NVIDIA's $96 Billion Quarter: What the Print Settled, and What It Didn't

Official member sample
This note is a complete member edition, published free so you can judge the depth before you pay: the opinion, the target, the reasoning, and the break conditions are all in the open here. Members receive editions like this across all eighteen covered names as they publish. Membership details. Designated September 7, 2026.

The Print

Correction, September 6, 2026: the original version of this table compared against incorrect prior-year figures: non-GAAP EPS of $1.05 (growth shown as +111%) and non-GAAP gross margin of 75.7% (described as roughly flat). Per NVIDIA's own Q2 FY2027 release, the year-ago quarter's non-GAAP EPS was $1.01, making growth +120%, and the year-ago non-GAAP gross margin was 72.5%, so margin expanded 2.5 points rather than holding flat. The errors came from pulling prior-year comparisons from stale coverage rather than the release's own comparative column; every comparative figure in our notes is now reconciled to the current release before publication. The corrected figures happen to make the quarter look better, which is beside the point: they were wrong, and this box is the record of it. One further precision fix, same date: an earlier version said Vera Rubin had reached full production; the release's language is that it is ramping into full production, and the text now matches it. And a structural precision fix, September 7: earlier text described the OpenAI project guarantee as collateralized. The 10-Q describes phased credit support limited to defined portions of lease and power payments, terminating if OpenAI reaches a satisfactory credit rating; that is a different protection structure than collateral, and the text now follows the filing. A sourcing precision fix, same date: the China figure is the CFO's product-specific statement (less than 1% of data center revenue shipped in H200 products to China-based customers), now quoted precisely rather than generalized to all of China data center revenue.

One quality note we would flag in either direction, so we flag it here: GAAP EPS of $2.46 came in above non-GAAP EPS of $2.22. That inversion reflects mark-to-market gains on NVIDIA's equity stakes in AI companies, which flow through GAAP income and are excluded from non-GAAP. The $2.22 is the number that describes the operating business, and it is the one we use.

The Guide That Matters

A 70% growth outlook for fiscal 2028 would be remarkable on its own; consensus sat near 44%. What makes it analytically interesting is the qualifier. CEO Jensen Huang's formulation was that the company has supply for 70% growth, while management noted customer forecasts point to demand roughly doubling. Read carefully, that is a company guiding to its manufacturing constraint while disclosing that demand exceeds it.

This distinction matters for the bubble debate. Channel-stuffing cycles guide to demand they hope materializes. Supply-rationed cycles guide to capacity they know they can build, with a demand buffer above it. The second structure is what NVIDIA described, and it is consistent with what we observe across the ecosystem: memory sold out into 2027, grid interconnection queues measured in years, and AI cloud contract pricing rising triple digits in under a year at the infrastructure providers we cover.

Vera Rubin, the next platform generation, is ramping into full production, with racks running at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, and Nebius, per the company's release. The annual architecture cadence is, in our view, the real moat: a custom chip program that takes three years to develop is aiming at a target that moves twice in that window.

The Backstop, Now With Terms

In late July, press reports of a roughly $250 billion NVIDIA backstop for a 10 GW OpenAI campus produced a sharp selloff and a wave of vendor-financing comparisons, most of them invoking Lucent, some invoking Enron. We wrote at the time that the direction of concern was legitimate, the magnitude of the reaction was not, and that the terms, when finalized, would be the test.

The terms arrived with the quarter. NVIDIA disclosed maximum customer guarantee exposure of $108.5 billion, dominated by $105 billion supporting one 4.25 GW OpenAI project in Ohio. Three observations:

First, the economics matched the reporting; the scope did not. $105 billion against 4.25 GW is about $24.7 billion per GW, essentially the same rate the $250 billion for 10 GW reports implied. What was reported as a commitment to a full campus is, in disclosed form, a capped and phased exposure to a defined first tranche.

Second, the mechanism did not spread. Roughly 97% of disclosed guarantee exposure relates to a single customer and project. The historical failure mode of vendor financing, the one that broke Lucent, is guarantees marching down the customer list until the vendor is underwriting its entire demand curve. One flagship guarantee, phased across nine construction stages and limited to defined portions of lease and power payments for powered AI capacity in a market where such capacity carries multi-year queues, is a strategic financing decision. It is now also a disclosed, measurable line investors can track every quarter. We prefer a market that can see the number to one trading on reports of it.

