site
← Back to Market Storm
Market StormPublished August 31, 2026 · by Josh Escusa

NVIDIA made $59.7 billion last quarter. $24 billion actually arrived.

AI-infrastructurecapexearnings-qualitymacrothesis
AI CAPEXThe hyperscaler buildoutRebuilt after NVIDIA’s latest quarterly report — checked against company filings through 31 August 2026

If you read nothing else

The AI spending is real. More and more of it is being paid for with promises instead of cash. NVIDIA now lends customers the money to buy its own chips, huge future bills are missing from what these companies say they owe, and record spending bought fewer computers than three months ago.

  1. NVIDIA reported $59.7 billion in profit last quarter. Only $24.1 billion actually reached its bank account — 40 cents of every dollar, down from 86 cents.
  2. NVIDIA used to get paid in 45 days. Now it is 60, and for the first time it says some big customers can take a full year. $15.5 billion of last quarter's sales is still unpaid.
  3. Seven companies have signed leases worth $1.24 trillion on data centres that have not been built. Until the buildings open, none of it counts as money they owe.
  4. NVIDIA, Meta and Alphabet promised to pay $198 billion of other companies' rent if those companies cannot. They set aside $815 million — a dollar for every $240.
  5. Computers got cheaper almost every year for eighty years. This year they got 10.9% more expensive, so a record $400 billion bought fewer of them than last quarter.

01

What happened

Two things are true at once.

The build is real and it is enormous. Seven companies have signed $1.24 trillion of rent on data centres nobody has handed over yet. Spending on computers and software hit 4.96% of the US economy — a record in eighty years of records.

And the company at the centre of it has stopped collecting most of what it earns.

For a year the risk in this build sat with the buyers — rent signed but not started, promises that do not count as debt yet. That is still true and it got bigger. What changed this quarter is on the other side of the deal, at the top of the chain, at the one company everyone else buys from.

The next section is the three ways that shows up: it lends, it guarantees, it owns. After it: the rent, the prices, the case against, and the one number that would settle it.

02

The numbers that matter

The figures the rest of this rests on, and which way each one cuts.

Cash NVIDIA collected per $1 of profit
40c
from 86c
$24.1bn of cash against $59.7bn of profit. Lowest of fourteen quarters.
Days NVIDIA waits to be paid
60
from 45
Terms of up to one year now disclosed for large customers.
Rent signed, buildings not handed over
$1.24T
seven companies
Meta $347bn · Microsoft $329bn · Oracle $260bn · and four more.
Other companies' rent promised
$198B
$0.8bn set aside
NVIDIA $108.5bn, Meta $46.0bn, Alphabet $43.8bn. A dollar per $240.
NVIDIA shares in private customers
$47.9B
from ~$2bn in 2024
Marking them up supplies about a fifth of reported profit.
Computer prices
+10.9%
largest rise on record
They fell about 17% a year through the 1990s.
What a US worker produces per hour
+0.64%
first half, annualised
Capital growing 3.0% a year against 5.6% in the dot-com boom.
Claims that survived being challenged
43 of 45
94 filings opened
Two refuted. None unverifiable. 32 needed a correction.

The one question

Every argument about an AI bubble is really an argument about who gets hurt if it stops. The doubters point at share prices, and at how fast the equipment wears out. Both were checked against the filings. Neither is where the weak spot is. It is here: the company selling the chips is now paying its customers' bills — lending them the money, owning a piece of them, and promising to cover their rent. Meanwhile $1.24 trillion of rent sits signed and uncounted. So: is this a boom paid for sensibly, against real orders already signed — or has the danger quietly moved to the one place a quarterly profit number never shows?

03

NVIDIA is paying its own customers’ bills

NVIDIA said it earned $59.7 billion in the three months to 26 July. The money that arrived was $24.1 billion.

40c

Cash collected for every dollar of profit reported. Three months earlier it was 86 cents — and this is the lowest of the fourteen quarters that can be measured.

Profit is what the accounts say you earned. Cash is what reached the bank. Both are in the same filing. NVIDIA publishes cash flow only every six months, so the quarter is a subtraction — $74,421m for the half, minus $50,344m for the first quarter. Arithmetic, not opinion.

