SK Group Chairman Chey Tae-won declared Tuesday at Computex 2026 in Taipei that SK hynix will double its total wafer production capacity over the next five years — a major strategic reversal from March, when Chey told reporters at Nvidia's GPU Technology Conference that adding capacity was not something the company had planned. Hours later, Nvidia CEO Jensen Huang walked up to the SK hynix booth, picked up a marker, and wrote "Please Make More" on an HBM4E wafer on display — the world's largest AI chip designer publicly asking its top memory supplier to move faster.
The two messages — one from the supplier's chairman and one from its most important buyer — landed at the same trade show on the same afternoon, producing the most direct public exchange about AI memory supply that either company has made.
HBM Shortage Forecast Holds Through 2030
Chey repeated a forecast he first made in March: the supply gap in the memory market is likely to persist until 2030, with wafer supply running more than 20 percent below demand across the industry. The reason is structural — not cyclical. High-bandwidth memory, the chip type that Nvidia's AI accelerators require to function, consumes far more wafer capacity per bit than conventional DRAM. Every new generation of AI hardware demands more HBM in faster configurations, from the same finite pool of manufacturing lines. As demand for AI accelerators accelerates and data center operators expand, the pressure on that pool grows faster than any single fab can relieve.
Chey was direct about the difficulty of adding supply. Building a new memory fab from an existing site takes at least three years; a greenfield facility takes more than five — placing even an immediate construction start near the tail end of the shortage window his own forecast describes. The cost cannot be fixed in advance, he said, because prices for land, equipment, and electricity are all moving simultaneously. What he did commit to was the spending itself: SK hynix would spare no investment needed to build the products its customers want.
That commitment represents a significant shift. In March, Chey had described new capacity as something that "couldn't be added on demand." The reversal in June reflects both the persistence of demand — SK hynix's entire 2026 HBM production is already sold out — and pressure from customers who have reportedly offered to purchase the company's own EUV lithography scanners and prefund new production lines because available capacity has fallen to near zero.
Jensen Huang Visits SK hynix Booth, Signs HBM Wafer in Taipei
Huang's visit to the SK hynix booth on Tuesday was his second consecutive day of meeting SK hynix management. The two executives had sat down the evening before at a private dinner in Taipei that Nvidia organized for Korean technology partners — the first time the company had convened such an event specifically for Korean firms at Computex. On the HBM4E wafer, Huang signed and wrote "Please Make More." On a SOCAMM next-generation memory module displayed alongside it, he wrote "Love SOCAMM." The handwritten notes functioned as a public statement: the CEO of the world's dominant AI chip designer, whose platforms consume more HBM per unit than any predecessor, asking its primary memory supplier to accelerate output.
Chey described the relationship with Huang as one built on trust, and declined to detail their private conversations. Asked whether SK hynix would remain Nvidia's leading HBM supplier as the Vera Rubin platform enters mass production — a platform for which SK hynix reportedly secured roughly 60 to 70 percent of HBM4 orders — he gave a cautious answer. He said he hoped so, noting the final supply decision belongs to Nvidia, but that SK hynix has always been a key partner. Huang confirmed separately that the Vera Rubin platform had entered full-scale production.
HBM4E Race: Why One Customer Changes the Calculation
On the competition for next-generation chips, Chey's remarks contained a pointed observation. Asked about the roadmap for SK hynix's HBM4E — the seventh-generation HBM standard that Samsung announced on May 29 it had already shipped in sample form — Chey said the schedule depends on the customer. "There's only one customer for HBM4E right now," he said, referring to Nvidia. The remark was widely read as a response to Samsung's early-sampling milestone: in Chey's framing, leading on samples matters less than leading on qualified production once actual demand exists. SK hynix plans to deliver its own HBM4E samples in the second half of 2026, with mass production scheduled for 2027.
Samsung also unveiled an HBM5 mockup at the same Computex show, signaling its intention to compete in the generation beyond HBM4E. The dual-generation display underscored how compressed the HBM roadmap has become under AI demand pressure.
