SANTA CLARA, Calif. July 26 — In January 2025, Nvidia secured an exclusive multi-year agreement with SK Hynix to supply next-generation High Bandwidth Memory (HBM4) chips, a deal that locks up a critical component for the company’s upcoming AI GPUs. The arrangement was announced jointly by Nvidia CEO Jensen Huang and SK Hynix CEO Kwak Noh-jung during a press event at Nvidia’s Santa Clara headquarters.
The memory deal is tied directly to the Stargate AI infrastructure project, a joint venture disclosed on January 21, 2025, by SoftBank, OpenAI, and Oracle. Nvidia serves as a key hardware supplier for Stargate, which aims to build massive AI data centers across the United States.
Under the exclusive agreement, Nvidia’s next-generation GPUs — including those based on the Blackwell architecture — will have dedicated access to SK Hynix’s HBM4 memory. The new standard offers double the bandwidth of the previous HBM3E generation, measured at 1.6 TB/s compared to 819 GB/s. Analysts estimate the total value of the memory supply deal at over $100 billion over the next five years.
The figure underscores Nvidia’s strategy to secure essential components amid surging demand driven by the rapid expansion of AI computing. SK Hynix is the first memory maker to begin mass production of HBM3E, a milestone it reached in March 2024.
The company plans to start volume production of HBM4 by the second half of 2026 — a timeline that aligns with the delivery schedules for the Stargate project, which plans to deploy Nvidia GB200 GPUs by late 2026.
Exclusive Memory Deal
The agreement gives Nvidia guaranteed access to HBM4 chips that are not available to other customers in the same volume. SK Hynix currently controls about 50% of the global HBM market, followed by Samsung and Micron. By locking in supply from the dominant producer, Nvidia aims to mitigate supply-chain risks that have historically plagued the semiconductor industry.
The multi-year nature of the deal provides both companies with long-term production visibility and investment certainty. On the timeline, the partnership was formalized in January 2025, just days after the Stargate joint venture was made public.
The proximity of the two announcements suggests the memory agreement was negotiated in parallel with the infrastructure project’s planning stages. Sources familiar with the deal have indicated that exclusivity extends through the Blackwell GPU generation and into subsequent architectures, though specific terms were not disclosed.
Stargate Infrastructure Project
The Stargate project represents one of the largest concentrated investments in AI computing infrastructure ever announced. With a price tag announced by SoftBank, OpenAI, and Oracle, the effort brings together financial backing, software expertise, and cloud infrastructure experience. Nvidia’s role as the designated GPU supplier centers on the company’s ability to deliver high-performance chips paired with the latest high-bandwidth memory.
Deploying Nvidia GB200 GPUs by late 2026 will require a steady stream of HBM4 modules. SK Hynix’s production ramp for the new memory standard is therefore a critical path item.
The company’s experience in mass-producing HBM3E — it was the first to reach that milestone — gives it a head start on the transition to HBM4. Analysts expect that Samsung and Micron will follow with their own HBM4 production runs, but SK Hynix’s exclusive agreement with Nvidia for the Stargate volumes gives it an early revenue advantage.
Supply Chain Strategy
Locking up HBM4 capacity is part of a broader pattern at Nvidia. As AI workloads drive exponential growth in computation and memory bandwidth requirements, the company has moved to secure not just chip fabrication capacity but also advanced packaging and memory. The Stargate project alone will consume tens of thousands of HBM4 stacks over the next two years.
SK Hynix’s decision to grant exclusivity likely reflects the scale and predictability of Nvidia’s demand. The over $100 billion estimated value of the deal over five years provides the memory maker with a guaranteed revenue stream that justifies the capital expenditure needed to build out HBM4 production lines.
For Nvidia, the arrangement accelerates delivery timelines for its largest customer — the Stargate consortium — and ensures that the GPUs reaching data centers will be paired with the fastest memory available. Looking ahead, the second half of 2026 will test SK Hynix’s ability to ramp HBM4 volume production on schedule. If the company meets its target, Stargate’s initial data centers could begin receiving GB200 GPUs later this year.
The joint venture’s success will depend on the seamless integration of Nvidia’s hardware, SK Hynix’s memory, and the software stack running on top. For now, the record shows a carefully orchestrated supply chain designed to move at the speed of AI demand.
