Nvidia and SK Hynix have signed a multi-year agreement covering both the design and manufacture of next-generation memory chips for AI. The deal, announced on Sunday during Nvidia CEO Jensen Huang’s tour of South Korea, gives SK Hynix a formal co-development role in the high-bandwidth memory that will power Nvidia’s most advanced accelerators, starting with the Vera Rubin platform now entering full production.

The agreement arrives at a moment when memory, not GPUs, has become the binding constraint on AI infrastructure expansion. Arm CEO Rene Haas said last week that memory is “probably the toughest” bottleneck the industry has to resolve. Nvidia is trying to fix that by locking in its supply chain years in advance.

What the deal covers

The partnership extends beyond a standard supply agreement. Nvidia and SK Hynix will co-develop next-generation memory for what Nvidia calls “AI factories,” the large-scale data centre clusters used for training and inference. The scope covers infrastructure, physical AI, and memory specifically designed for Vera Rubin, Nvidia’s most powerful accelerator platform.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

Vera Rubin is built around clusters of Vera central processing units and Rubin graphics cores, allied with terabytes of HBM4 in each server system. It delivers reportedly 3.5 times the training performance and five times the inference performance of its Blackwell predecessor. Shipments are expected to begin in Q3 2026, with over 350 supply chain partners across 30 countries involved in production.

“Together, we will co-develop the next generation of memory for AI factories and support the accelerating global expansion of AI infrastructure,” Huang said in a statement.

The HBM4 race

Huang confirmed for the first time last week at Computex in Taipei that all three major memory manufacturers, Samsung, SK Hynix, and Micron, have been cleared to supply HBM4 for Vera Rubin. But the multi-year co-development pact with SK Hynix signals a deeper relationship than a standard vendor qualification.

Industry analysts estimate SK Hynix holds roughly 60% to 70% of HBM4 volume allocated to Vera Rubin, with Samsung at approximately 25% to 30% and Micron supplying the remainder. The new agreement makes SK Hynix’s position more durable. Long-term supply commitments make it easier for the company to expand capacity and gradually increase market share.

The competition between the three is fierce. Memory chip stocks have rocketed over the past year on surging prices, and all three firms are racing to deliver 16-layer HBM stacks that Nvidia has reportedly requested for delivery as early as late 2026.

Huang’s Seoul tour

The SK Hynix deal was part of a broader series of announcements during Huang’s high-profile visit to South Korea. Nvidia also announced that SK Telecom will build a new gigawatt-scale AI cloud powered by Nvidia chips, with the first data centre coming online early next year. Naver, Korea’s dominant internet platform, will use Nvidia’s AI models to expand its data centre capacity and build additional gigawatt-scale AI factories. Doosan Group will use Nvidia’s physical AI technology to power its industrial robotics.

Huang threw the ceremonial first pitch at a Korean Baseball Organisation game between the Kiwoom Heroes and Doosan Bears on Saturday, dined with SK Group Chairman Chey Tae-won, and visited gaming studios Krafton and NC Corp to promote adoption of Nvidia’s RTX Spark chip, its new entry into the PC market.

The pattern mirrors Nvidia’s approach across Asia. At Computex in Taipei, Huang announced the Cosmos Coalition, new partnerships with Chinese robotics firms, and the launch of Cosmos 3. In Seoul, the focus shifted to memory and infrastructure. The message is consistent: Nvidia is building an ecosystem, not just selling chips, and it is doing so country by country.

Why memory matters more than GPUs right now

The AI industry’s demand for compute is well documented. Less discussed is the memory wall. Each Vera Rubin NVL72 configuration connects 36 CPUs and 72 GPUs, and the entire system requires terabytes of HBM4. Advanced packaging technology, specifically TSMC’s CoWoS process that integrates GPU dies with HBM into a single package, is a key constraint. Memory supply, not silicon fabrication, is now the pacing factor for how fast Nvidia can ship its highest-end systems.

Haas’s comment about memory being the toughest bottleneck reflects a structural shortage that analysts at TrendForce have described as a “memory supercycle.” HBM capacity is projected to remain tight through at least 2028, and some forecasts extend the constraint to 2030. For Nvidia, locking SK Hynix into a multi-year co-development deal is not just about securing supply. It is about ensuring the supply exists at all.