A report has surfaced suggesting that Micron may be among the parties in NVIDIA's multi-year memory supply agreements for AI applications. Edgewater Research's "Compute Industry Update," dated August 20, states—based on anonymous channel information—that NVIDIA has likely signed multi-year DRAM and HBM agreements with both SK hynix and Micron.
What this report newly connects is the group of 16 undisclosed Strategic Customer Agreements (SCAs) that Micron has not named customers for, and the company's already-announced mass-production relationship with NVIDIA for HBM4 and SOCAMM2. SK Group and NVIDIA announced on July 25 a long-term partnership to jointly develop and optimize next-generation AI memory, including HBM, with NVIDIA securing stable supply—but Micron has not announced itself as a party to any multi-year contract.
However, neither NVIDIA nor Micron has publicly disclosed a multi-year contract between the two companies. Edgewater's original note was intended for clients, and the screenshots that have circulated publicly contain no contract dates, terms, or volumes. The same goes for pricing, prepayments, or product-specific allocations. At this stage, it's necessary to distinguish between the confirmed mass-production relationship and the still-unconfirmed, report-stage contractual relationship.
The confirmed SK hynix partnership versus the report-stage Micron contract
Regarding NVIDIA and SK hynix, a long-term AI memory partnership has already been officially announced. In the July 25 announcement, both companies stated they would jointly develop and optimize next-generation AI memory, including HBM, with NVIDIA securing stable supply. However, this is part of a broader cooperation framework that includes a letter of intent (LOI), and the supply volume, unit price, and contract duration for HBM have not been disclosed.
| Relationship | Disclosure status | Confirmed details | Unconfirmed details |
|---|---|---|---|
| NVIDIA and SK hynix | Officially announced | Long-term AI memory partnership; joint development and optimization including HBM; secured stable supply | Supply volume, price, contract duration |
| NVIDIA and Micron | Report stage (Edgewater) | Possible multi-year DRAM/HBM agreement | Whether a contract exists; duration, volume, price, product scope |
Edgewater's wording does not assert two separate, confirmed contracts. It uses the phrase "likely signed" and refers to a single "agreement" in the singular. It remains unclear whether this refers to the same document as SK hynix's July announcement, or an additional contract altogether.
Micron withheld customer names across 16 SCAs
During its fiscal Q3 2026 earnings call, Micron disclosed that it had signed 16 Strategic Customer Agreements (SCAs). These are typically five-year contracts running from 2026 through the end of 2030, though automotive-related agreements are generally three-year contracts. SCAs are take-or-pay agreements that obligate customers to purchase a set volume—DRAM contracts may include HBM as needed, and NAND is also supplied under multi-year terms.
According to the company, these 16 contracts account for roughly 20% of DRAM volume and about one-third of NAND volume over the contract period. The remaining contract revenue based on minimum pricing for 14 of these agreements totals approximately $100 billion, with deposits and related financial commitments reaching $22 billion. In exchange for securing volume, customers take on purchase obligations and financial conditions.
That said, what Micron has disclosed is only the composition of the 16 SCAs—four large customers, three mid-sized customers, and smaller automotive-sector customers—without naming any individual customer. Whether NVIDIA is among that group of contracts, and if so, whether it covers HBM, SOCAMM, standard DRAM, or NAND, remains unknown. The Edgewater report has drawn attention precisely because it links the previously anonymous SCA group to NVIDIA, but Micron's own earnings materials alone cannot confirm that connection.
From HBM to CPU-side memory: why sourcing is split across products
In March, Micron announced it had begun mass shipments of its HBM4 36GB 12-High memory, designed for NVIDIA's Vera Rubin platform, starting in the first calendar quarter of 2026. It offers pin speeds exceeding 11Gb/s and bandwidth exceeding 2.8TB/s. At the same time, the company is mass-producing 192GB SOCAMM2 modules, which it says will support up to 2TB per CPU and 1.2TB/s of bandwidth in both the Vera Rubin NVL72 system and standalone Vera CPUs.
HBM serves as high-bandwidth, high-capacity memory located near the GPU for computation. SOCAMM, LPDRAM, and DDR5, by contrast, serve as main memory on the CPU side. In June, TrendForce estimated that the three major memory makers combined could meet only about 60% of NVIDIA's LPDRAM demand, explaining that reductions in SOCAMM capacity reflect supply constraints rather than falling demand. This estimate pertains specifically to LPDRAM and is not a figure representing HBM supply volumes or demand forecasts.
Because memory in AI systems does not serve a single role, there is a rational basis for NVIDIA to source from multiple suppliers across different product lines. However, that rationale does not prove the specifics of any individual contract. Nor should a change in SOCAMM configuration be treated as evidence of declining product shipments or weakening HBM demand.
2028 marks the start of supply improvement, not supply-demand balance
In the same note, Edgewater revised its forecast for NVIDIA's 2027 HBM demand downward, from the 31–33 billion units (in 1Gb-equivalent terms) cited in its previous note dated August 13, to the high-20-billion range. This revision aligns with an expected reduction in memory configuration for Rubin Ultra, but it pertains to HBM bit-demand forecasts specifically—it is not a conclusion that NVIDIA's overall product shipments or standard DRAM demand will decline.
Micron expects DRAM and NAND supply tightness to persist beyond 2027, while explaining that industry-wide supply will gradually improve by 2028. The company has also explicitly stated that it cannot predict when supply will catch up with demand. It would be premature to read 2028 as the year the shortage resolves. Supply improvement and supply-demand balance are not the same point in time.
Micron's own capacity expansion timeline also involves lags. Its Idaho ID1 facility is expected to produce its first wafers by mid-2027, while ID2 is slated for late 2028. Meaningful shipments from the existing Taoyuan, Taiwan facility are also expected around mid-2027. A first wafer marks a production milestone—it does not mean a new facility can immediately meet required volumes.
Building new fabs takes time and requires talent, permits, and power infrastructure. Process miniaturization is also becoming more difficult, and according to Micron, the ratio of front-end wafer inputs needed to produce the same bit volume rises with each HBM generation transition, squeezing supply available for non-HBM products. Multi-year contracts may improve visibility into procurement volumes, but they do not guarantee NVIDIA's full required volume, nor do they immediately ease broader market tightness.
Three conditions to watch when verifying the contract
The first point to confirm is whether Micron acknowledges NVIDIA as one of the counterparties in its SCAs. Next is whether the agreement covers HBM alone, or extends to SOCAMM, standard DRAM, and NAND as well. Shortages in AI server components are not limited to a single part, and the scope of products involved changes the implications for the supply chain.
Finally, there is the matter of disclosed transaction terms. Beyond how supply volume and pricing are determined, contract duration and deposit amounts also need verification. The $100 billion and $22 billion figures Micron disclosed represent totals across the entire group of SCAs—not amounts specific to NVIDIA. Until these details are disclosed, this report serves as evidence suggesting NVIDIA may be solidifying a long-term commercial relationship with Micron—but it does not settle questions of supply priority or pricing outcomes.