SK Hynix Becomes Nvidia’s AI Memory Co-Designer in Seoul Deal

Nvidia and SK Hynix signed a multiyear AI memory partnership in Seoul on Monday that covers the chipmaker’s entire product roadmap, from Vera Rubin AI supercomputers due in data centers this quarter to RTX Spark laptops shipping this fall and Jetson Thor robotics modules. Beyond supply commitments, the deal embeds Nvidia’s own AI engineering tools inside SK Hynix’s semiconductor fabrication processes, a scope that reaches well past any conventional purchase agreement.

Hours before the signing, South Korea’s benchmark Kospi index had dropped as much as 8.8%, triggering a 20-minute circuit breaker, as investors shed AI-linked stocks following weaker-than-expected quarterly guidance from Broadcom. The chipmaker’s founder and CEO Jensen Huang, in Seoul since June 5 for a schedule that included five separate sessions with SK Group Chairman Chey Tae-won, told reporters after their meeting Monday that anyone watching the selloff “should be very happy because now you can buy at a discount.” The deal he had signed that morning concerns a timeline measured in product generations.

Five Days in Seoul

Huang arrived June 5 and his first stop was a gaming venue, not a boardroom. At T1 Base Camp in Seoul’s Mapo district, he gifted Lee “Faker” Sang-hyeok, the six-time League of Legends World Champion and T1 team captain, a personally signed GeForce RTX 5090 graphics card. Some 500 fans packed into the venue collected signed vouchers for the incoming RTX Spark laptop, the consumer AI platform Huang had unveiled at Computex in Taipei days earlier. That evening he sat down with SK Group Chairman Chey Tae-won, LG Group Chairman Koo Kwang-mo, and Naver Chairman Lee Hae-jin at a barbecue restaurant near Hongik University for what Korean media tagged a “Sam-So” gathering, named for samgyeopsal and soju.

It was his second South Korea trip in seven months. At the Asia-Pacific Economic Cooperation (APEC) summit in October 2025, he pledged more than 250,000 advanced GPUs for South Korean AI infrastructure alongside Samsung, SK Hynix, and Hyundai Motor. This trip ran through Saturday individual meetings with those same three companies, a ceremonial first pitch at a Korea Baseball Organization (KBO) League game at Jamsil Stadium on Sunday, then a second dinner in Gangnam that night alongside the SK Group chairman and SK Hynix CEO Kwak Noh-jung.

The memory theme had been running for a week before any deal was formally announced. At Computex the week prior, Huang stopped by SK Hynix’s exhibition booth and left a handwritten note on a displayed HBM (high-bandwidth memory) wafer: “PLEASE MAKE MORE.” Arriving in Seoul on June 5, he confirmed publicly that Samsung, the South Korean company, and Micron Technology had all cleared HBM4 (sixth-generation high-bandwidth memory) qualification for the platform. He also told reporters the memory shortage would persist for “quite a few years.”

What the Co-Design Agreement Covers

The multiyear technology partnership announced June 8 spans four product lines and runs two tracks simultaneously: conventional memory supply, and the co-engineering of semiconductor design and manufacturing processes themselves.

On the supply side, the South Korean company will co-develop sixth-generation memory for Vera Rubin AI supercomputers and Vera CPUs, targeting Q3 2026 commercial delivery. Vera Rubin systems deliver 10 times the agent throughput of the prior Grace Blackwell platform, per the chipmaker’s own published specifications. The company will also supply LPDDR5X (low-power DRAM, the memory format used in thin laptops and compact desktops) for Nvidia’s RTX Spark laptop platform, due from ASUS, Dell, HP, Lenovo, Microsoft, and others this fall. Jetson Thor, the compute module for robotics and autonomous systems, completes the four-platform scope.

The co-engineering track is where the agreement departs from a supply contract. Both companies will apply CUDA-X, Nvidia’s suite of GPU-accelerated computing libraries, and PhysicsNeMo, the AI physics simulation framework, to accelerate TCAD (technology computer-aided design, the modeling methodology that simulates transistor behavior before committing to fabrication) workflows and internal engineering codes at the South Korean company’s factories. They will also deploy Omniverse, the 3D industrial simulation platform, alongside cuOpt, the GPU logistics optimization library, to build factory digital twins – persistent virtual models of fabrication equipment and production flows – with the stated goal of fully autonomous fab operations.

“Together, we are co-developing the next generation of memory for AI factories and applying AI to how we design and manufacture semiconductors,” said SK Group Chairman Chey Tae-won at Monday’s announcement. The press release adds that the agreement “builds on years of deep co-engineering collaboration.”

Samsung’s 25 Percent Problem

Samsung Electronics also cleared qualification for the new memory generation – the chipmaker’s CEO confirmed all three major vendors on June 5, saying they were “all racing to support” the platform. The Korean rival began mass production of sixth-generation memory in February and announced updated HBM4 product details publicly this week.