Cash collected per dollar of NVIDIA's reported profit

It shows how much of the profit NVIDIA reports actually arrives as money, and it just fell by half in three months.

078.8157.740.3Q1 FY26Q2 FY27
Cash collected per dollar of NVIDIA's reported profit
Labelcents of cash per dollar of profit
Q1 FY26146
Q2 FY2658.2
Q3 FY2674.4
Q4 FY2684.2
Q1 FY2786.3
Q2 FY2740.3

NVIDIA cash flow statements, filed XBRL facts

There are three reasons the cash is not arriving, and they stack.

One: it lends. NVIDIA now waits 60 days to be paid, up from 45. For the first time it told investors it offers “longer payment terms ranging from 90 days up to one year” to large customers.

The label matters more than the number. It files those terms under “financing arrangements” — the exact phrase last year’s annual report used to say it had none.

How long NVIDIA waits to get paid

Two extra weeks of waiting is the plainest sign that the company selling the chips has started lending its customers the money.

032.264.459.6Apr 2024Jul 2026
How long NVIDIA waits to get paid
Labeldays
Apr 202443.2
Jul 202442.8
Oct 202445.9
Jan 202553.4
Apr 202545.7
Jul 202554.1
Oct 202553.3
Jan 202651.4
Apr 202645.4
Jul 202659.6

NVIDIA balance sheets and revenue, filed XBRL facts

At the previous quarter’s pace, about $15 billion more would have been cash instead of an IOU.

The usual explanation is shipment timing — chips sent late in the quarter, paid for early in the next. Timing does not require a new accounting label, and it does not usually arrive with a $105 billion guarantee signed nine days earlier.

Two: it guarantees. On 17 August NVIDIA promised to cover $105 billion of payments owed by a company tied to OpenAI. Counting the rest, the most it could be called on for is $108.5 billion. Money set aside against that: nothing.

Meta has promised up to $46 billion on a venture it values at $2.9 billion. Alphabet has backstopped $43.8 billion and put away $815 million.

Promised to cover, versus money set aside

Three companies have promised to cover about $198 billion of other people's rent. Between them they have set aside $815 million — a dollar for every $240 promised.

NVIDIA
nothing set aside
$108.5B
Meta, Louisiana
nothing set aside
$46.03B
Alphabet backstops
$0.8bn set aside
$43.79B
Promised to cover, versus money set aside
Label$ billions
NVIDIA$108.5B
Meta, Louisiana$46.03B
Alphabet backstops$43.79B

NVIDIA 10-Q Notes 8 and 10; Meta and Alphabet 10-Qs at 30 June 2026

$1 per $240

Money set aside against money promised, across the three companies.

None of this is hidden. It is in the filings, in plain words. It simply is not counted as debt, because under the rules it is not debt until somebody fails to pay.

Three: it owns. NVIDIA holds $47.9 billion of shares in privately held companies. Many are its own customers. Two years ago that was about $2 billion.

NVIDIA's stake in privately held companies it sells to

In under two years NVIDIA went from owning almost nothing in its own customers to holding $47.9 billion of them, and marking those holdings up now supplies a fifth of its profit.

$0B$25.9B$51.7B$47.9BOct 2024Jul 2026
NVIDIA's stake in privately held companies it sells to
Label$ billions
Oct 2024$2.2B
Jan 2025$3.4B
Apr 2025$3.2B
Jul 2025$3.8B
Oct 2025$8.2B
Jan 2026$22.3B
Apr 2026$42.3B
Jul 2026$47.9B

NVIDIA Form 10-Q, quarter ended 26 July 2026

Marking those stakes up now supplies roughly a fifth of NVIDIA’s reported profit.

Put the three together. NVIDIA sells to companies it owns a piece of, lends them the money to buy, and guarantees their rent. Each one is legal, disclosed, and defensible on its own. Together they mean a growing share of NVIDIA’s sales is money NVIDIA also supplied.

04

A trillion dollars of rent, off the books

Rent signed for data centres that do not exist yet

These seven promises add up to about $1.24 trillion and appear on nobody's books today, because the buildings have not been handed over.

