Nvidia and SK Hynix Sign Multi-Year Agreement to Develop Next-Generation Memory
During Jensen Huang's visit to Korea, the companies officially announced a technology partnership. The collaboration aims to develop advanced HBM (High Bandwidth Memory) solutions for AI data centers, strengthening the alliance between the GPU leader and the memory leader.
Analysis: The Nvidia-SK Hynix Alliance — When "Chimaek" Decides the Fate of the AI Industry
Jensen Huang flew to Seoul not for sightseeing, but for the one resource now more critical than GPUs: memory.
While global media debates the processor race, insiders know the real bottleneck in AI infrastructure is memory. From June 5-8, 2026, Jensen Huang toured Seoul in what felt more like a diplomatic mission than a business trip. He met with the heads of SK Group, LG Group, Hyundai Motor Group, Samsung Electronics, and even threw the first pitch at a baseball game. But the main event happened over dinner at Kkanbu Chicken, where Huang and SK Group Chairman Chey Tae-won discussed the future of HBM4 and HBM4E amid the sizzle of frying oil.
Officially, a multi-year technology partnership between Nvidia and SK Hynix for joint development of next-generation memory was announced. Unofficially, Nvidia just admitted that without SK Hynix, its future Rubin Ultra GPUs simply won't take off.
[The Core]: What's Really Happening
The official story: Nvidia and SK Hynix will jointly develop HBM4 and HBM4E for "AI factories." Huang spoke eloquently about "goal alignment" and "supporting the global expansion of AI infrastructure." But the real essence of the deal is not "collaboration" — it's Nvidia's capitulation to the only supplier that can deliver the volumes it needs.
Understand this simple fact: a GPU without memory is a beautiful piece of silicon that can't compute anything. Today, every advanced AI system hits not a compute wall, but a "memory wall" — the bottleneck between the processor and the data it processes. Arm CEO Rene Haas recently called memory "probably the hardest bottleneck" the industry must solve.
What actually happened: Jensen Huang flew to Korea and told Chey Tae-won: "I'll buy everything you can produce. All your competitors have qualified too, but you are my main partner." This is not risk diversification, as the press writes. It's cementing SK Hynix's monopoly as the "first-tier supplier" for Nvidia's most important product — the Vera Rubin platform.
Why does this matter? Because the volumes of HBM4 that Nvidia will need are so enormous that SK Hynix has already announced plans to double production capacity within five years. And all of that is for one customer. Analysts estimate SK Hynix controls 60-70% of HBM4 supply for Vera Rubin, Samsung 25-30%, and Micron the remainder.
Timeline and Context
To understand why this deal became inevitable, trace the chain of events over the past two months. Some went almost unnoticed by the public.
May 2026, Computex, Taipei: Jensen Huang appears at the SK Hynix booth and personally writes on an HBM4E wafer with a marker: "PLEASE MAKE MORE." This is not a marketing gesture — it's the cry of a man who knows his business is hitting the physical limits of production. At the same show, he officially confirms that all three memory makers (SK Hynix, Samsung, Micron) have passed HBM4 qualification for the Vera Rubin platform.
Late May 2026: Samsung delivers HBM4 samples to Nvidia, demonstrating a "technological breakthrough." But insiders know: the gap between "passed qualification" and "landed a major contract" is a chasm. SK Hynix is already shipping HBM4 in volumes sufficient to start production.
June 4-5, 2026: Huang arrives in Seoul. At the airport, he tells journalists: "I came to thank all of Nvidia's partners and customers." Then he meets with SK Hynix and SK Group leadership over "chimaek" (fried chicken and beer) — already the second time this year. That evening, it becomes known that all three Korean HBM makers (SK Hynix, Samsung, and US-based Micron) have received the "green light" for shipments. Huang adds that he is "preparing a surprise" for Korea.
June 7-8, 2026: Official announcement of a multi-year technology partnership. The wording is evasive, but the essence is clear: SK Hynix becomes not just a supplier, but a co-developer of HBM for Nvidia.
In parallel: SK Hynix showcases at Computex the first sample of 12-layer HBM4E with 48 GB capacity and 4 TB/s speed — 33% higher density and 37% higher clock frequency than HBM4. Mass production is planned for 2027.
Here's the key: Samsung also has HBM4 solutions, and their products are "technically" not inferior to SK Hynix. But SK Hynix wins not on technology — but on scale and trust. They have already proven they can consistently deliver HBM3 and HBM3E in huge volumes. Samsung is still recovering from issues with the previous generation.
Who Wins and Who Loses
Winner #1: SK Hynix. This is obvious. The company that a few years ago was the "second player" after Samsung in memory is now the king of HBM. Its global HBM market share in Q1 2026 was 57-58%. A multi-year contract with Nvidia means this status is locked in for years. Investors have taken note: SK Hynix shares have risen 240% year-to-date and trade near an all-time high.
Winner #2: Nvidia (strategically). Yes, Nvidia has to share profits and accept supplier terms. But there is no alternative. Without guaranteed HBM4 supply, the Vera Rubin platform (which is supposed to start shipping in Q3 2026) simply cannot launch in planned volumes. Huang made a deal that gives him "supply visibility" for years ahead. In an environment where AI infrastructure demand grows in waves, each larger than the last, this is invaluable.
Winner #3: SK Hynix shareholders. See point 1. 240% growth in a year is no accident. It's the market realizing that SK Hynix has become the "indispensable supplier" for the leading company of the AI era.
