Nvidia's 70% FY2028 guide: real supply-constrained demand or partly self-funded?

The story works until this line in the quarterly report Nvidia filed alongside Wednesday's earnings. The company just issued its first-ever year-ahead forecast — roughly 70% revenue growth in fiscal 2028, a trajectory toward $700 billion in annual sales — and the shares rallied more than 4% after hours. The same 10-Q, in the notes that no headline quotes, disclosed two quieter numbers: $366 billion of future purchase commitments, and up to $108.5 billion of guarantees on other companies' data-center leases.

Set the units before anything else. Nvidia's fiscal 2027 ends in January 2027; "fiscal 2028" is calendar 2027. Growth of about 70% on a roughly $410 billion base puts the guide around $700 billion — about $100 billion above the $570 billion that analysts had been modeling, according to Fortune's coverage. Management framed the number as a ceiling imposed by manufacturing, not by customers. Jensen Huang's line, as reported: demand is "much greater than 70%," but supply "allows us to confidently deliver 70%." For a company that had never guided a year ahead, that is a statement of stunning certainty.

It is worth checking where that certainty lives before believing it. It lives in three stacks on the balance sheet.

The commitment stack. As of July 26, Nvidia's outstanding future commitments totaled $366 billion: $279 billion of supply-and-capacity agreements (memory, wafers, manufacturing), plus cloud-service agreements, not-yet-started data-center leases, committed equity investments, and capex. The supply number alone nearly tripled in half a year, from $95.2 billion at fiscal 2026 year-end. The same sixfold surge from $16.1 billion to $95.2 billion a year earlier drew a warning from Michael Burry, who compared it to Cisco's dot-com-era purchase commitments. These contracts are how Nvidia "knows" it can deliver 70%: it has signed away the money to buy the parts. The supply constraint is therefore partly a decision — Nvidia sized the guide to supply it has already locked, not to an unmeasured demand pool.

The demand side: who pays. Here is the part that muddies the "external demand" story. Over the first half, Nvidia's equity portfolio grew to roughly $94 billion — $42.8 billion of marketable equity securities plus $51.2 billion of non-marketable stakes — versus about $35 billion in January. It spent $42.4 billion buying equity securities in six months, investing in the very buyers of its chips: about $30 billion into OpenAI, plus stakes in CoreWeave, SK Group, infrastructure and glass suppliers. On top of that sit the $108.5 billion of guarantees — $105 billion of capped residual-value support behind a SoftBank-run Ohio campus leased by OpenAI, and $3.5 billion backing AI-cloud partners' lease obligations — and arrangements like the CoreWeave deal where Nvidia is obligated to buy unsold capacity. Then consider the composition of demand: Huang said non-hyperscaler buyers — AI labs, neoclouds, sovereign projects — are now about half the business and growing 100% a year, and Kress said frontier AI labs will contribute about a quarter of next year's revenue. The fastest-growing half of the revenue is exactly the segment Nvidia's own equity and guarantees help finance. That does not make the demand fake. It makes it partly vendor-supported.

The earnings optics. Follow the same dollars into the income statement. In the first half, Nvidia reported $118 billion of GAAP net income — but $23.7 billion of that was non-cash gains from equity securities, paper markups on the same portfolio that finances its customers. Strip them out and roughly $0.81 of the $4.85 in first-half GAAP diluted EPS disappears, about seventeen cents of every reported dollar. The tell showed up in the second-quarter numbers: GAAP EPS of $2.46 exceeded non-GAAP EPS of $2.22 — technically unusual, because adjusted earnings normally sit above reported. Non-GAAP strips the investment winnings; the official number counts them.

The conservation problem. Now the cash-flow bridge, because earnings can be marked to model but cash has a harder time being narrated indefinitely.

First half FY2027 ($B)
Operating cash flow	74.4
minus capital expenditure	(4.4)
Free cash flow	70.0
Net purchases of equity securities	(35.2)
Cash dividends	(6.3)
Share repurchases	(39.0)
New debt issued (June 2026 notes)	+24.9

Free cash flow of $70 billion could not cover $35 billion of ecosystem equity purchases plus $45 billion of dividends and buybacks. The roughly $10 billion gap was bridged by $25 billion of new senior notes issued in June. In other words, the shareholder cash program is already partly financed by borrowing. That matters because the buyback runway — about $99 billion remaining, roughly fifteen months at the first-half pace — is the discretionary pool that a guarantee draw, an equity write-down, or a purchase-commitment provision would have to drain. The question for a buyer is not whether Nvidia can afford a $700 billion year without help. It is how much of the price depends on the same balance sheet that manufactures a fifth of reported earnings.

