Nvidia Now Underwrites the Demand It Forecasts

I.The quarter that broke its own pattern

Nvidia has spent a decade refusing to guide beyond the next quarter. On 26 August 2026 it stopped. Alongside record second-quarter results — $96.2 billion of revenue, up 106% year over year, with $89.0 billion from data centre — management projected roughly 70% revenue growth for fiscal 2028 and $108.0 billion for the current quarter. Jensen Huang was explicit that this was new: the company had never guided a year in advance before.

The arithmetic is startling. Fiscal 2027 is tracking to roughly $410 billion on reported and guided quarters; 70% growth implies something close to $700 billion in fiscal 2028, against a sell-side consensus nearer $570 billion. That is a company adding, in one year, more revenue than Nvidia earned in total two years ago.

The interesting question is not whether the number is achievable. It is why a company with a $2 trillion backlog and no competitive threat felt the need to say it out loud.

II.Where the cash didn’t go

The answer sits three pages into the CFO commentary. Days sales outstanding — how long Nvidia waits to be paid — rose to 60 days, from 45 the prior quarter. Accounts receivable reached $63.1 billion, roughly two-thirds of a full quarter’s sales sitting uncollected. Operating cash flow was $24.1 billion against $59.7 billion of net income, and free cash flow fell to about $21.3 billion from roughly $48.6 billion three months earlier. Revenue grew 18% sequentially; cash collection went backwards.

Some of that gap is benign. Inventory rose to $31.6 billion, but the growth is in raw materials and work in process rather than unsold finished goods — a company hoarding inputs, not stuck with product. Some of it is not. Analysis of the filing notes that Nvidia’s February annual report said it had entered into no customer financing arrangements; six months later it discloses terms running from 90 days to a full year for investment-grade buyers at data-centre scale, and warns these will keep affecting the timing of operating cash flows.

III.Nvidia became the credit

Extended terms are the small version of the story. The large version is contingent. Nvidia disclosed $108.5 billion of maximum guarantee exposure: $105.0 billion capped credit support tied to SB Energy’s Ohio campus hosting OpenAI compute, phasing across nine construction stages with twenty-year lease terms per phase, plus $3.5 billion behind AI cloud partners’ leases. Against that $3.5 billion, partners have escrowed a reported $712 million. Separately, the company has committed $25 billion to equity investments in AI firms, and holds $279 billion of supply and capacity purchase commitments — up from $119 billion — against $320.3 billion of total assets.

Nvidia’s own risk language explains the logic: a meaningful part of its addressable market cannot yet secure long-term infrastructure contracts or investment-grade financing on its own. So Nvidia lends its rating to the transaction. The guarantee on the Ohio campus releases only when the tenant reaches a satisfactory credit rating — which is, in one clause, the whole thesis. Revenue lands within two years; the credit risk runs for twenty.

IV.The binding constraint moved from compute to credit

For three years the limiting factor in AI was silicon: who could get allocation. That constraint has largely cleared, and a different one has taken its place. Kress told analysts she expects hyperscaler capital expenditure above $800 billion in 2026 and $1.3 trillion in 2027, with a backlog above $2 trillion behind it. Nobody funds that from operating cash flow.

So it is being funded from the bond market and from structures that never touch a balance sheet. Hyperscaler-related bond issuance reached about $225 billion by mid-2026, a roughly tenfold jump year to date, with full-year issuance projected near $400 billion. A Nikkei study cited in the same reporting puts “hidden” obligations — long-term GPU and server purchase agreements plus data-centre leases, disclosed in footnotes rather than as debt — at $1.65 trillion, above the $1.35 trillion carried on balance sheets; Moody’s separately identifies $1.2 trillion of off-balance-sheet liabilities. The market has noticed: one investor on earnings day observed that Nvidia’s credit spreads trade more like a BBB issuer’s than an AA issuer’s.

Read that way, the year-ahead forecast is not guidance. It is collateral. A public 70% number from the sole supplier makes every lender, lessor and IPO underwriter in the chain more comfortable financing the counterparty — which is precisely what the chain needs to keep moving.

V.What this does to the wider economy

This is now a macro exposure, not a sector story. Epoch AI estimates AI-related data centre construction, compute hardware and networking equipment reached about 0.8% of US GDP in Q1 2026, with computing infrastructure overall near 1.5% — roughly double the 2015–2022 norm, and the leading driver of growth in US private investment.

The transmission channels are worth separating. Capital flows are the loudest: the buildout is absorbing an unusual share of investment-grade issuance, which raises the cost of capital for unrelated borrowers. Prices are moving through memory, where high-bandwidth supply is a live constraint. Productivity is the quiet one. Federal Reserve economists concluded in a 17 July 2026 FEDS Note that micro-level productivity gains are not yet adding up in aggregate, that reported adoption is broad but shallow in usage intensity, and that historically such gains lag investment by years. On labour, their read is that AI is showing up in slower hiring of younger workers rather than layoffs, with 20–24 year-old unemployment the series to watch.

The uncomfortable arithmetic: capital is being committed at a pace that assumes the productivity payoff, while the official statistics cannot yet see it. That gap is normal for a general-purpose technology. It is also exactly what makes the financing structure matter — because debt has a schedule and diffusion does not.

VI.The demand underneath is real, and younger than the paper

None of this means the revenue is imaginary. It is the opposite: the end demand is growing faster than any software business on record. Anthropic’s annualised revenue passed $65 billion at the end of July 2026, from $47 billion in May and about $9 billion at the end of 2025; OpenAI has roughly doubled to $40 billion over the same window. Both have filed confidentially for IPOs, with Anthropic potentially pricing this autumn.

The mismatch is duration, not direction. A twenty-year lease guarantee is being written against counterparties whose revenue history is measured in quarters and whose comparability is contested — the two companies do not necessarily calculate run-rate the same way. Huang’s defence is that agentic workloads consume 15–100 times the compute of a human user, so today’s demand curve understates tomorrow’s. He may well be right. The point is that if he is wrong, the loss no longer lands only on the buyer’s balance sheet.

VII.How it plays out: three scenarios

Horizon two to five years. Probabilities are analytical judgment, not forecast output, and sum to 100%.