Nvidia Reportedly Locks In Multi-Year DRAM and HBM Supply

Nvidia is reportedly moving to reserve memory years ahead of delivery, securing conventional DRAM and high-bandwidth memory (HBM) from SK hynix and Micron for AI accelerators. The report comes as memory suppliers shift toward longer commitments and warn that AI-driven demand is running ahead of available capacity. 4451

The important distinction is between what is confirmed and what remains reported. Nvidia and SK hynix publicly announced a multiyear technology partnership covering next-generation memory development and manufacturing. 22 The separate claim that Nvidia signed comparable multiyear supply agreements with both SK hynix and Micron comes from industry reporting; the public material available here does not establish the deals’ exact volumes, prices, or terms. 44

What Nvidia reportedly secured

The reported agreements cover two related but different memory needs:

• HBM: High-bandwidth memory sits close to an AI accelerator and supplies data at the speed required for demanding training and inference workloads.
• Conventional DRAM: Standard system memory is also required across the servers and platforms surrounding an accelerator.

Securing both types of memory would give Nvidia more visibility into a key input for accelerator production. It would not eliminate supply-chain risk: a contract cannot instantly create qualified wafers, packaging capacity, or production yields. But it can give Nvidia a stronger position when suppliers are allocating capacity years ahead.

Nvidia has also confirmed that Samsung, SK hynix, and Micron have passed qualification for HBM4 intended for the Vera Rubin platform, although Nvidia has not disclosed official volume allocations. 24 That makes the reported agreements strategically significant without proving that any one supplier will provide a specific share of future demand.

How the agreements connect to the SK hynix partnership

The June 7 Nvidia–SK hynix announcement was broader than a conventional purchasing contract. The companies described a multiyear partnership to align next-generation memory development with Nvidia’s AI-infrastructure roadmap and expand supply for the AI-factory buildout. 22

The partnership includes memory co-development for future Nvidia platforms, including Vera Rubin systems, and spans both design and manufacturing. 1732 In practical terms, that gives SK hynix a role earlier in the product cycle: the supplier can work with Nvidia on memory technology and manufacturing requirements before systems move into large-scale deployment.

That distinction matters because HBM is not simply interchangeable inventory. It must be designed, stacked, packaged, tested, and qualified for a particular accelerator platform. A co-development relationship can therefore reduce uncertainty around compatibility and timing, while a supply commitment can improve visibility into future capacity.

Why memory contracts are getting longer

Memory manufacturers are facing an unusual combination of high AI demand and long lead times for new capacity. Building fabs, installing equipment, improving yields, and qualifying advanced HBM products all require substantial investment before additional usable supply reaches customers.

That environment creates a natural incentive for three- to five-year agreements rather than one-year purchasing cycles:

• Buyers receive greater supply visibility for products that are difficult to replace quickly.
• Suppliers gain demand visibility that can support investment in wafer capacity, stacking, and packaging.
• Both sides can plan product road maps around known technology and qualification requirements.

Reporting on the market says 2027 memory pricing was not yet settled and that suppliers and customers were favoring longer agreements with volume and pricing frameworks instead of relying only on annual negotiations. 44 The implication is not that another specific Nvidia contract is confirmed; it is that uncertainty itself is encouraging more capacity reservation across the industry.

What the shortage outlook says about AI demand

Micron CEO Sanjay Mehrotra has said customers recognize that shortages in memory and storage will take considerable time to improve, even as industry supply is expected to improve gradually in 2028. 51 That outlook is consistent with a market in which new capacity arrives incrementally rather than resolving the imbalance immediately.

Some industry reports describe HBM capacity as sold out through 2026 and warn that tightness could continue into 2027 or later. 3639 The precise duration remains uncertain, so “through 2028” should be treated as a forecast rather than a settled deadline. The stronger conclusion is that AI memory constraints are expected to persist well beyond a single product cycle.

This changes the role of memory in the AI supply chain. It is no longer merely a peripheral component purchased after the accelerator design is complete. Access to qualified HBM and DRAM can affect how many accelerators a company can ship, when new platforms can be deployed, and how quickly data-center operators can expand.

Micron’s own long-term customer commitments

Micron’s fiscal third-quarter 2026 results illustrate the broader shift toward contracted memory capacity. The company reported revenue of $41.46 billion and guided fiscal fourth-quarter revenue to approximately $50 billion, plus or minus $1 billion. 5051

Micron also said it had completed 16 Strategic Customer Agreements. The typical agreements run for five years, from calendar 2026 through the end of calendar 2030; automotive agreements generally run for three years. Together, the 16 agreements represent roughly 20% of Micron’s DRAM volume and one-third of its NAND volume over the period. 63

Other reporting puts the minimum contracted revenue associated with the agreements at approximately $100 billion, while also citing more than $22 billion in customer financial commitments. 5255 These figures describe Micron’s customer-agreement program generally, not a confirmed Nvidia-specific contract. The available reporting does not reliably establish the number or value of any additional completed agreements beyond the 16 disclosed by Micron.

Why Micron’s U.S. expansion will take years to help

Micron’s initial U.S. expansion plan called for approximately $200 billion across semiconductor manufacturing and research and development: about $150 billion for manufacturing and $50 billion for R&D in Idaho, New York, and Virginia. 8

The schedule shows why that investment cannot quickly remove the shortage. Micron expects initial wafer output from its first Idaho fab in mid-2027, while production is expected to ramp primarily in 2028. 25 A fab’s first wafers are not the same as full, qualified production. Construction, equipment installation, yield learning, advanced packaging, and customer qualification all stand between groundbreaking and dependable high-volume supply.

That timing gap is especially important for HBM. Even if a new facility produces DRAM wafers, the memory must still move through the specialized stacking, packaging, testing, and qualification processes required for AI systems. The conclusion that new U.S. capacity will take years to materially ease the shortage is therefore an inference from the announced schedule and the complexity of the supply chain, not a claim that the investment will fail to add capacity.

Micron has since raised its U.S. investment target to more than $250 billion through 2035, reinforcing the scale and duration of the capacity buildout. 7

The strategic takeaway

The Nvidia reports point to a broader change in semiconductor purchasing: AI companies are increasingly reserving memory capacity and aligning with suppliers before products reach mass deployment.

SK hynix’s confirmed partnership shows the technology-development side of that shift. Micron’s 16 strategic customer agreements show how suppliers are seeking multiyear demand commitments. Nvidia’s reported DRAM and HBM deals, if the reporting is accurate, show how an accelerator designer is trying to protect its production roadmap against a shortage that may persist into the next generation of AI infrastructure.

The exact Nvidia–Micron terms remain unverified. The strategic direction is clearer: in the AI era, qualified memory is becoming infrastructure—and access to it may be as important to shipping compute as the accelerator itself.