Nvidia Ramps Up Memory Procurement: Can Micron and Samsung Become the Biggest Beneficiaries of AI Computing Expansion?

TradFi
Обновлено: 18-11-2026 08:22

On November 17, 2026, at the close of US markets, according to Gate stock market data, Nvidia (NASDAQ: NVDA) was up at $155.46 and briefly rose as high as $162.95 after the close. The direct trigger for the post-market jump was its third-quarter earnings report released that day: revenue of $115.8 billion, $107.6 billion from the data center business, and an increase in the expected revenue growth rate for fiscal 2028 from 70% to 75%. At the first earnings call since Jensen Huang stepped down, interim CEO and CFO Colette Kress’s key judgment came down to just four words: demand far exceeds supply.

Source: Gate stock market data

More than the revenue numbers themselves, the thing worth dissecting is a commitment beyond the balance sheet. A filing Nvidia submitted to the US SEC shows that its supply commitment rose from $279.0 billion in the prior quarter to $390.0 billion, with a quarter-on-quarter increase of $111.0 billion. The company also explicitly said the incremental amount is "mainly related to multi-year memory supply procurement." This figure extends the AI compute narrative beyond GPU shipment volumes and into HBM, DRAM, substrate/boards, and advanced packaging. It also puts Micron (NASDAQ: MU) and Samsung Electronics at the center of discussions about who benefits across the supply chain.

What key signals did Nvidia’s third-quarter earnings release?

For $115.8 billion in revenue, the year-over-year growth rate was 103%, and the quarter-over-quarter growth was about 20%. That beat the market’s prior expectation of roughly $109.3 billion. Data center revenue came in at $107.6 billion, up 21% quarter over quarter and up 110% year over year, keeping its share of total revenue at over 90%. Adjusted gross margin was 74.9%, and the revenue guidance for the fourth quarter was $126.1 billion to $133.9 billion.

What truly changed market pricing was not the realized Q3 numbers, but the growth guidance for fiscal 2028. Nvidia raised what had already surprised the market from 70% to 75%, and it said clearly that "both demand growth and supply increases are supporting the upward revision." The fact that increased supply can serve as a reason for the upward revision is itself a signal. It suggests the bottlenecks that had been restraining shipment pacing over the past several quarters are gradually being opened up. Kress also disclosed that the ramp speed of the next-generation Vera Rubin chips is "the fastest in Nvidia’s history," already creating about $20 billion in revenue.

Source: Nvidia earnings report

The jump in the supply commitment and the increase in growth guidance occurred in sync. This indicates Nvidia’s upstream locking is not just passive inventory replenishment. Instead, after demand visibility improved, the company proactively used multi-year contracts to secure production capacity certainty. Memory is the largest-volume piece in the equation, with the longest expansion cycle and the lowest supply elasticity.

Why does a $390.0 billion supply commitment point to growing memory demand?

The supply commitment rose from $279.0 billion to $390.0 billion, an increase of $111.0 billion in a single quarter. In the filing, Nvidia attributed the incremental amount to "multi-year memory supply procurement." A chip design company with annual revenue in the hundreds of billions of dollars, one-quarter’s incremental memory-related commitments exceeding $100 billion does not mean ordinary stocking. It means pre-booking HBM and DRAM capacity for the next several years—effectively "buying out" that capacity early.

HBM supply constraints have structural causes. HBM is manufactured by stacking multi-layer DRAM wafers and using through-silicon via (TSV) processes. Compared with conventional DRAM, it uses far more wafer area and packaging resources per unit of capacity, and its yield ramp takes longer. When HBM squeezes advanced DRAM capacity, the supply of traditional server DRAM tightens at the same time. In Q4 2026, the price of general DRAM contracts is expected to rise 10% to 15% quarter over quarter. PC DRAM contract prices are expected to increase by 13% to 18%. In some server memory spot markets, prices are already 30% to 50% higher than contract prices.