Meta
Includes ~$68bn signed in July 2026
$347B
Microsoft
$329.1B
Oracle
$260B
Amazon
$137.2B
Alphabet
$85.2B
NVIDIA
$45B
CoreWeave
$35.5B
Rent signed for data centres that do not exist yet
Label$ billions
Meta$347B
Microsoft$329.1B
Oracle$260B
Amazon$137.2B
Alphabet$85.2B
NVIDIA$45B
CoreWeave$35.5B

Company 10-K and 10-Q contractual obligation tables, 2026-05-31 to 2026-07-26

Seven companies have signed leases on data centres nobody has handed over yet. Until handover, the rent does not appear in what they owe.

The previous version of this report said $830 billion across four companies. That was too narrow. Oracle alone adds $260 billion — disclosed in its own annual report as “not reflected on our consolidated balance sheet”. Meta signed another $68 billion in July.

05

Record spending, fewer computers

This is the finding that surprised this report most, and it needs no finance vocabulary at all.

Computers got more expensive.

The price of computers, against a year earlier

Computers got cheaper almost every year for eighty years, and this year they got sharply more expensive instead.

0%5.9%11.8%10.91%2023Q22026Q2
The price of computers, against a year earlier
Label% change versus a year earlier
2023Q20.19%
2023Q3-0.08%
2023Q4-0.51%
2024Q1-0.66%
2024Q20.39%
2024Q30.61%
2024Q40.85%
2025Q10.65%
2025Q21.58%
2025Q32.27%
2025Q41.8%
2026Q15.1%
2026Q210.91%

BEA national accounts, computer price index (series B935RG)

For eighty years they got cheaper almost every year — about 17% a year through the 1990s. In the year to June 2026 prices rose 10.9%, the largest jump on record.

$400bn

What America spent on computers over the year — a record. Adjusted for those prices, the quantity it received actually fell.

A record bill buying less is a different story from a record buildout.

06

What the doubters get wrong

Three popular arguments did not survive the filings.

It is not the telecoms bubble again. Spending on communication equipment and structures peaked at 1.39% of the economy in 2000. Today it is 0.69% — half. The comparison people reach for measures the wrong thing.

The record is narrower than it sounds. Computers and software together are 4.96% of the economy, a record. Strip out software and equipment is 2.45% — still below the 2.91% of late 2000.

It is not “more machines, more output”. Output per hour grew just 0.64% annualised in the first half. Capital is growing 3.0% a year against 5.6% in the dot-com boom. This is a smaller machine-buying wave than 1998’s.

07

What would settle it

The discipline: name in advance what would break each side of the case.

Next quarter’s NVIDIA cash flow statement, and nothing else.

If the money starts arriving again, this quarter was timing and the case loses its only crack. If it does not, the company at the top of the chain is paying for the demand it reports.

After that, in the order they matter: the guarantees — $198 billion promised with $815 million behind it, where nothing moves until somebody misses a payment and then all of it moves at once. The circle, where each link is defensible and the loop is the risk. The rent, $1.24 trillion landing as buildings are handed over, on a schedule nobody publishes in full. And the prices — if computers keep getting dearer, every dollar buys less and the payback gets longer.

The most useful split in this whole report is not growth. It is who pays cash and who does not. Microsoft built $115.9 billion of capacity with no new debt at all. Oracle borrowed about 98% of a 162% increase. Same buildout, opposite balance sheets — and only one of them needs the credit window to stay open.

The bull breaks if…

  • Any company has to bring a data-centre venture of the Meta type onto its balance sheet, or an auditor escalates the concern it already flagged — the off-the-books obligation becomes an on-the-books one across several giants at once.
  • The signed-but-not-started lease balances keep growing at the 2026 rate while the contracts fail to turn into cash, widening the gap between what is owed and what is earned.
  • The weakest borrowers stop being contained — ordinary high-yield borrowing costs follow the riskiest tier wider, which is how trouble travels from the edge to the buildout.
  • Amazon or Alphabet posts a second straight year of spending more cash than it makes while guiding spending higher again.

The bear fails if…

  • The economy-wide efficiency measure turns positive and stays there for several quarters — the gains show up in the national data instead of only in studies of individual tasks.
  • The borrowed share falls back toward 2025 levels as profits catch up with spending, with Microsoft the template rather than the exception.
  • Oracle-style contract backlogs start converting at the rate the filings imply, particularly the roughly 12% due within a year.
  • The signed leases start on schedule and simply appear as ordinary lease obligations without incident — the commitment was always real and always disclosed.

