Nvidia's $96 Billion Quarter and the $279 Billion Problem

Nvidia reported $96 billion in revenue last quarter. That number alone would qualify as the largest corporate quarter in history for most companies. It also tells you almost nothing about the problem Nvidia's management is trying to solve.

The number that matters is $279 billion in future purchase commitments, up from $119 billion just three months earlier. The jump is not a demand signal. It is a supply security move. Nvidia is pre-paying for memory at a scale that will reshape who captures the most value in the AI hardware stack.

Nvidia's second quarter of fiscal 2027 closed at $96.2 billion in revenue, more than doubling from a year ago. Data Center—the segment feeding hyperscaler training runs and enterprise AI deployments—accounted for $89 billion of that, up 117%. Gross margin sat at 75%. The stock jumped in after-hours trading.

Then management laid out what's coming next, and the picture changed.

CFO Colette Kress guided for gross margins to fall to 74% in the current quarter, then drop further to 71%–72% by fiscal year-end. She called the squeeze a "timing gap"—Nvidia is absorbing higher memory costs through January before passing them to customers through negotiated price increases. Margins should settle in the 72%–73% range in fiscal 2028 once those increases take effect.

At Nvidia's revenue scale, one percentage point of gross margin equals roughly $1 billion per quarter. The trough could cost the company $5–6 billion in quarterly gross profit versus the 75% rate. That is not a rounding error. It is the price of keeping the supply chain intact.

The mechanism behind the pressure is straightforward. High-bandwidth memory—the only type fast enough to feed data to AI accelerators—now accounts for an estimated 30% to 40% of the cost to build an AI accelerator, up from under 20% two generations ago. Three companies make it: SK Hynix, Samsung, and Micron. Demand from Nvidia alone is pushing against the limits of what those three can produce.

Management described "extreme pricing conditions in memory" on the earnings call. Nvidia has signed a multi-year memory partnership with SK Hynix and doubled its supply commitments to lock in capacity before prices climb further. The company is choosing supply security over margin preservation because the alternative—running out of memory and missing customer deliveries—would be far more costly at this growth rate.

The commitments break down roughly like this: about $92 billion for the rest of the current fiscal year, then $87 billion and $88 billion in the next two fiscal years. $267 billion of the total is scheduled within three years. This is not a speculative bet on distant demand. It is a near-term procurement lock designed to feed the Vera Rubin product launch.

There is another layer to the supply constraint that few headlines caught. Nvidia's flagship Rubin Ultra chip—initially previewed with 1 terabyte of HBM memory—will ship with just 192 gigabytes, roughly one-fifth of the original plan. The downgrade came from cutting the memory stack height from 16-high to 8-high and reducing the number of memory cubes. Even the standard Rubin configuration at 288 gigabytes will have more memory than Rubin Ultra.

Nvidia says these are "optimized SKU variations" rather than forced downgrades. The full-spec versions remain in production for hyperscale customers that can absorb the cost. But the existence of lower-memory variants at the flagship level signals something: memory scarcity is real enough that even Nvidia cannot get the chips it originally designed in full volume.

This mirrors a pattern the semiconductor industry knows well. When AMD designed its MI450 for Meta, it offered fewer compute dies and half the memory of the full MI455X variant—because supply constraints forced the customer to choose between shipping now with less capability or waiting for a chip that might not arrive. Nvidia is now running the same playbook on its own product.

Here is what the economics look like from both sides of this trade.

Nvidia's revenue is still exploding. ACIE—AI Compute, Infrastructure, and Edge, which serves enterprise and mid-market customers—grew 138% year-over-year and 25% sequentially to $40 billion. Hyperscale revenue hit $49 billion, up 13% sequentially. Sovereign AI—governments building their own AI infrastructure—tripled year-over-year and grew 35% in the quarter. The revenue opportunity per gigawatt is rising from roughly $18 billion in the Hopper era to an estimated $40 billion with Vera Rubin.

But the value chain is shifting upstream. HBM now costs more, and the three memory makers are capturing a bigger slice of every AI accelerator sale. Micron reported an 84.9% non-GAAP gross margin in its latest quarter—higher than Nvidia's 75%. That has never happened before. The supplier is now more profitable on its core product than the company selling the finished accelerator.

Nvidia's response—absorb the cost, lock in supply, raise prices later—is rational but it carries execution risk. If memory prices keep climbing past what Nvidia budgeted for, the 71%–72% trough could go deeper. If the price increases Nvidia has negotiated with customers don't stick, the fiscal 2028 recovery to 72%–73% doesn't materialize. And if a competitor with a different memory architecture sidesteps the bottleneck entirely, the $279 billion commitment becomes leverage rather than advantage.

The stock trades at a market cap of $5.2 trillion, roughly 27 times trailing earnings and about 58 times forward earnings on consensus expectations. The PEG ratio sits near 0.22, which looks cheap by any conventional growth-stock metric. But those multiples assume the revenue trajectory continues uninterrupted—and that the margin trough is temporary.

Management gave a useful framing for demand. Customer forecasts suggest demand is roughly doubling, but Nvidia guided for about 70% revenue growth next year. The gap between what customers want and what Nvidia delivers is not a demand problem—it's a supply one. Foundry capacity, data-center power, and memory availability are all bottlenecks. Nvidia grows fast, but not as fast as the market wants it to.

The question for an investor is not whether Nvidia wins. The question is whether the near-term return curve still justifies the allocation. The company is spending $279 billion in commitments and taking a margin hit to protect a position that is already dominant. That discipline makes sense operationally. It does not automatically translate into stock performance over the next six to twelve months.

While demand is clearly not the problem, the margin trough, the memory dependency, and the sheer scale of capital already deployed into supply chains mean the stock's return path depends on execution details most headlines won't track. Whether Q4 margins hold at 71%–72% tells you if this is a timing gap or a structural shift. Whether Rubin Ultra reaches customers at scale with the downgraded memory tells you if the product cycle can sustain the growth rate. Whether memory prices stabilize before Nvidia's price increases kick in tells you if the fiscal 2028 recovery is real.

The AI infrastructure build-out is still accelerating. Nvidia remains the company at the center of it. But the margins that made the stock one of the best performers in market history are no longer guaranteed—they are being negotiated, quarter by quarter, with three memory suppliers who now know they hold the bottleneck.