Nvidia’s procurement commitments, in essence, lock in both price and volume certainty in a market where the supply curve rises steeply. But the party it locks in can only come from just three HBM mass-production suppliers worldwide.

Nvidia expands memory purchases—can Micron and Samsung benefit?

Nvidia’s forecast for global HBM4 market share in 2026 roughly falls within these ranges: SK hynix at about 50% to 54%, Samsung Electronics at about 28% to 32%, and Micron at about 18% to 22%. Micron has already sold out its 2026 HBM supply. Its HBM4 36GB 12-layer stacked product entered mass production in the first quarter. Pin speeds exceed 11 Gb/s, bandwidth surpasses 2.8 TB/s, and it is about 2.3x better than HBM3E.

From the perspective of "benefit certainty," Micron’s upside comes from a two-layer stack. First, there’s revenue-side locking: multi-year contracts turn HBM capacity into predictable cash flows and reduce the tail risk of traditional storage cycles—things like sudden price crashes leading to inventory impairment. Second, the product mix shifts upward: Micron’s capital expenditure guidance for fiscal 2026 was raised from $18 billion at the start of the year to more than $27 billion. The incremental spend is concentrated on advanced DRAM and HBM. In the second fiscal quarter, Micron’s gross margin reached 74.9%. In the third quarter, guidance is as high as 81%.

Samsung’s logic is closer to "share repair + vertical integration." Samsung is the only HBM supplier that completes DRAM manufacturing, logic chip manufacturing, and advanced packaging across the full process within its own fabs. Among the three, Samsung’s HBM4 validation progress is the fastest. Its challenge, however, is that expanding HBM is crowding out general DRAM capacity, and general DRAM prices are also moving higher. Samsung needs to decide how to allocate capacity and profit between HBM share and general products. Nvidia’s $39.0 billion commitment is a tailwind for both companies on the order side. But Micron has more direct "pure HBM elasticity," while Samsung has a broader revenue base from "full-stack capabilities + general DRAM price increases."

Why carrier boards, advanced packaging, and CPO are becoming more important

If HBM is the bottleneck at the memory layer, then carrier boards and advanced packaging are the bottlenecks at the physical integration layer. Around the same time, Nvidia announced it would invest more than $100 million in Kinsus, a packaging substrate manufacturer in Taiwan, for equipment procurement. The goal is to expand Kinsus’s advanced ABF carrier board capacity dedicated to Nvidia by more than 3.5x. New capacity is expected to come online in the second half of 2027, with full ramp targeted for the third quarter of 2028. Nvidia also amended its existing substrate supply agreement with Samsung Electro-Mechanics to improve operating capability at existing plants.

On advanced packaging, CoWoS remains the core technology for integrating compute dies and HBM. TSMC is the main supplier, but downstream packaging capacity remains tight. Nvidia is investing tens of billions of dollars to help Amkor Group/UTAC (ASE Investment Holdings)旗下矽品 expand CoWoS capacity. Packaging expansion cycles are measured in years. That means even if HBM supply is secured, packaging capacity could still become a secondary constraint on shipment pacing.

CPO’s significance is that it brings "interconnect" into the same expansion framework. Nvidia’s third-quarter networking revenue exceeded $21 billion. CPO switches have entered mass production, and Lambda has announced plans to投入生产部署. When the number of GPUs per rack increases and cluster scale grows, constraints on power consumption and signal integrity for electrical interconnects will quickly amplify. Optical interconnects will shift from "optional" to "required." Mass production of CPO means Nvidia is extending supply chain bottlenecks from the compute chips themselves down layer by layer into memory, carrier boards, packaging, and optical modules—and locking them in step by step.

How long can AI compute demand keep going?