08

What this means for you

If you do not trade stocks, this is the part that still reaches you.

Here is the whole thing in one habit, and it takes ten seconds on any company.

Find the profit. Then find the cash. If they disagree, the disagreement is the story.

NVIDIA said it earned $59.7 billion last quarter. $24.1 billion arrived. Both numbers are in the same document, a page apart. Neither is wrong and neither is hidden. But only one of them is money, and the gap between them is customers who have the chips and have not paid yet.

That is not fraud and it is not a warning. Companies extend credit all the time. What makes it worth knowing is that NVIDIA had not done it at this scale before, said so in writing, and changed what it calls the practice in the same filing.

The second habit is about big numbers generally. When you hear a record amount was spent on something, ask whether they got a record amount of it. America spent more on computers this year than ever. Computer prices also rose 10.9% — the first real increase in eighty years. So the bill set a record and the pile of computers did not. Those are different sentences, and almost every headline uses the first to imply the second.

How this reads against the other reports

This is the report the other eight are the evidence for.

Every company write-up in this section found a version of the same thing: the number in the headline is not the number the filing supports. Microsoft's earnings growth was powered by a paper gain on a stake in a private company. Nebius has never made an operating profit, and every dollar of bottom-line profit it has reported is a revaluation of its holding in something else. Alphabet's record quarter was 69% a paper gain on shares it had agreed not to sell. Three instances of one pattern — and this report is that pattern stated directly.

Ranked by how much each company spends on buildings and equipment relative to what it earns, the ladder still holds: Palantir at 0.75%, Microsoft around 32%, Alphabet at 37.5%, Amazon at roughly 105% of its operating cash, SpaceX at 235%, CoreWeave at about 290%, and Nebius's first half at 828%. AMD sits outside it entirely — it sells the chips everyone else is buying.

What this report adds is the other half of that ladder: where the money comes from. Microsoft is at the low end and pays cash. Oracle is at the high end and borrows 98%. Same buildout, opposite balance sheets — and only one of them needs lenders to keep saying yes.

How this was researched

Every Market Storm report is produced by STORM — a multi-agent research method adapted from Stanford's STORM. Several AI agents each take a different stake (a fundamentals analyst, a short-seller, an industry engineer, a valuation watcher), interview each other while grounded in live web search, and surface the load-bearing claims. A separate skeptic pass then tries to refute each of those claims against primary sources. What survives is written up here — with the caveats it earned. The finance is the payload; the method is the point.

Who researched this

5 agents took opposing stakes, then every load-bearing claim went to a separate pass told to refute them

  1. 01Capital-markets analyst

    How the build is financed and where the obligation actually sits.

  2. 02Macro economist

    Whether the spending shows up as output, and how this compares to past buildouts.

  3. 03Semiconductor cycle analyst

    NVIDIA as the single best real-time read on whether demand is real.

  4. 04Short seller

    The specific mechanism that breaks first, tested rather than asserted.

  5. 05Plain-English translator

    What a non-specialist needs to understand and the single clearest number for each point.

Perspectives
5
Turns each
3
Claims tracked
45
Refuted-tested
45
Primary docs opened
94

Verification ledger

A separate skeptic pass tried to refute every load-bearing claim against primary sources. Where it bit:

11 confirmed32 partly-true2 corrected

Checked against the filings themselves. Forty-five load-bearing claims were surfaced and all forty-five went to agents told to disprove them, opening 94 primary documents between them — NVIDIA’s Q2 FY2027 10-Q and its 17 August 8-K, the Microsoft, Amazon, Alphabet, Meta, Oracle and CoreWeave filings, BEA’s national accounts, the San Francisco Fed’s productivity file and the BLS series. Eleven survived untouched, thirty-two needed a correction to a number, a basis or a caveat, two were refuted, and none was left unverifiable. The corrections that change what a figure means are below.

CorrectedThe previous version of this report drew the perimeter too small

It put rent signed but not started at $830 billion across four companies. Checked again with the net cast wider, it is $1.24 trillion across seven. Oracle alone adds $260 billion, disclosed in its own annual report as “not reflected on our consolidated balance sheet”. Meta signed a further ~$68 billion in July. NVIDIA carries $45 billion of its own. The direction was right and the size was understated by half a trillion dollars.

