← Back to blog
GeneralAbout 6 min read

The 74 Percent Gross Margin Has a Shelf Life

Published Nov 17, 2026
The 74 Percent Gross Margin Has a Shelf Life

Seventy-four percent. That is Nvidia's gross margin guidance for the quarter it reported on November 17, down from 75 percent in the prior quarter, with a further fall to 71 or 72 percent guided for the current one. Management has said margins recover to 72–73 percent in fiscal 2028, once price increases take effect.

Three numbers, one direction, and a story that is much more interesting than a two-point decline usually is. Because at Nvidia's scale, a point of gross margin is not a rounding error. On roughly $303 billion of trailing twelve-month revenue, one point is about $3 billion a year. Three points is $9 billion — more than most semiconductor companies earn in total, and more than Nvidia spent on research and development not that long ago.

So why is it falling, and why does it matter more than the revenue beat?

Force one: memory, and the HBM4 squeeze

Nvidia does not make memory. It buys HBM from SK Hynix, Samsung and Micron, and it buys conventional DRAM for the rest of the system. Both have been climbing, and the climb has been steeper than management expected when it set its fiscal 2027 framework.

The memory squeeze is not only a cost problem. It is a supply problem, and supply problems hit revenue before they hit margin.

KeyBanc cut its 2026 Rubin GPU shipment estimate from 1.7–1.8 million units to 1.5 million, citing HBM4 availability. Management expected Vera Rubin to be about twenty percent of data center revenue in the quarter just reported. If HBM4 does not arrive in the volumes the ramp assumes, the shortfall gets filled with GB300 — older architecture, different margin, and a different mix than the model assumes.

This is the part analysts will push on during the call. Not "is demand there" — everyone agrees demand is there — but "can you build the thing you sold."

Macro view of stacked HBM memory dies on a mirrored surface, gold contacts gleaming

Force two: foundry pricing, arriving next year

The second force is newer and less discussed. Chip designers say they have received wafer pricing notices from TSMC and other foundries for next year, in the range of ten to twenty percent. If those increases are not passed through, the impact on gross margin is four to five points.

Four to five points, on a company guiding 74 percent, would take margin into the high sixties — a level Nvidia has not seen since before the AI boom. That is why the pricing question matters more than the headline number: Nvidia has enormous pricing power, but exercising it takes time, and the contracts for next year's racks are being signed now.

There is a lag built into this. Nvidia raises prices for new orders, but a large share of revenue in any given quarter comes from orders booked months earlier at earlier prices. So the cost arrives first and the price arrives later, which is exactly the shape management described: trough in fiscal Q4, recovery through fiscal 2028.

Force three: the mix is getting heavier

Here is the structural piece, and it is the one that does not get enough attention.

Nvidia no longer sells GPUs. It sells racks — GB300, and soon Vera Rubin — that bundle chips, memory, networking, power delivery, cooling and firmware into a system that a hyperscaler can drop into a hall. That is a better business in absolute dollars, because the content per rack is enormous. It is a worse business in percentage margin, because a larger share of the bill of materials is pass-through content bought from suppliers at supplier margins.

If you sell a $30,000 board with $10,000 of your own value added, your gross margin looks like a semiconductor company's. If you sell a $3 million rack with $900,000 of memory and networking inside it, your gross margin looks more like a systems company's — even though your absolute gross profit per unit is much higher.

This is not deterioration. It is arithmetic. But it means that comparing Nvidia's 74 percent today to its 78 percent in 2023 is comparing two different businesses, and the market has been slow to internalize that.

The $145 billion question

All of this sits on top of a commitment that would be reckless for any other company and is merely aggressive for this one: $145 billion in total supply, including inventory purchase commitments and prepayments, as disclosed on the August call.

That number is a forecast in the form of a contract. Nvidia has ordered the memory, reserved the wafer starts and prepaid the assembly. If customers take delivery on schedule, the commitment converts into revenue at the guided margin. If a data center hall slips by two quarters because the substation is late, the inventory does not vanish — it sits on the balance sheet, and the next platform ramps on top of it.

This is why the power constraint story matters to a chip company. SoftBank, a partner on one of Nvidia's buildouts, took a $3.7 billion impairment on power equipment tied to data center construction. You can have the GPUs and the customers and the capital and still be waiting on a transformer. Every quarter of delay costs someone money, and the party holding $145 billion of prepaid inventory is not the one with the most flexibility.

What would make this a problem rather than a cycle

Two things.

The first is if the margin recovery depends on price increases that customers refuse. Hyperscalers are the most sophisticated buyers on earth, and they now have alternatives — AMD, their own ASICs, and for some workloads, older Nvidia generations that are still perfectly good. A supplier that needs to raise prices to protect margin in a market with three viable options has less power than its share suggests. Watch whether the fiscal 2028 recovery language survives the next two calls unchanged.

The second is if memory stays tight for longer than a year. HBM4 is a hard manufacturing problem, and the fabs that make it are also being asked to supply every other AI accelerator on the market. If supply loosens in 2027, the margin story is a one-year dip. If it stays tight into 2028, Nvidia is running a systems business at semiconductor multiples, and the 11x forward earnings starts to look less like a bargain and more like a correct assessment.

The boring conclusion

Nothing in this margin trajectory is alarming on its own. Memory costs go up and down. Foundries raise prices when they have leverage. Systems businesses have lower percentage margins than chip businesses, always.

What makes it worth watching now is that all three forces arrive in the same four quarters as a leadership transition. A founder could stand on a stage, explain that the trough is temporary, and be believed. An interim CFO-turned-CEO has to do it with a spreadsheet, in a room full of people who just downgraded the stock ten times.

Seventy-four percent is not a crisis. But it is the number that will decide whether the market believes the next three years of this company's story, and it is being reported by someone who has never had to sell that story alone.

Related articles