Jensen Huang touched down in Seoul on Sunday, June 7, 2026, delivering a stark warning to the global tech industry: the AI memory shortage is not a temporary blip — it will persist for "the next several years."
Key Highlights
- Huang met SK Group Chairman Chey Tae-won over informal "chimaek" dinner in Gangnam, Seoul
- "Everything across the entire industry supply chain — from wafers to silicon photonics to cable connectors — is in short supply," Huang said
- At Computex 2026 on June 2, Huang wrote "Please Make More" on an HBM4E wafer at SK Hynix's booth
- Wafer supply is running more than 20% below demand, with the gap expected to persist through 2030
- SK Hynix plans to double DRAM wafer capacity from 550,000 to 1 million wafers per month by 2030
- SK Hynix supplies an estimated 50–70% of Nvidia's HBM4 requirements for Blackwell GPUs
- Nvidia and SK Hynix are formalizing a multi-year strategic partnership
The "Please Make More" Moment
On June 2, 2026, at Computex in Taipei, Jensen Huang walked up to SK Hynix's exhibition booth, picked up a marker, and wrote three words on a gleaming HBM4E wafer on display: "Please Make More." The photograph spread instantly across the industry, crystallizing what had been an abstract supply constraint into a visceral image — the world's most powerful chip designer publicly asking its top memory supplier to accelerate faster than physics typically allows.
The gesture was symbolic, but the underlying crisis is structural. HBM (High Bandwidth Memory) chips are the memory technology that powers Nvidia's GPU clusters. They sit directly atop the GPU die and feed data faster than any conventional DRAM could manage. Without HBM, there are no Blackwell systems. Without Blackwell systems, there is no frontier AI infrastructure.
Seoul Dinner and a Partnership in the Making
On June 7, Huang flew to Seoul for meetings with Korea's top chipmakers. Over "chimaek" — the beloved Korean combination of fried chicken and beer — Huang dined with SK Group Chairman Chey Tae-won at Kkanbu Chicken in Gangnam. Huang was also scheduled to meet Samsung Electronics Vice Chairman Jun Young-hyun on June 8, having already had a prior dinner in California with Samsung Chairman Jay Y. Lee.
The conversations produced a framework for an expanded Nvidia–SK Hynix partnership that industry sources say will include multi-year volume commitments spanning several GPU generations, accelerated HBM4 and HBM4E production ramps, joint yield optimization programs, and custom memory configurations for AI factories.
SK Hynix currently supplies an estimated 50 to 70 percent of Nvidia's HBM4 requirements — making it a critical dependency for both Blackwell accelerators and the upcoming Vera Rubin platform.
A Shortage That Runs Deeper Than Memory
Huang's warning extended well beyond DRAM chips. "Demand is enormous. Everything across the entire industry supply chain — from wafers to silicon photonics to cable connectors — is in short supply," he told reporters in Seoul.
This is significant. Silicon photonics are the optical interconnects that link GPU clusters across racks. Cable connectors are the physical links between compute nodes. If both are in short supply alongside HBM, it signals that the AI buildout has outrun the industrial base required to support it — across multiple technology categories simultaneously.
According to analysts, the AI memory supply gap is expected to remain above 20% through 2030. SK Hynix currently processes 550,000 DRAM wafers per month. The company has now committed to doubling that to 1 million wafers per month by 2030 — a reversal from March 2026, when Chairman Chey told reporters at Nvidia's GTC that adding capacity was not in the company's plans.
Impact on the AI Infrastructure Race
The timing matters. Huang projected "much greater growth" in H2 2026 versus H1, with "substantial growth through next year" — a trajectory that depends on the memory supply situation improving. Any slippage in HBM availability could delay Blackwell and Vera Rubin shipments to hyperscalers, constraining the GPU supply that AI companies are already fighting to secure.
For the MENA region, where Saudi Arabia, the UAE, and Qatar are each committing billions to sovereign AI infrastructure, the memory shortage adds a new constraint. Projects like Saudi Arabia's HUMAIN AI initiative — which committed $3 billion to xAI alone — rely on Nvidia GPU allocations that are themselves limited by HBM availability. A shortage lasting through the early 2030s means that access to AI compute will remain a geopolitical asset, not merely a procurement decision.
What's Next
SK Hynix's HBM4E samples are expected in the second half of 2026, with mass production scheduled for 2027. Nvidia and SK Hynix planned to brief media on their formal cooperation agreement on June 8. The expanded partnership — covering joint AI factory development in Korea — is designed to structurally lock in memory supply across multiple generations of Nvidia GPU platforms.
Huang's message in Seoul is unambiguous: the AI memory crunch is a multi-year constraint that will shape which companies and countries can build AI infrastructure at scale — and which cannot.
Source: Seoul Economic Daily