CorrectedA credit-market claim compressed its own dates

The claim said the gap between the riskiest borrowers and the rest widened “in nine sessions” and had risen steadily. It was six sessions, and the widening was a trend rather than a straight line — the gap actually narrowed in four of the twelve months. The direction holds; the tidiness did not.

CorrectedA filing sweep missed two documents and misdated a third

A claim that NVIDIA’s 10-Q was the only substantive filing in the window was wrong twice: NVIDIA also filed an 8-K on 26 August carrying the earnings release and CFO commentary, and Alphabet filed an N-PX. Amazon’s prospectus was dated two days outside the window it was placed in. Worth naming because three checks with three different start dates were mistaken for three confirmations of one fact.

Partly-trueThe guarantee is $108.5bn, and the scaling was slightly off

The $105 billion of guarantees NVIDIA signed on 17 August equals 45.9% of its shareholders’ equity, not 47.4%, and 1.86 times its cash and marketable debt securities. It is the $108.5 billion total — including the AI-cloud guarantees — that reaches 47%. The 8-K also does not say whether the obligation is direct or off-balance-sheet: it is tagged under an item covering both, and the body is a one-line cross-reference.

Partly-trueThe $15 billion of “financed” sales depends which quarter you measure from

Had NVIDIA collected at the previous quarter’s pace, roughly $15.1–15.5 billion more would have been cash. That is real, and it is baseline-dependent: the quarter it is measured against had the lowest days-to-collect in ten quarters, so it flatters the comparison. Measured against the ten-quarter average the figure is smaller. The 60-day reading is also not an all-time high — NVIDIA ran 72–75 days in mid-2022, when revenue was falling rather than doubling.

Partly-trueThe productivity gap is opening for the opposite reason to the one usually given

The gap between output per hour and underlying efficiency did widen, to 2.80 points. But it widened because efficiency collapsed, not because output surged: output per hour grew just 0.64% annualised in the first half. And the “more machines, not more efficiency” story does not survive its own data — capital input is growing 3.0% a year against 5.6% during the dot-com boom. This is a smaller machine-buying wave than 1998’s, not a larger one.

Open questions

Q1

What happens to how much debt these companies appear to carry if one data-centre venture of the Meta type has to come onto a balance sheet? Ernst & Young has already flagged the judgment, and none of the companies discloses what that change would do to their numbers.

Q2

What are Anthropic’s and OpenAI’s real economics? Neither is a public company, so there is nothing filed to read — every revenue and profit figure circulating for the two biggest model labs is unaudited and unverifiable, including the ones that would support the optimistic case here.

Q3

How much of the ~$830B of signed-but-not-started leases can actually be cancelled, and at what cost? Microsoft notes some are "subject to contractual conditions"; none of the four companies says what walking away would cost.

Q4

Does the Texas grid operator’s eligibility cull represent demand that disappears or demand that refiles? ~294 GW of a 498 GW pile failed on paperwork, and whether that is speculative junk or real projects missing a deadline changes the forecast by hundreds of gigawatts.

Research, not advice. It is the output of an AI research method applied to public information, and it may contain errors. Nothing here is a recommendation to buy or sell any security. The author may hold positions in companies covered. Do your own research.

Sources

48 documents consulted for this report — 26 of them filings or first-party disclosures. Every link was checked before publication.

Filings and primary documents

What the company told a regulator. Every load-bearing figure traces here.

