On Thursday, October 22, Nvidia fell 4.4 percent. By Friday’s close of $210.96 the stock had posted its third consecutive losing week, sat 12 percent below the high it set on October 6, and was down 7.6 percent for the month, its worst month since last November (CNBC, October 24, 2026). Nothing in that week’s news explained the selling. At GTC Berlin the company announced a shipping container holding seven Vera Rubin racks, which cuts the time to bring a data center online from six months to fourteen days, and a supply-chain expansion that includes a $1.1 billion capacity advance to a Taiwanese testing house (CNBC, October 21, 2026; NVIDIA, October 21, 2026). Those are the announcements of a company that expects to sell an enormous amount of hardware.
The selling was about something else. It was about the balance sheet that has been assembled underneath the hardware. In the same month, the specialized GPU clouds that rent Nvidia’s machines, and that Nvidia has partly financed, have fallen at least 30 percent. CoreWeave’s credit default swaps now trade past 11 percent, and its 2031 and 2032 bonds yield above 15 percent. An Australian neocloud in which Nvidia holds a stake, Firmus, withdrew its initial public offering (CNBC, October 24, 2026).
So the market spent the week of GTC Berlin buying the product and selling the financing. That divergence is the whole story of the next eighteen months, and it resolves on a schedule.
What the industry has promised
Begin with the commitments, because they are on the record and none of them is disputed.
| Commitment | Figure | As of |
|---|---|---|
| Top-five hyperscaler capital spending, 2026 | about $800 billion | August 2026 earnings call |
| Top-five hyperscaler capital spending, 2027 | about $1.3 trillion | August 2026 earnings call |
| Nvidia revenue growth, fiscal 2028 (roughly calendar 2027) | about 70 percent | August 2026 earnings call |
| Cloud industry contracted backlog | more than $2 trillion | August 2026 earnings call |
| Nvidia off-balance-sheet commitments | more than $500 billion | August 2026 filing |
| Neocloud installed capacity, end of 2026 | 8 gigawatts, from about 3 at the end of 2025 | August 2026 earnings call |
Sources: the Nvidia second-quarter fiscal 2027 earnings call and CNBC, October 24, 2026. J.P. Morgan’s own research puts the 2026 figure at roughly $900 billion across five hyperscalers, revised upward from $736 billion during the year (J.P. Morgan, October 19, 2026).
Nvidia’s quarter, for scale: revenue of $96.2 billion, more than double the year before, with $89 billion of it in data centers. Guidance for the current quarter is $108 billion, and management expects to grow about 70 percent in the next fiscal year while describing that outlook as supply-constrained. Arithmetic on the reported quarter gives a run rate near $385 billion. Seventy percent on top of that is roughly $650 billion.
That is one company’s revenue. The spending it implies across its customers is larger. At $1.3 trillion of 2027 capital expenditure, the industry is committing, in a single calendar year, more than the five largest technology companies spent cumulatively in the four years after ChatGPT launched, which J.P. Morgan measures as roughly equal to the 2026 figure alone.
What has to arrive to pay for it
The revenue side of the ledger has one genuinely spectacular data point, and it deserves to be stated before any caveat. OpenAI’s annualized revenue reached nearly $70 billion by late September, having risen more than 70 percent since the beginning of the third quarter (Reuters via Investing.com, September 29, 2026). A year earlier it was above $20 billion (TechCrunch, November 6, 2025). Tripling in twelve months at that scale is not a story about demand failing to materialize.
It is, however, a story about a gap that is closing more slowly than the spending is growing. In February the company told investors it planned to spend about $600 billion on compute through 2030, against roughly $280 billion of projected revenue in that year, after disclosing $1.4 trillion of infrastructure commitments months earlier (CNBC, February 20, 2026). In September the Financial Times reported a plan to burn $278 billion in cash between 2026 and 2030 (Reuters, September 18, 2026). The company is financing the gap with a $30 billion round at a valuation near $1.4 trillion, described as bridge capital in place of an IPO.
Set one consultancy’s arithmetic beside that. Bain & Company puts the annual revenue required to fund AI infrastructure at about $6 trillion by 2031, against roughly $1.5 trillion of annual infrastructure spending. Its own accounting of where that revenue can come from credits existing applications with $1.2 trillion to $1.8 trillion. The remaining $4.2 trillion is assigned to products that do not exist yet: advertising inside models that replace search, autonomous trucking and drones, physical AI, and a final category Bain describes as new products and uses that do not exist today (Bain & Company, September 29, 2026).