Third, and most important: the next $500 billion is not NVIDIA's balance sheet. NVIDIA announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR intended to mobilize more than $500 billion of capital as an alternative financing source for its customers, subject to definitive agreements. Until those agreements close and capital is actually drawn, this is announced intent, not funded demand, and we track it as such. Whatever one thinks of AI capex in aggregate, this is structurally the opposite of the Lucent pattern: the vendor originating demand while institutional credit markets, which price risk professionally and answer to their own investors, hold it. The circularity critique of July described a loop with no external capital. The disclosed architecture routes the marginal financing dollar through the largest external capital pools in the world.

The Honest Negatives

Our July commentary argued that credibility in this sector belongs to whoever names the risks before the market does. In that spirit:

The margin trough is real. NVIDIA guided gross margin to decline and bottom at 71-72% in the fourth quarter, in part on memory prices. The industry's binding constraint has migrated from chips to power to memory, exactly the sequence our constraint-migration framework anticipates, and it now taxes even the strongest player in the ecosystem. If memory inflation persists deeper into fiscal 2028 than guided, earnings estimates across the sector, ours included, are too high. This is a watch item, not a footnote.

Concentration is extreme and increasing. Data Center is 92% of revenue. The largest guarantee, the largest equity investment framework, and a meaningful slice of demand involve a single counterparty. NVIDIA is the best-protected creditor OpenAI could have, but entanglement is entanglement, and no disclosure changes that.

Crowding has not gone anywhere. July demonstrated that levered positioning in AI names can produce 40%-plus monthly drawdowns in funds while fundamentals improve. NVIDIA remains the most widely held expression of the AI trade. Investors should expect violent drawdowns on financing headlines regardless of business performance, and should size positions accordingly. That is risk management, not pessimism.

China is gone from the numbers. The CFO's formulation, quoted precisely: less than 1% of total data center revenue was shipped in H200 products to customers based in China in the quarter. We treat that as an honest baseline (any recovery is upside), but it also means the geopolitical option has already been exercised against the P&L once, and the mechanism that did it still exists.

What Would Prove Us Wrong

We publish break conditions on every framework we maintain. For our constructive NVIDIA view, the conditions are:

1. Guarantee contagion. The disclosed maximum-exposure line now makes this measurable quarterly. If customer guarantees spread materially beyond the single disclosed project toward the broader customer list, that is the Lucent signature, and it would change our view regardless of reported growth.

2. The demand ratio turning. Token volume growth across the industry has exceeded token price deflation by a wide margin (usage disclosures across the ecosystem have shown order-of-magnitude annual growth). If volume growth slows below price deflation on a sustained basis, AI compute revenue shrinks while capex is still landing, and every infrastructure thesis, including this one, breaks.

3. Profitable-buyer capex rolling over. The majority of AI infrastructure spend is funded by the operating cash flows of highly profitable companies. Those companies raised guidance again this cycle. Two consecutive quarters of hyperscaler capex guide-downs would tell us the buildout's foundation is cracking before any NVIDIA print would.

4. The margin floor breaking. Management guided a 71-72% trough. A print below 71%, or a push-out of the recovery, would signal memory costs are not cyclical but structural, and our earnings path would be wrong.

5. A cut to the fiscal 2028 outlook. The 70% guide is now the reference point. Any reduction is a material negative event, whatever the explanation.

Our Opinion on NVIDIA

Three Douglas Research maintains a constructive view on NVIDIA following these results, with an internal price target of $320 against a recent price near $218.60, implying meaningful upside if, and only if, the fiscal 2028 guide is approximately delivered and the margin trough recovers as described. At roughly 15x our fiscal 2028 earnings estimate, the market is pricing NVIDIA like a mature company in the year its own guidance describes acceleration. We could be wrong in the specific ways enumerated above, and we track each of them quarterly. This is our opinion, not a recommendation; our full reasoning, assumptions, and estimates are available in our research process description.