What multi year DRAM and high bandwidth memory supply agreements has Nvidia reportedly signed with SK Hynix and Micron, why are these deals being pursued amid an AI driven global memory shortage expected to persist through at least 2028, how do they relate to Nvidia’s co developm

Nvidia has publicly announced a multi-year technology and supply partnership with SK Hynix, including joint work on next-generation AI memory; reports also say Nvidia has locked in multi-year DRAM and HBM supply with both SK Hynix and Micron, though the reported Micron arrangement’s precise volume, pricing, and duration have not been publicly disclosed.

• Why Nvidia is securing supply: AI accelerators require large amounts of DRAM and especially HBM, while memory makers cannot add leading-edge capacity quickly. Micron CEO Sanjay Mehrotra has said DRAM and NAND demand “significantly” exceeds supply and expects the imbalance to extend beyond calendar 2027, driven by AI demand and structural supply constraints. This supports the view that tightness could last into 2028, but that date is an industry forecast, not a guaranteed endpoint.

• SK Hynix partnership: Nvidia and SK Hynix’s announced multi-year agreement is more than a purchase contract: it covers co-development and scaling of advanced memory for Nvidia’s AI systems. That tighter roadmap integration helps ensure that future HBM generations are qualified, packaged, and available alongside Nvidia GPU-platform launches.

• Contracting model is changing: Memory has traditionally been sold through short, often annual agreements. Suppliers are increasingly seeking three-to-five-year commitments from large AI and cloud customers, trading assured demand and customer financing/visibility for allocation certainty. This reduces spot-market exposure for both sides but can also lock buyers into unfavorable pricing if the cycle reverses.

• Micron’s position and CEO comments: Mehrotra has characterized memory as strategically central to the AI era, not a peripheral component: AI systems’ performance and scale are constrained by access to memory bandwidth and capacity. Micron reported that customers were seeking materially more supply than it could commit, with one report citing requests roughly 50% above available commitments.

• Micron’s five-year agreements: As of its fiscal-Q3 2026 reporting, Micron had signed 16 strategic customer agreements, reportedly worth about $22 billion; reporting described them as five-year arrangements with pricing mechanisms intended to preserve strong margins. Those are broader customer agreements, not confirmed public details of a dedicated Nvidia five-year contract.

• Financial results and investment: Micron reported record fiscal Q3 2026 results, including $41.46 billion in revenue for the quarter ended May 28, 2026. Its then-announced U.S. plan was about $150 billion for domestic memory manufacturing plus $50 billion for research and development—roughly $200 billion in total. Micron subsequently said it was increasing planned U.S. investment to more than $250 billion through 2035, so $200 billion is no longer the most current headline figure.

• When Idaho adds supply: Idaho Fab 1 is targeted for first wafers in mid-2027, but meaningful production contribution is expected mainly in 2028; Idaho Fab 2 is targeted for late 2028. Thus, these projects do not provide a near-term remedy for the present HBM/DRAM bottleneck.

• Implications: The agreements make long-term supply more predictable for Nvidia and improve revenue visibility, utilization, and pricing power for memory manufacturers. They also suggest that leading-edge HBM will remain allocation-driven and expensive while new fabs ramp. However, they do not eliminate semiconductor cyclicality: if cloud providers and AI customers reserve capacity based on overly optimistic deployment forecasts, simultaneous demand slowing and new-fab output could create excess inventories, contract renegotiations, lower utilization, and a sharp future price correction. The key uncertainty is whether actual AI-memory consumption grows fast enough to absorb capacity coming online from 2028 onward.