The allocation math sits heavily on one side. Supply-chain analysts cited by Yonhap expect the South Korean partner to hold roughly 60 to 70 percent of next-platform sixth-generation memory volume, with the Korean rival at 25 to 30 percent and Micron Technology covering the balance. Official volume figures have not been published. A TrendForce HBM4 supply analysis from January 2026 cited yield consistency and large-scale production track record as the deciding factors, putting expected allocation at close to two-thirds of the chipmaker’s demand. UBS, in a market outlook the company published in April 2026, predicted approximately 70 percent of that volume, above the roughly 50 percent some earlier forecasts had assumed. Counterpoint Research put the South Korean company’s global high-bandwidth memory shipment share at 53 percent in Q3 2025, ahead of the Korean rival at 35 percent and Micron at 11 percent.

• 60-70% of next-platform memory volume estimated for the South Korean partner, per Yonhap and UBS
• 53% global high-bandwidth memory share in Q3 2025, per Counterpoint Research
• $600+ per stack, expected price for sixth-generation memory, per industry sources cited by TrendForce
• $43 billion projected global high-bandwidth memory market by 2027, per Citi Research

The Korean rival restructured its DRAM development teams under a unified engineering organization this year to compete in this cycle, and at least one executive reported customers saying “Samsung is back” on the new product generation. Both companies hold certified supplier status for the current platform. The co-design agreement covers the platforms after it; the Korean rival’s publicly disclosed arrangement with the chipmaker covers supply for this generation.

Why the Kospi Circuit Breaker Fired

The selloff traced back to one quarterly report. On June 3, Broadcom guided Q3 AI chip sales at $16 billion – more than 200 percent above the prior year, but roughly 7 percent below the market’s $17.2 billion estimate. Full-year AI chip guidance of $56 billion also fell short of the consensus $57.6 billion. Broadcom shares fell about 13 percent on June 5, and the damage swept through semiconductor stocks globally.

South Korea’s exposure ran deep. The Kospi had posted cumulative gains exceeding 100 percent at its year-to-date peak, built almost entirely on chip stocks. Retail margin debt in the country reached 37.74 trillion won (roughly $24.2 billion at the exchange rate of about 1,554 won per US dollar as of that week), per market data cited by financial analytics platform TradingKey. Once foreign funds turned sellers, forced margin liquidation amplified the exit: foreign investors had net sold approximately $22 billion in Korean stocks since May, the South Korean memory company accounting for about $12 billion of those outflows. By Monday morning, the Kospi had fallen 8.8% and a circuit breaker halted trading for 20 minutes. The Korean rival’s shares dropped as much as 11 percent; the South Korean memory company slid 10 percent.

We’re at the beginning of it, and whatever happened to the stock market, you should be very happy because now you can buy at a discount.

The chipmaker’s founder and CEO said those words to reporters Monday after meeting with the SK Group chairman. South Korean President Lee Jae Myung separately stated that day he believed the domestic market remained undervalued. Shares of both chipmakers pared some losses after the remarks, though the index had still posted one of its sharpest single-session drops in months.

When AI Designs Its Own Memory

Monday’s manufacturing provisions address a supply constraint that can’t be solved by ordering more chips. Each high-bandwidth memory generation stacks more DRAM dies on a base logic chip using through-silicon vias – microscopic vertical connectors etched through the silicon itself. The current sixth generation uses 12-die stacks. The chipmaker has reportedly asked suppliers to target 16-die configurations for Q4 2026 delivery, per reporting by Electronic Times. Getting there requires narrowing wafer thicknesses from roughly 50 micrometers today down to about 30 micrometers, a fabrication challenge that stretches qualification timelines and limits how quickly output can scale.

CUDA-X libraries and PhysicsNeMo are built for exactly that class of bottleneck. The TCAD workflows that predict transistor behavior before a wafer is processed are compute-intensive and slow to iterate on. GPU-accelerated AI shortens the cycle between each process change and a validated production output. The Omniverse digital twin layer adds another layer: factory engineers can simulate an entire production line before touching physical equipment, compressing the cost of each trial. These are software tools the chipmaker already sells to hyperscale data centers. Embedding them inside a memory partner’s factory workflow closes a loop: the chips training today’s AI models also help engineer the memory that goes into the next generation of those chips.

“It is a foregone conclusion that AI will be infrastructure for the world, just like the internet was infrastructure for the world,” the chipmaker’s CEO told reporters during his Seoul visit. The co-design agreement plants that thesis inside a memory fab. Next-generation AI supercomputer systems are set for commercial delivery to data center customers this quarter, and every memory stack inside those servers will carry the South Korean partner’s name.

Disclaimer: This article is for informational purposes only and does not constitute investment advice. References to stock market movements, executive commentary on market conditions, and valuation estimates are reported as news and should not be relied upon as guidance for individual investment decisions. Figures are accurate as of publication. Consult a qualified financial professional before making financial decisions.