Nvidia's Record Quarter Confirms the AI Boom Isn't Peaking. The Costs Just Moved to Memory.

Nvidia reported its fiscal second quarter on Aug. 26, and the scorecard held its end of the hype: revenue of $96.2 billion, up 106% from a year earlier; data-center revenue of $89 billion, up 117%; adjusted earnings of $2.22 a share against roughly $2.10 expected. It was the fourteenth straight quarter above the company's own midpoint — a solid beat, if no longer the 20%-style blowouts of two years ago.

Then the stock did something revealing. It dipped more than 1% as investors drilled into the gross-margin outlook, reversed to finish the extended session higher, and kept climbing. By Friday Nvidia traded near $228, within about 4% of its 52-week high, with a market value above $5.4 trillion. Price is sentiment, not proof — but the reversal was the market sorting two messages inside one report.

The first message was demand.

The company guided to about $108 billion for the current quarter — its first projection above $100 billion, against a consensus near $104 billion. More importantly, it handed out a rare multi-year number: revenue should grow roughly 70% in fiscal 2028, versus an analyst average near 44%. CFO Colette Kress called the outlook supply-constrained, noting that customers' own forecasts implied growth could double from a year earlier if the supply existed. That is the sentence that answers the market's central fear — that the AI buildout would decelerate. The demand-peak doubt that had pulled the stock roughly 12% below its May high lost its main witness this quarter.

The support ran down the whole chain. Vera Rubin, the next platform, is in full production with orders from every major customer, expected to be about a fifth of data-center revenue this quarter and projected by management as its fastest ramp ever. AWS committed to deploy an additional 2 million Nvidia GPUs through 2028. AI labs should account for about a quarter of next year's business, and the neo-clouds — CoreWeave, Nebius — plan to exit the year with more than 8 gigawatts of Nvidia capacity, up from 3 gigawatts. AMD and Intel rose about 1% on the read-through: a longer, deeper compute cycle, not a demand top. The one discount built into the outlook is China — management is assuming no data-center revenue there.

The second message was the gross-margin step-down, and it is really a memory story.

Gross margin printed 75% this quarter, is guided to 74% for the third quarter, and is expected to bottom around 71%-72% in the fourth before recovering to roughly 72%-73% in fiscal 2028. Nvidia's own word for the cause is "extreme" memory pricing, and the underlying numbers explain the tone. Memory is now roughly a quarter of the cost of a high-end AI server rack, up sharply from two years ago. Server DRAM contract prices roughly doubled in the first quarter of 2026. SK Hynix sold out all of its 2026 high-bandwidth-memory capacity before the year began.

This is bottleneck migration in real time. As GPU supply eased, the scarce, expensive layer moved to memory — which is exactly why Asian chip shares led the response to Nvidia's report, with the KOSPI up 1.8%, SK Hynix up 4.4% and Samsung up 2.8%. Nvidia is using its own pricing power to hand part of the bill back: days before the report, customers were told server prices would rise more than 15% in many cases for systems shipping early next year, covering the Vera Rubin and Grace Blackwell lines, with memory costs cited as the driver.

Judge the reset as the second question, not the first. Does a three-point gross-margin step-down matter while revenue compounds at 70%? Not much — gross profit grows straight through it. The real question is where the value went. Memory makers acquired pricing power that used to sit inside Nvidia's model, and that reshapes which names in the supply chain deserve a new valuation anchor. This is a transfer of pricing power, not a collapse in growth quality.

The third message was the balance sheet, and it is the part to keep in front of you.

In August, Nvidia launched financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR meant to mobilize over $500 billion of third-party capital for customers to buy its own systems, using the compute as collateral — then packaged a $105 billion credit-and-compute deal for an OpenAI data-center complex in Ohio that could reach roughly 10 gigawatts. The financing removes the customer capital constraint — the industry's one genuine brake on the buildout. Nvidia's credibility rests on real heft: roughly $66 billion of net cash, trailing free cash flow near $127 billion, and a fresh $26 billion returned to shareholders last quarter while inventory climbed to about $32 billion to pre-build for Vera Rubin.

But the structure also moves risk onto Nvidia's books. Its revenue now depends partly on customers taking on debt its partners arranged, and payments are already stretching — days sales outstanding rose to 60 days. Part of the Nvidia story has quietly become a credit-cycle story, and that is the new risk boundary.

What to watch now.

One strong quarter does not change a regime; the confirmation queue is visible. Watch four things. First, the fourth-quarter gross margin actually lands near 71%-72% and then recovers toward 72%-73% in fiscal 2028 — a floor is only a floor if it holds. Second, memory prices stop inflating; the Korean memory makers and Micron are now the cheapest read on Nvidia's own cost line. Third, the $500 billion financing program converts from memorandums of understanding into deals that draw real capital, and the neo-cloud capacity gets used rather than just built. Fourth, hyperscalers absorb the price increases instead of accelerating their custom-silicon programs — three hyperscalers accounted for more than half of revenue last quarter, so pushback would surface fast.

The signal that breaks the narrative is not one margin print. It is hyperscaler capex decelerating — the four largest spent about $166 billion combined in the June quarter, up 87% from a year earlier — or memory costs forcing that 70% multi-year guide down. Until then, the statement of the quarter stands: the AI buildout is not peaking. It is getting more expensive, and the industry will now decide, quarter by quarter, who pays.