Nvidia's Growth Cap Is Priced in Memory

Nvidia closed Wednesday at $209.66, down 1.6 percent into the print. After the close it reported $96.2 billion of revenue — about $4 billion above the LSEG consensus near $92.2 billion — with data center at $89.0 billion and gross margin at 75.0 percent. Guidance is $108.0 billion, plus or minus 2 percent, with no China data-center compute. Shares then rose as much as 8 percent Thursday once Colette Kress said fiscal 2028 revenue should grow about 70 percent. Street models had been nearer 44 percent.

The hangover story had treated that print as a repricing of AI monetization, with Nvidia still named the sector leader even as its multiple sat near 27 times earnings, a 2019 low. Five days earlier, Samsung Electronics approved a 2026 shareholder return of 90 to 110 trillion won — roughly $65 billion to $80 billion — half of three-year free cash flow, the largest ever by a Korean company. One firm compounding a platform, the other distributing the cycle: the pairing wrote itself. Korean memory stocks then rose with Nvidia on Thursday.

The residual is a cost, not a contest

If compute gaps and energy constraints are moving the sector, the 10-Q has to show a shortfall or a cost increase that reaches Nvidia’s earnings path. Wall Street has shrugged at a clean beat before. This time the shrug preceded the print; the print reversed it.

Kress was specific. Nvidia is “experiencing extreme pricing conditions in memory,” with increases that “exceeded our prior expectations and are headed even higher into next year.” Third-quarter gross margin is guided to 74.0 percent, plus or minus 50 basis points, then a fourth-quarter bottom at 71 to 72 percent before 72 to 73 percent in fiscal 2028 — a 300-to-400 basis-point reset from the quarter just reported. Supply commitments more than doubled, from $119 billion to $279 billion, “primarily related to the procurement of memory.” Of that, $92 billion comes due in the rest of fiscal 2027.

The growth number is a cap, not a demand fade. Customer forecasts “point to our growth doubling next year”; Nvidia’s outlook is 70 percent because it is supply-constrained. Jensen Huang put it in supply-chain language: the firm has supply for about 70 percent growth; demand is “much higher.” That is a 30-point fulfillment shortfall on fiscal 2028, in the guide. TSMC has already said AI-chip supply stays tight for years. Memory is the invoice that reached the income statement.

The 10-Q prices a 300-basis-point memory reset and a 30-point supply gap. It does not price a Samsung dividend.

Power is in the footnotes, not yet in cost of goods. Nvidia disclosed SB Energy guarantees at the PORTS-Pike campus in Ohio, capped at $105 billion, plus $3.5 billion of other land, power, and shell guarantees. Morgan Stanley has a 38-gigawatt U.S. data-center power gap through 2028 against 68 gigawatts of needed additions. Those delay when racks go live. They did not move this week’s margin guide. Whether the apps ever pay for the hall is a later test.

Samsung is collecting the same bottleneck

Samsung’s return is half the free cash flow from the memory supercycle inflating Nvidia’s bill. It is not a substitute for Vera Rubin. The company is also spending a record 110 trillion won — about $73 billion — on facilities and research this year, began HBM4 mass production in February, and has said 2026 HBM4 capacity is sold out. SK hynix remains the purer Nvidia HBM supplier; Samsung sells stacked memory and the conventional DRAM that tightens when wafers move. Kress named Samsung, SK hynix, and Micron as the vendors she is pressing for capacity. Hyperscalers hedging into custom silicon are buying a different accelerator, not a different memory physics.

What still gets mis-specified is a leadership contest. Nvidia can print $96 billion and sit near a 2019 multiple because the path now includes known cost inflation and a known fulfillment cap. Samsung can pay a record dividend because that inflation is its revenue. Energy can delay a campus without a 400-basis-point miss this year. China’s cheaper-token path and Groq-style inference capacity change how much silicon a workload needs; they do not cancel the purchase order already booked.

The testable claim is the fourth-quarter print. If gross margin bottoms at 71 to 72 percent and fiscal 2028 growth holds near 70 percent while memory commitments stay elevated, the hangover was a cost-of-goods story. If margin holds and growth misses because customers delay halls for interconnection, the megawatt gap arrives in the numbers. Treasury yields can still move any morning’s tape. They cannot explain a $160 billion increase in memory purchase obligations.