Nvidia booked $96.2 billion in revenue for the quarter, up 106% from a year earlier, with $89 billion of it from data centers. Jensen Huang broke his own practice of declining to comment on future demand and guided to 70% data center revenue growth in fiscal 2028, a year that unusually starts at the end of January 2027. Chip supply is short of what buyers want, and the near-term worry about demand cooling is dead. The disclosures worth reading sit further into the 10-Q, where Nvidia's own balance sheet has become part of how the AI buildout gets financed.
Four things were worth checking going in: the state of data center demand, how cleanly the Vera Rubin transition is going, customer concentration, and the structure of ecosystem financing. Demand accelerated in both of Nvidia's largest customer groups. Large cloud providers bought $48.7 billion of product, up 13% from the prior quarter. Revenue from AI cloud services, industrial customers and enterprises rose 25% for the quarter and 138% for the year, to $40.3 billion. The second group growing faster than the first is the meaningful part: the acceleration is not confined to the hyperscalers, who are the ones underwriting the frontier labs, Anthropic and OpenAI among them. Together the two groups make up the data center customer segment.
On the earnings call, CFO Colette Kress described the 70% fiscal 2028 growth figure as a supply-constrained forecast. In a CNBC interview with Jim Cramer, Huang said demand remains very high and is still accelerating.
The criticism in the numbers is margin. Gross margin falls from 75% to 74% in the current quarter and then settles at 71% to 72% in the fourth, and the reason is a memory shortage pushing costs up.
Vera Rubin, the platform replacing Blackwell, reached full production in August 2026, with working systems already installed at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius. It should account for 20% of data center revenue this quarter, which Kress called the fastest product ramp in Nvidia's history. Buyers are not holding back purchases to wait for the new architecture, and Rubin's arrival did not stall Blackwell demand, which is the failure mode a generational transition usually produces. What is left is execution: turning a large forecast into shipped systems while the supply chain runs at its limit and the company tries to stand up customer deployments and protect margin at the same time.
The concentration numbers are where the filing gets less comfortable. One customer accounted for 16% of revenue last quarter. Five customers accounted for 70% of receivables. Average collection has stretched to 60 days, and for certain investment-grade customers Nvidia has extended payment terms out to a full year. Extending credit is not by itself a sign of weak revenue quality; a financially strong supplier can put its balance sheet to work, granting deferrals or accepting slower collection to move more volume, and that credit helps customers fund their purchases. It also moves the risk onto Nvidia's books. The issue here is scale and concentration, and the thing to watch next quarter is whether the customer base broadens past a handful of buyers.
Then there are the direct commitments. The 10-Q says Nvidia is helping customers get access to data center capacity, land and power, with $36 billion set aside for AI cloud services obligations, an activity whose outcome Nvidia does not control. That sits alongside a further $366 billion in future obligations, of which $279 billion is supply and capacity. Nvidia also provides guarantees to customers with maximum potential obligations of $108.5 billion, and $105 billion of that relates to a single 4.25-gigawatt OpenAI project in Ohio. To spread the risk, Nvidia recently signed agreements to build an alternative source of capital for customers, working with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR toward more than $500 billion. Kress addressed the circular financing concern directly: third-party capital providers will assess risk on their own, and Nvidia's compute holds its value and can be redeployed across projects.
That last exchange is the one I would push on. The $500 billion vehicle changes who supplies the money, not who absorbs a shortfall. A guarantee is a guarantee no matter how many co-investors stand next to it, and $105 billion of the $108.5 billion guarantee book rests on one site for one customer. Every other concentration figure in the filing is milder than that. One customer at 16% of revenue is a normal risk for a supplier in a boom; five customers at 70% of receivables is tighter, because receivables are the money not yet collected; the guarantee book is a single name. Kress's answer is about the quality of the counterparties, which is a reasonable thing to say about Apollo or KKR and a separate question from what happens if 4.25 gigawatts in Ohio does not earn what its sponsor projected.
The margin path deserves the same scrutiny. Calling 70% growth a supply-constrained forecast turns guidance into a floor, and that framing holds precisely as long as memory is the binding constraint. Memory is also what is taking gross margin from 75% to 74% to 71-72%. The same shortage that lets Nvidia present its outlook as conservative is costing it three to four points of margin on the way there, and if memory eases, the conservatism argument eases with it.
Demand was the question going into the print and it was answered: the buildout does not hit a demand wall in the coming months, the platform transition is landing without technical trouble, and the growth is spreading beyond the hyperscalers. What accumulated instead was exposure. The number to track next quarter is not revenue. It is whether the receivables concentration comes down while the guarantee book stays where it is.