SK hynix Capital Position and Pending New York Listing
The capacity commitment arrives as SK hynix is negotiating a new position in global capital markets. Last week, the company crossed $1 trillion in market capitalization for the first time — a 244 percent gain since October 2025 — joining Samsung and Micron in a group of memory chipmakers whose combined value now reflects the market's assessment of AI infrastructure demand for the rest of the decade. Chey confirmed that SK hynix has also filed for a listing of American depositary receipts on the New York Stock Exchange, targeting a debut in the second half of 2026. The proceeds are expected to support the capacity expansion and reduce the company's dependence on the Korean Exchange as its primary funding venue.
Goldman Sachs raised its 2028 operating profit forecast for SK hynix by 24 percent following Tuesday's announcement, citing sustained AI demand as the driver. The company reported first-quarter 2026 revenue of 52.6 trillion won — roughly $35.5 billion — at an operating margin of 72 percent, the highest in its history.
On pricing, Chey drew a distinction. He said memory prices should track supply and demand but warned that sudden spikes would damage the broader ecosystem's ability to grow sustainably. A supplier with 58 percent of the global HBM market — a share Samsung and Micron together match at roughly 21 percent each — holds the pricing leverage to make that statement mean something.
What Does the AI Memory Chip Shortage Mean for Infrastructure Buyers?
For enterprise buyers, data center operators, and AI infrastructure planners, the central implication of Tuesday's announcement is timing. Chey's five-year doubling timeline — placed against his own three-to-five-year fab construction lead time — means no new supply from this expansion reaches the market before 2029 at the earliest under the most optimistic scenario. The shortage Chey forecasts through 2030 will not be resolved by capacity that cannot come online until the shortage window is nearly closed.
In the near term, analysts differ on exactly how long the worst pressure lasts. Sanchit Vir Gogia of Greyhound Research described the constraint as "structural reallocation of the memory market driven by AI infrastructure economics" and argued the 2030 forecast was plausible given AI infrastructure growth trajectories. Shrish Pant of Gartner offered a more qualified read, noting that the 2030 horizon assumes uninterrupted AI demand and that traditional DRAM prices could improve by 2028 as facilities under construction in South Korea and the United States begin to contribute. Both agreed that meaningful price relief is unlikely before late 2027.
The consumer implication is already visible. The same reallocation of wafer capacity toward high-margin HBM has tightened supply for conventional memory used in smartphones, laptops, and servers, driving price increases across categories. For anyone procuring memory at scale, the message from Computex is that multi-year contracts, early commitments, and supply-chain redundancy are not optional hedges — they are structural necessities for the rest of the decade.
Frequently Asked Questions
Why is SK hynix doubling its wafer capacity now?
AI demand for high-bandwidth memory has outpaced supply faster than SK hynix anticipated, with the company's entire 2026 HBM production already sold out. Chairman Chey Tae-won cited a structural shortage he expects to persist through 2030, driven by AI accelerators that require far more wafer capacity per chip than conventional DRAM, making expansion a strategic necessity rather than an optional investment.
How long will the AI memory chip shortage last?
SK hynix Chairman Chey Tae-won forecast the shortage will persist until at least 2030, with wafer supply running more than 20 percent below demand industry-wide. Gartner's Shrish Pant indicated meaningful price relief is unlikely before late 2027, while Greyhound Research's Sanchit Vir Gogia characterized the constraint as a structural reallocation of the memory market rather than a temporary cycle.
What is high-bandwidth memory and why does AI need it?
High-bandwidth memory is a form of DRAM in which multiple chips are stacked vertically and connected through microscopic copper pathways, enabling data transfer rates far exceeding standard memory. AI accelerators from Nvidia, AMD, and Google require HBM because the models they run must move enormous volumes of data between compute and memory at every processing step; without it, the logic die inside a GPU cannot be fed data fast enough to run large AI workloads efficiently.
Will SK hynix remain Nvidia's primary HBM supplier for Vera Rubin?
Chey said he hoped so but acknowledged the final supply decision rests with Nvidia. SK hynix reportedly secured roughly 60 to 70 percent of HBM4 orders for the Vera Rubin platform, and Huang confirmed Vera Rubin had entered full-scale production. Samsung, which shipped HBM4E samples first and unveiled HBM5 mockups at Computex, is competing aggressively for a larger share of next-generation allocations.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