Read that last line without flinching. The largest single line item in the industry’s own justification is a category with no contents. That is not a criticism of Bain, which is doing the honest thing by naming the residual rather than hiding it. It is a description of the position: the buildout is a wager that innovation arrives on schedule, and the schedule is now short enough to be dated.
The bridge, and who stands on it
Between the spending and the revenue sits credit, and Nvidia has moved into the middle of it.
The company’s July filing disclosed $36 billion of off-balance-sheet agreements backing neoclouds, a guarantee of up to $105 billion on the OpenAI data center in Pike County, Ohio, and $20 billion of future leases signed as a tenant. Total off-balance-sheet commitments exceed $500 billion, most maturing by fiscal 2029. Its equity portfolio was worth $99 billion at the end of the second quarter. In August it partnered with six asset managers on a $500 billion financing effort that treats compute as an asset class, including up to $125 billion of residual-value guarantees on GPU financings (CNBC, October 24, 2026).
None of this is improper. A supplier with $70 billion of half-year free cash flow and an investment-grade rating can lower its customers’ cost of capital, and lowering it is one way to sell more machines. The question is what the guarantees are worth if the thing they guarantee stops being valuable.
Huang’s own framing of the constraint has shifted. In Berlin he said, “The demand of the capital is much greater than the availability of the capital” (CNBC, October 23, 2026). That is a sentence about financing, not about chips. A year ago the constraint was supply. Now the constraint is the willingness of lenders to fund a business whose collateral depreciates on somebody else’s schedule.
David Trainer of New Constructs put the risk in the plainest form anyone has: “If the demand for compute doesn’t meet these really super lofty expectations, we’re going to have a problem.” His firm calculates the net present value of Nvidia’s off-balance-sheet liabilities at more than $153 billion, up by over $100 billion in a year. Moody’s says guarantees are the fastest-growing form of off-balance-sheet financing in 2026, and that the exposure investment-grade companies can carry is under daily review (CNBC, October 24, 2026).
Why the reckoning is not now
A forecast that ignores the bull case is a mood, not an argument. Four facts defend the current quarter, and three of them are hard.
Capacity is sold. Amazon’s chief executive has said the company is sold out for 2026 and largely for 2027, and Nvidia describes its own outlook as limited by supply rather than demand. Contracted backlog across the cloud industry exceeds $2 trillion. A business with two years of signed orders does not fail on a Tuesday.
Revenue is compounding fast. The OpenAI number above is the strongest single refutation of the bear case, and it improved by more than 70 percent in a single quarter.
Margins are intact. Nvidia guides to 74 percent gross margin this quarter, a trough of 71 to 72 percent next quarter on rising memory costs, and 72 to 73 percent for the fiscal year after. Those are the numbers of a company with pricing power, not one discounting into a glut.
And the depreciation argument, the one I made in July, has not yet reached an income statement. Accounting changes land on a schedule, and the schedule is set by product transitions.
Why the second quarter of 2027
Here is the calendar. Every date in it is either scheduled or expected, and each one narrows the question.
| Date | Event | What it decides |
|---|---|---|
| Expected October 28 and 29, 2026 | Alphabet, Microsoft, Meta, and Amazon report | First guidance on 2027 capital spending |
| November 17, 2026 | Nvidia’s third quarter | Whether $108 billion arrives, and what margins do |
| Late January to February 2027 | Hyperscaler fourth quarters; Nvidia’s fiscal year-end report, expected February 24 | The full 2027 capex number, and Nvidia’s fiscal 2028 guide in writing |
| First half of 2027 | DSX Pod deliveries begin; Vera Rubin ramps in volume | Whether supply stops being the binding constraint |
| May 19, 2027 (expected) | Nvidia’s first quarter of fiscal 2028 | The first honest test of the 70 percent growth claim |
| Second quarter of 2027 | The collision | Whether the revenue that justifies 2027’s spending is visible |
Dates for Nvidia’s reports from the Wall Street Horizon earnings calendar; the fiscal 2028 growth figure and the 2027 capex estimate from the August earnings call.
Three mechanisms converge in that window, and they reinforce one another.
Supply stops being the excuse. Nvidia says its customers’ forecasts point to growth doubling next year, and that it will deliver about 70 percent because it cannot make more. That is a defensible position while capacity is scarce. Once Rubin volume ships and containerized factories cut installation to two weeks, scarcity stops explaining a shortfall. The first quarter in which a miss can only mean demand is the one reported in May 2027.
