NVIDIA's $96B Quarter, and What It Costs to Deliver

Last week I mapped five ways the AI trade could break. This print is the first real checkpoint, and the first thing to say is that the bull case won the quarter. Revenue grew 106% to a record $96.2 billion, and management did something it has refused to do for two years: it guided the full year ahead, to roughly 70% fiscal 2028 revenue growth, and said supply, not demand, is the limit. That is the headline, and it deserves to be the headline.

The Q3 revenue guide also beat, $108 billion against a street near $103.9 billion. The one soft note was the 74% margin guide, below the ~75% the street wanted and a one-point sequential step down. Underneath the beat, NVIDIA extended customer payment terms, built working capital and added commitments and guarantee exposure around the buildout. The growth is real and enormous. This piece is about the growth first, and then about what it is costing NVIDIA to deliver it.

Headline numbers vs. consensus

The growth is the story

Start with the magnitude, because it is easy to lose in the detail. Revenue was a record $96.2 billion, up 106% year over year, a company this size still doubling, which almost nothing at this scale does. Data Center did $89.0 billion of that, up 117%. EPS was $2.22, up 120%.

Then the part that moved the stock. Management guided the full year ahead for the first time, to roughly 70% fiscal 2028 revenue growth, and framed the number as supply-limited rather than demand-limited. Jensen Huang said it flat: “even though our demand is much greater than 70%, our supply allows us to confidently deliver 70%.” Read that carefully. The 70% is what NVIDIA can build, not what customers want to buy. Demand is higher. Vera Rubin is now in full production, which is what makes management confident enough to commit to a number this far out.

The demand is also broadening, which matters for durability. The fastest-growing customers were not the hyperscalers. The bucket NVIDIA calls AI Clouds, Industrial and Enterprise, the neoclouds, sovereigns and model labs, grew 138% to $40.3 billion, against 13% sequential growth for hyperscale at $48.7 billion. The Q3 guide of $108 billion sits about $4 billion above the street and assumes zero China compute; Hopper shipments to China ran under 1% of Data Center revenue, so nothing in the guide leans on a market a policy memo could erase.

The demand case did not weaken this quarter. It got stronger. What changed is the cost and the financing behind the ramp, and that is the rest of the piece.

What it costs: where the margin goes

Last week I called NVIDIA’s 75% gross margin a shortage price. So the number I cared about was the margin guide, and it steps down for more than one quarter. Management walked the whole path: 75% now, 74% next, a trough near 71 to 72% in Q4, then a partial recovery to roughly 72 to 73% through fiscal 2028. That is two-plus points below today. The NVIDIA CFO attributed the reset to memory costs rising faster than expected. Supply commitments jumped from $119 billion to $279 billion in a quarter, primarily related to memory procurement.

The reset is that NVIDIA is not guiding back to 75%. It is guiding to 72 to 73% as the new floor, absorbing part of the memory pass-through to keep the ramp moving. A dominant franchise is spending a point or two of margin for volume and share, which it can afford. But it lowers the profit the whole trade is discounting against. Watch that floor. A step from 75 to 72 is a normal ramp cost. A slide from 72 toward the high 60s means pricing power is actually eroding.

The cash line turned

Operating cash flow was $24.1 billion, down from $50.3 billion last quarter. Cash from operations fell by more than half even as revenue reached a record.

This is not a demand-collapse story. Operating income was $63.7 billion. It is a working-capital story, and the working capital is the tell. Receivables rose to $63.1 billion and days sales outstanding went from 45 to 60 in one quarter, because NVIDIA extended payment terms “on large, multi-quarter agreements with certain investment-grade customers.” Inventory climbed to $31.6 billion while the company staged for Rubin, and NVIDIA issued $25 billion of senior notes.

Cash, cash equivalents and marketable debt securities rose to $56.6 billion, even after returning $26 billion to shareholders, helped by the $25 billion note issuance. The near-term pressure is clear: a working-capital build and extended customer terms delay cash collection. That does not prove demand is weak. It does make the cash conversion of that demand worth watching.

NVIDIA is taking on customer-credit exposure

The receivables are one symptom. The commentary explains why:

“AI clouds and model makers are seeing extraordinary demand for AI infrastructure, yet many are growing faster than their balance sheets and long-term credit profiles can support. In response, we have entered into arrangements that help select customers secure the land, power and data center capacity needed to support their growth.”

The quarter shows three distinct mechanisms. First, NVIDIA is extending payment terms to selected investment-grade customers, which is trade credit, not a loan book. Second, it has $108.5 billion of guarantee exposure, including $105 billion of credit support for one Ohio campus under long-term leases. Third, it has large strategic-investment and supply-commitment exposures elsewhere in the AI buildout. These instruments have different economics and accounting. Together, they leave NVIDIA more exposed to customer credit and infrastructure execution than a conventional chip supplier.

