Nvidia's $96 Billion Quarter Was a Supply Story — and the AI Trade Just Heard It

Nvidia (NVDA) reported its fiscal second-quarter results after Wednesday's close, and the tape said more than the headline. Revenue was $96.2 billion, up 106% from a year earlier and about $4 billion ahead of the $92.2 billion analysts expected. Nvidia shares rose about 4% after hours, to roughly $220. The more revealing move was in a smaller name: Nebius Group (NBIS), an Amsterdam AI-cloud company that buys Nvidia's chips and rents out the compute, had closed the regular session down 3.6% — then jumped more than 6%, to roughly $227, once Nvidia's numbers landed.

Decompose that $96.2 billion and the quarter becomes a different event. The data center business — servers, networking, and the chips Nvidia sells for AI computing — generated $89.0 billion, up 117% from a year ago. That is roughly 92 cents of every dollar Nvidia earns. Everything else it still makes, gaming graphics chief among them, added up to about $7.2 billion, around 7% of revenue. Nvidia is no longer a chip company with an AI division; it is a single-segment AI-infrastructure company. On that basis non-GAAP profit was $2.22 a share against a $2.10 consensus.

The part of the report that mattered to the whole AI trade was Nvidia's explanation of why it grew, because 2026's bear case against AI stocks was, in essence, overspending: hyperscalers and cloud startups borrowing enormous sums to build computing capacity they may later be forced to discount or idle. The earnings call answered with a different bottleneck. Management said Nvidia has supply for roughly 70% growth, that demand runs "much higher" than that supply, and that "the entire supply chain is under strain." The current-quarter outlook makes the same point in numbers: Nvidia guided to $108 billion — up about 12% sequentially and close to double the $57 billion it posted for last year's third fiscal quarter — and that forecast explicitly assumes zero data-center compute revenue from China, the biggest market export controls block it from serving. Remove China from the forecast entirely and the business still accelerates. The constraint is manufacturing capacity, not customers.

Even a supply story has a price, and it shows up in a line most readers skim: gross margin. Nvidia kept 75 cents of every revenue dollar in the quarter — a 75.0% gross margin — and guided it down to about 74% for the current quarter. That one-point drift is the accounting trace of the strain its chief executive described, the cost of paying up for memory and foundry capacity in order to buy the bottleneck open.

Now to Nebius, because that six-point jump is the read-through. Nebius sits on the opposite side of the same chain. Nvidia prices the scarcity and pockets the margin; Nebius raises capital to buy the hardware and resell it by the hour. Its entire model is a leveraged bet on precisely what Nvidia confirmed: that demand for Nvidia-grade compute outruns supply. The confirmation is not arm's-length. Nvidia owns a 9.3% stake in Nebius, and the earnings release named Nebius among the partners already running racks of the new Vera Rubin platform, alongside Google Cloud, Microsoft Azure, Oracle, and CoreWeave. Nebius stock came into the report about 28% below its late-June high of $299.86, damaged through the summer's AI-infrastructure selloff. A demand signal landing in a beaten-down, high-beta name moves it further than it moves the company that printed the signal — the same reason observers framed the night as Nvidia's number outpacing "fears about AI overspending."

But the structure that produced the pop also locates the risk. Nebius's own June-quarter report, two weeks earlier, was dramatic and incomplete at once: revenue up 454% to $582 million, AI-cloud revenue up 514%, adjusted EBITDA swung to a $236 million profit — yet still a net loss on the bottom line. The buildout is funded with priced capital: a $5 billion convertible-note sale priced a week before Nvidia's report, adding to $775 million in secured debt Nebius raised in July against its GPU infrastructure and contracted customer cash flows. Convertible notes are debt a holder can swap into new stock, so if the shares stay high the raise becomes dilution. Set that ledger against Nvidia's: it returned $26 billion to shareholders last quarter and still holds roughly $99 billion of buyback authorization. The seller of scarce chips sends cash back to its owners; its biggest customers raise cash to buy the chips. Nvidia has also increasingly helped its own customers get funded — structuring and backstopping deals so new data centers can be built — which means some of the demand it books is demand it helped finance. The circle has a calendar date attached: Nvidia's Nebius stake unlocks in September, the scheduled moment when a holder of that size becomes free to sell.

Read the after-hours prices as a measurement. Nvidia up about 4%; the operators up about 6%. The gap is the difference between owning the company that prices the scarcity and owning the businesses that are financed on the scarcity persisting. The quarter pushed the demand question — the central fear of the summer selloff — out by another quarter. It did not answer who absorbs the cost if the buildout slows. That moves to observables: whether Nvidia can source enough hardware to actually ship $108 billion and whether the one-point margin dip at 75% becomes a trend as memory and foundry prices climb; and, on the operators' side, the September unlock, the dilution schedule of the $5 billion convertible, and whether the next round of financing is available on terms anywhere near these. The earnings settled how much demand exists. The financing question — which layer of the chain carries the risk — is the one that decides who gets paid, and it is still open.