The company response, in its strongest form. Huang says demand genuinely exceeds what Nvidia can build. Kress acknowledges the criticism directly — "some will call this circular financing," she said, per the call — and answers that "independent capital still underwrites every deal on its own merits," and "we're not making loans." The $500 billion financing platform announced with Apollo, BlackRock, Blackstone, and others uses third-party money, not Nvidia's. And the heaviest buyers need no help: the four hyperscalers are expected to spend about $1 trillion a year on AI infrastructure by 2027, up from roughly $630 billion planned for 2026 — that is external, cash-funded demand that converts to Nvidia revenue through ordinary purchases. Memory costs are genuinely tight, which is why Nvidia is passing through roughly 15% price increases even as gross margin is guided down to 72%-73% next year. That scarcity is real.

But part of the response reframes rather than reconciles. The strongest single counterweight is the OpenAI guarantee: OpenAI indemnifies Nvidia for any guarantee payment, yet the indemnity is a claim against the same entity whose default would trigger the payment. In the conservative number that still matters, Nvidia's maximum gross guarantee exposure is $108.5 billion before it sells or re-lets a single data center — about 10% larger than the entire buyback authorization.

What would prove the guide durable versus vendor-supported? Evidence, not vibes, and the next few filings will supply it. Four checks.

First, guarantee draw disclosures: any quarter in which Nvidia must actually perform under the OpenAI or AI-cloud guarantees is a Level 4 event. Absence of draws is the benign reading — but watch whether the guarantees grow faster than the revenue they enable. Second, the purchase-commitment line: Nvidia provisions for excess purchase obligations through cost of revenue; the H20 fiasco in fiscal 2026 created a $4.5 billion provision. If the $366 billion stack of commitments converts to revenue, the provision line stays small; if demand softens, it swells. Third, collections: account receivables rose $24.6 billion in the first half to $63 billion, and Nvidia is extending payment terms to investment-grade customers out to one year. Watch whether the financed cohort's cash arrives on time or whether receivables keep growing faster than revenue. Fourth, the buyback ledger: repurchases funded from operating cash flow are the sign of sustainable demand; buybacks funded by new debt (as partially happened in June) are the sign of a covenant on borrowed time.

Three cases, and the shareholder invoice runs through all of them. In the resolved case, hyperscaler capex converts to cash, commitments convert to revenue without big provisions, no guarantee is drawn, and buybacks continue from cash flow — then roughly $700 billion of fiscal 2028 revenue at a 40%-45% free-cash-flow conversion produces around $300 billion of free cash flow, a mid-single-digit FCF yield on a ~$5 trillion market value, and the guide is genuinely durable. In the persistent-but-lawful case, growth continues while the financed cohort never quite cash-converts on its own: some provisions appear on commitments, equity gains stop growing so GAAP EPS decelerates as paper winnings fade, and the $99 billion buyback gets funded increasingly with debt — the red flag becomes a margin-then-multiple story. In the materially misstated case, a trigger event — OpenAI insolvency, or a broader crack in the closed loop — produces guarantee draws and portfolio write-downs that divert cash from the buyback and turn the FCF-yield arithmetic negative; that is a tail risk, not a finding, and treating it as established would be exactly the mistake this genre punishes.

Level of the evidence: two (anomaly with corroborating comparison) toward three (a documented discrepancy across independent records — the GAAP-versus-non-GAAP inversion, the equity-gains share of net income, and the cash bridge that required new debt). Not fraud, not accusation: Nvidia's numbers add up; they just add up on a balance sheet that is also buying the demand and then booking the paper gains. The next settling documents are already scheduled: the first 10-Q that discloses a guarantee draw or a swelling provision line, and the first clean quarter where buybacks are funded by operating cash rather than new bonds. Watch the number that made this story — the $23.7 billion of non-cash equity gains — and ask yourself each quarter how much of the EPS you are being shown is actually cash that arrived.