The most direct evidence on the demand side comes from backlog orders. Nvidia disclosed that cloud computing industry backlog orders rose from $2.0 trillion to $2.5 trillion within three months. Renewal pricing on some multi-year leases for Hopper GPUs can be as high as twice the original price. Since B300 was launched, its lease rates have risen by more than 60%. GPU leasing prices are a real-time pricing signal for compute supply and demand. If renewal prices for existing cards can double, that indicates the shortage is not limited to new-card delivery—it runs across the entire pool of in-use assets.

Capital expenditure data provides another form of cross-validation. In the first three quarters of 2026, the combined capital expenditures of five major cloud providers—Microsoft, Amazon, Google, Meta, and Oracle—were about $560 billion, up 98% year over year. Nvidia’s management expects the five cloud giants’ total spending for all of 2026 to approach $800 billion, and to rise to as high as $1.3 trillion in 2027. This scale of spending inertia forms the demand base for Nvidia’s upward revision to its fiscal 2028 growth guidance.

That said, risks to long-term demand can also be validated. The three storage giants’ combined capital expenditures in 2026 are close to $120 billion. The timing when new capacity is released is concentrated in 2027 to 2028. The share of HBM in DRAM wafer capacity has risen from less than 5% in 2022 to about 20% by the end of 2025, and it is expected to reach 35% in 2027. On the supply side, expansion is moving forward at a pace unseen in history. If cloud providers’ capital expenditure growth slows after 2027, the prices and gross margins supported today by supply constraints will first come under pressure in general DRAM, and then transmit to HBM contract renewal pricing.

Conclusion

The core message of Nvidia’s $39.0 billion supply commitment is not simply that "Nvidia bought lots of memory." It is that constraints in AI infrastructure are spreading beyond compute capability itself—from GPU compute power—to memory bandwidth, carrier board layers, packaging capacity, and optical interconnect bandwidth. Micron and Samsung sit upstream in this constraint chain. Multi-year contracts give them rare visibility into revenue. But capital expenditures on the supply side are also accumulating at record speed. For those tracking linkages between the AI supply chain and the crypto market, the variable truly worth monitoring is not the size of the outperformance in a single quarter’s revenue. Instead, it is how quickly the time gap between the HBM capacity release schedule and the growth rate of cloud providers’ capital expenditures narrows.

FAQ

What exactly does Nvidia mean by "supply commitment"?

A supply commitment refers to the aggregated disclosures in Nvidia’s earnings materials for already-contracted procurement obligations over the coming years. Within this quarter’s $39.0 billion, the incremental portion mainly comes from multi-year memory procurement contracts rather than one-time spot orders. It reflects early locking in of HBM and DRAM capacity.

Which has more upside elasticity: Micron or Samsung?

Micron has a higher share of HBM revenue and its capacity is already sold out, so its profit elasticity tied to upward moves in HBM pricing and volumes is more direct. Samsung has more room to repair its HBM share and it also has a hedge from general DRAM price increases and vertical integration. The benefit directions align, but the driver mix differs.

What does mass production of CPO switches mean for the supply chain?

CPO packages an optical engine with a switching chip together. This reduces power consumption and improves signal integrity. Mass production means bottlenecks in AI clusters expand from "whether there are enough chips" to "whether there is enough interconnect bandwidth." Optical modules, optical engines, and advanced packaging all gain new demand, but deployment timing and yield uncertainties remain.

What is the most verifiable risk for AI compute demand?

Capital expenditure growth on the supply side. The three storage manufacturers’ combined capital expenditures in 2026 are close to $120 billion, with new capacity concentrated to be released in 2027 to 2028. If cloud providers’ capital expenditure growth had slowed earlier than expected, storage prices and HBM contract renewal pricing would face pressure first. That is the clearest negative time-sensitive variable in the current chain of reasoning.

The content herein does not constitute any offer, solicitation, or recommendation. You should always seek independent professional advice before making any investment decisions. Please note that Gate may restrict or prohibit the use of all or a portion of the Services from Restricted Locations. For more information, please read the User Agreement

Поделиться

sign up guide logosign up guide logo
sign up guide content imgsign up guide content img
Sign Up
Log In