  1. 1NVIDIA Form 8-K filed 2026-08-17
  2. 2NVIDIA Form 10-Q, period ended 2026-07-26
  3. 3SEC EDGAR — NVIDIA filing index
  4. 4NVIDIA Form 10-Q Exhibit 10 (agreement), period ended 2026-07-26
  5. 5NVIDIA Form 8-K filed 2026-08-17 — 0001045810-26-000069-index
  6. 6NVIDIA Form 8-K filed 2026-08-17 — sbeoainvidia-portsrelease.
  7. 7Alphabet — SEC filing d140593d424b5.htm
  8. 8Alphabet — SEC filing d32286d424b2.htm
  9. 9CoreWeave — SEC filing crwv-20260630.htm
  10. 10SEC EDGAR — Alphabet filing index
  11. 11Alphabet — SEC filing goog-20260630.htm
  12. 12Alphabet — SEC filing d171253d8k.htm
  13. 13Alphabet — SEC filing d159970dfwp.htm
  14. 14SEC EDGAR — Amazon filing index
  15. 15Amazon — SEC filing tm2619352d3_fwp.htm
  16. 16SEC EDGAR — CoreWeave filing index
  17. 17Amazon — SEC filing tm2619352-2_424b5.htm
  18. 18Microsoft — SEC filing msft-20260630.htm
  19. 19Amazon — SEC filing amzn-20260630.htm
  20. 20Oracle Form 10-K, period ended 2026-05-31
  21. 21Meta Form 10-Q, period ended 2026-06-30

Company disclosures

Releases, decks, transcripts and engineering posts — the company speaking, unaudited.

  1. 22NVIDIA IR — Investor
  2. 23NVIDIA Newsroom — NVIDIA Guarantees SB Energy’s PORTS-Pike Technology Campus in Ohio to Exclusively Host NVIDIA AI Compute
  3. 24NVIDIA IR — Default
  4. 25NVIDIA Newsroom — NVIDIA Sets Conference Call for Second-Quarter Financial Results
  5. 26NVIDIA IR — Default — default.aspx

Market and pricing data

Prices, multiples and market values, as of the dates given in the report.

  1. 27FRED (St. Louis Fed) — Fredgraph
  2. 28FRED (St. Louis Fed) — Fredgraph — fredgraph.csv?id=DGS30&cos
  3. 29FRED (St. Louis Fed) — Fredgraph — fredgraph.csv?id=A679RC1Q0
  4. 30FRED (St. Louis Fed) — Fredgraph — fredgraph.csv?id=BAMLH0A0H
  5. 31FRED (St. Louis Fed) — Fredgraph — fredgraph.csv?id=DGS30
  6. 32BEA — quarterly national accounts data file
  7. 33BEA — Seriesregister
  8. 34San Francisco Fed — Quarterly tfp
  9. 35FRED (St. Louis Fed) — Fredgraph — fredgraph.csv?id=OPHNFB
  10. 36FRED (St. Louis Fed) — Fredgraph — fredgraph.csv?id=PCU334413
  11. 37FRED (St. Louis Fed) — Private fixed investment: Nonresidential: Information processing equipment and software: Computers and peripheral equipment (B935RC1Q027SBEA) | FRED
  12. 38FRED (St. Louis Fed) — Gross Domestic Product (GDP) | FRED
  13. 39FRED (St. Louis Fed) — ICE BofA CCC & Lower US High Yield Index Option-Adjusted Spread (BAMLH0A3HYC) | FRED
  14. 40FRED (St. Louis Fed) — ICE BofA US High Yield Index Option-Adjusted Spread (BAMLH0A0HYM2) | FRED

Reporting and analysis

Third-party coverage, used for context and for checking claims against a second pair of eyes.

  1. 41idc.com — IDC - Global Memory Shortage Crisis: Market Analysis and the Potential Impact on the Smartphone and PC Markets in 2026
  2. 42Yahoo Finance — OpenAI Got $5.5B in SB Energy Warrants to Lease Nvidia
  3. 43CNBC — Nvidia backing $105 billion in financing for OpenAI data center in Ohio
  4. 44datacenterknowledge.com — Nvidia Backs OpenAI’s Ohio Data Center Buildout With $105B
  5. 45betanews.com — OpenAI and Nvidia's 20-year 8GW AI data center deal in Ohio
  6. 46wkzo.com — Nvidia to provide up to $105 billion guarantee for OpenAI’s Ohio data center | WKZO | Everything Kalamazoo
  7. 47techjacksolutions.com — Nvidia sb energy 1 5b investment openai ports pike
  8. 48fraser.stlouisfed.org — FRASER | Discover Economic History

Get the next Market Storm in your inbox

One email when a real market catalyst triggers a new report — the method, the numbers, and what the verification pass caught.

Just your email, just new builds — never sold or shared. Unsubscribe anytime by replying to any email. Privacy