Depreciation arrives. Nvidia says Vera Rubin delivers 30 times the throughput per megawatt and 35 times lower token cost than Grace Blackwell Ultra. Every one of those multiples is a statement about the economic value of the previous generation installed in the field. For a hyperscaler that owns its fleet, the effect shows up as an accounting charge and a shorter useful life. For a lessor whose loan is secured by that fleet, it shows up as collateral. Both effects land in the same year the new machines ship, which is 2027.
Credit reprices first. CoreWeave’s swap spreads widened during October, not during a recession, but during the week its supplier held a victory lap. If spreads stay at that level into the 2027 refinancing window, the marginal neocloud pays a rate at which no arithmetic works: borrow near 15 percent to earn a return on capital deployed near 9 percent, against hardware that loses value on a three-year cadence. The companies that cannot refinance do not announce it. They simply stop ordering.
Add one more pressure that is not about AI at all. Hyperscalers’ cost of capital has risen while their spending has, and every point on the discount rate makes a ten-year asset with a three-year economic life harder to justify. The industry has been rescued so far by the fact that its largest customers fund construction from cash flow. That rescue has a limit, and the limit is the first year in which free cash flow turns negative at a company whose stock is priced for growth.
What would prove me wrong
I hold this loosely enough to name the evidence that would move it.
If all four hyperscalers raise 2027 capital spending in their fourth-quarter calls while Nvidia lifts its fiscal 2028 guidance above 70 percent with margins holding near 73 percent, then demand is real, supply is the only constraint, and the wall moves to the right.
If CoreWeave’s credit spreads narrow back below 5 percent and its bonds trade above par, the funding bridge is working and the credit channel is closed.
And if OpenAI’s annualized revenue passes $150 billion before the middle of 2027 while its burn flattens, the gap between spending and revenue is closing faster than any of the arithmetic above assumes, and the whole argument needs revising rather than defending.
What would confirm the forecast is narrower and easier to see: a hyperscaler holding 2027 spending flat rather than raising it, a neocloud failing to refinance, or a GPU lessor marking down the residual value of a Blackwell fleet. Any one of those in the first half of 2027 turns the divergence of October into a repricing.
What a careful reader does with this
The honest core of the argument is not that artificial intelligence is worthless. It is that the industry has financed a five-year build against a revenue curve that arrives later than the debt does, and that the difference has been transferred to people who did not choose it: private credit funds, the pension money behind them, and index investors whose retirement accounts hold the seven companies at roughly a third of the S&P 500.
A reader cannot un-concentrate an index alone, but a reader can know what they own. The useful question between now and May is not whether AI works. It is which of the numbers above moves first, and the answer will be visible in a bond price before it is visible in a headline.
The wall is not made of concrete. It is made of contracts dated 2027, and the calendar is already set.
Sources
- Bain & Company. (2026, September 29). Global AI market could hit $6 trillion annually by 2031 through unlocking value and innovation.
- CNBC. (2026, February 20). OpenAI resets spending expectations, tells investors compute target is around $600 billion by 2030.
- CNBC. (2026, October 21). Nvidia announces plug and play Vera Rubin racks that can be shipped on a truck.
- CNBC. (2026, October 23). On the ground at Nvidia’s AI showpiece: Jensen-mania, robot ducks and neocloud distress.
- CNBC. (2026, October 24). Nvidia’s role as AI financier faces first big test as neoclouds take market hit.
- J.P. Morgan. (2026, October 19). What unprecedented AI capex means for markets and investors.
- MarketBeat. (2026, August 26). NVIDIA Q2 2027 earnings call transcript.
- NVIDIA. (2026, October 21). NVIDIA and supply chain partners expand capacity to accelerate AI infrastructure buildout.
- Reuters. (2026, September 18). OpenAI forecasts cash burn near $280 billion by 2030, FT reports.
- Reuters via Investing.com. (2026, September 29). OpenAI targets $30 billion funding at $1.4 trillion valuation, Bloomberg News reports.
- TechCrunch. (2025, November 6). Sam Altman says OpenAI has $20B ARR and about $1.4 trillion in data center commitments.
- Wall Street Horizon. (2026). Nvidia earnings date and conference call calendar.
PRH | huffmanwrites.org | © Philip Huffman