That does not make NVIDIA a bank, and the arrangements can be rational from a position of strength. But it changes what you own. A chip company is a bet on a product. One that guarantees customers’ leases and lets selected buyers pay later is also partly a bet on customer credit. The receivable book is concentrated: five direct customers account for 70% of it. The question shifts from whether demand is real to whether the financing structure behind it remains sound.

What it means for the AI trade

The most useful read from this print is not about NVIDIA. Demand is running higher than NVIDIA can supply, and this quarter tells you where the overflow lands. If NVIDIA can only build to 70% while demand is greater, the constraint itself is the signal: the bottleneck names capture what NVIDIA cannot ship.

Memory is the clearest beneficiary. The reason NVIDIA’s margin steps down is that memory costs are rising faster than expected, and supply commitments jumped to $279 billion, most of it memory. That is a squeeze for NVIDIA and a tailwind for the HBM makers. SK Hynix, Micron and Samsung sell into a market that is effectively sold out, at rising prices, with NVIDIA pre-committing volume years out. The point or two of margin NVIDIA gives up is pricing power for them.

The neoclouds are the second read. The fastest-growing bucket this quarter was AI Clouds and model labs, up 138%, and NVIDIA is now helping finance their land, power and capacity. CoreWeave and Nebius are the listed versions of that buyer, and their growth is being pulled forward by the same credit support that adds risk to NVIDIA’s balance sheet. Bullish for them now, more entangled if any of them stumbles.

Networking and custom silicon are the third. A supply-limited 70% guide means the whole rack scales with it. Broadcom and Marvell on custom accelerators and interconnect, Arista on switching, all read through from a buildout NVIDIA just guided higher and longer out. Marvell reports this week, and this print raises the bar for it.

The caveat is that these read-throughs and the risks are the same force. Sold-out memory, financed neoclouds and a longer buildout are exactly what load margin, cash and credit exposure onto NVIDIA. The demand got better for everyone downstream, and the payback math got harder for the company at the center. That is why the trade is broadening, and why I am watching NVIDIA’s receivables line as closely as its revenue.

What I am watching across the trade

NVIDIA’s balance sheet is the cleanest gauge the trade has. Because it now finances and pre-commits to the buildout, stress there tends to show before it reaches a neocloud headline. Each line below is an NVIDIA tell, read as a signal for the whole complex.

• Margin. Does the 72 to 73% floor hold, or does memory drag it lower? That line separates a ramp dip from eroding pricing power. The same memory cost is a tailwind for SK Hynix, Micron and Samsung, read from the other end, so watch whether HBM stays sold out or supply catches up.
• Cash. DSO past 60 days and a second soft quarter of operating cash flow would turn a wiggle into a trend.
• Concentration. How much of receivables sits with the top few names, and any rise in guarantees above $108.5 billion. Strain at the financed buyers, CoreWeave, Nebius and the private labs, tests that exposure first. The one I watch hardest.
• Financing. Signs that extended customer terms persist, guarantees climb and NVIDIA issues more debt to carry the buildout.
• Conversion. That $279 billion of supply commitments turning into revenue on schedule, rather than sitting while growth flattens.

I am still long, and this print is a reason to stay long. Demand is broadening, management sees roughly 70% fiscal-2028 growth, and margins remain strong even at the trough. The market was right to rally on the top line. But the top line was never my worry. NVIDIA has added customer-credit and execution exposure to a trade that already carries plenty of leverage and concentration. I take the beat, stay long, and watch the receivables line as closely as the revenue line.

This is analysis, not investment advice. Figures come from NVIDIA’s Q2 FY2027 press release and CFO commentary, published August 26, 2026, for the quarter ended July 26, 2026. Commitments and guarantees are not cash already spent, and their ultimate exposure depends on future events. Recheck before relying on them.

Sources

• NVIDIA Q2 FY2027 press release, published August 26, 2026 (headline results, Q3 guidance and buyback): NVIDIA Newsroom
• NVIDIA Q2 FY2027 CFO commentary, filed August 26, 2026 (segment detail, cash flow, DSO, commitments, guarantees and financing arrangements): SEC exhibit
• NVIDIA Q2 FY2027 Form 10-Q, filed August 26, 2026 (customer concentration and longer payment-term disclosure): filing
• Fiscal-2028 outlook, supply constraint and margin path, reported August 26, 2026: Yahoo Finance
• Consensus estimates and after-hours context — published August 26, 2026: Associated Press, CoinDesk