Winner (unexpectedly) #4: Samsung and Micron. Yes, they lost the race for share, but they are still in the game. Huang personally confirmed that Samsung's products have qualified, and Samsung is already shipping HBM4 samples. Micron is also on the list. Nvidia cannot afford to bet everything on one supplier — that's too risky. So Samsung and Micron will get roughly 30-40% of orders. Less than SK Hynix, but still billions of dollars. After Samsung's failure with previous HBM generations, returning to Nvidia's fold is a victory.
Loser #1: Intel and AMD. They don't have a memory partner like Nvidia does. Intel is trying to grow its foundry business and promote HBM-like solutions through its EMIB packaging technology, but it's nowhere near what Nvidia has built with SK Hynix. AMD also depends on the same suppliers, but lacks Nvidia's "scale effect" to dictate terms.
Loser #2: Nvidia's customers (Microsoft, Google, Amazon, Meta). They will pay more. Because HBM4 is scarce, and scarcity means high prices. SK Hynix already signals it cannot meet 100% of demand even at maximum capacity. Nvidia's customers will either pay a premium for priority shipments or wait in line. Their AI plans will slow down.
Loser #3: China. Beijing is closely watching this alliance. The US + South Korea + Taiwan (TSMC) close the supply chain for the most advanced AI components. China is left out. Its companies (Huawei, SMIC) are trying to develop alternatives, but they are years, if not a decade, behind. This contract is further proof that the "Western" (including Korea) technology bloc is consolidating.
What the Media Isn't Saying
Typical news feeds will reprint headlines about a "multi-year partnership" and move on. But there are three things you won't read in press releases.
Insight #1: This is less about HBM4 and more about HBM4E and 2027.
Notice the dates. SK Hynix is already shipping HBM4. HBM4E samples were shown at Computex, but mass production is only in 2027. And what happens in 2027? Nvidia launches the Rubin Ultra platform. That's what HBM4E is for. The contract signed now is a preliminary "reservation" for 2027 capacity. Huang understands: if he doesn't secure HBM4E today, it will be too late in a year — all of SK Hynix's capacity will be booked for years ahead.
Why does this matter? Because TSMC, which handles packaging (CoWoS) for Nvidia chips, also has limited capacity. The combination of HBM4E shortage and packaging shortage is a perfect storm for 2027. Nvidia is trying to "stake a place in line" now, 12-18 months before production starts.
Insight #2: Huang was bluffing about "supplier diversification."
At Computex, Huang said: "All three companies have qualified." Many took this as a signal that Nvidia is diversifying risk and won't depend on one supplier. But the multi-year contract with SK Hynix says otherwise. Yes, Samsung and Micron will get orders. But SK Hynix is the "first among equals." It gets the lion's share, and more importantly, it participates in co-development.
What does this mean in practice? SK Hynix will know the specifications of future Nvidia GPUs years in advance. They can optimize their memory specifically for Nvidia's architecture, creating a "technology gap" between Nvidia+SK Hynix and everyone else. This is not "diversification" — it's an "exclusive alliance" dressed up with public statements about openness.
Insight #3: "Physical AI" and robots are the real purpose of the visit.
Western media only discuss memory. But look at Huang's agenda in Korea. He met with Doosan Group (robotic manufacturing), Hyundai Motor Group (cars and robots), LG Group. And with SK Telecom, which is building a gigawatt-scale AI cloud platform and developing digital twins for SK Hynix factories.
What's happening here? Nvidia is betting on "physical AI" — robots that interact with the real world. And Korea is a global hub for robotics and automation. Huang came not just for memory. He came to tie Korean industry to Nvidia's ecosystem: Omniverse for digital twins, platforms for robot training, GPUs for AI factories. The SK Hynix contract is the "hook." Hyundai and Doosan robots are the "bait."
Forecast: Next 30 Days and 90 Days
Next 30 Days (through mid-July 2026)
1. Renegotiation of Samsung and Micron contracts. After the exclusive partnership announcement with SK Hynix, Samsung and Micron will demand written volume guarantees from Nvidia. Nvidia will likely agree — but with lower numbers than competitors want. Expect news of "tripartite negotiations" in the press. Samsung shares may correct 3-5% on fears they are losing the race.
2. Rally in semiconductor stocks. The multi-year contract news is already priced into SK Hynix (240% growth), but it will boost the entire sector. TSMC (which handles packaging) will get an additional boost. ASML (lithography equipment maker for HBM) will also rise — because SK Hynix needs to expand capacity.
3. Official announcement of a new SK Hynix fab. Under this multi-year agreement, SK Hynix will announce concrete expansion plans — likely in South Korea (North Chungcheong Province) or possibly in the US under pressure from the CHIPS Act. Investment will be at least $15-20 billion. Announcement expected within 30 days.
Medium-Term Forecast (90 days, through September 2026)
1. Official launch of the Vera Rubin platform. In Q3 2026 (i.e., the next 2-3 months), Nvidia will announce the start of shipments of Vera Rubin-based systems. At that point, it will become clear how well SK Hynix is handling volumes. If there are delays, Huang will personally fly to Seoul again. If not, Nvidia's stock price may break another psychological barrier ($5 trillion market cap?).
2. Micron may drop out of the race. Micron has less capacity than its Korean competitors. If SK Hynix and Samsung fill their fabs to the brim with Nvidia orders, Micron may lack volume to secure large contracts with other customers (AMD, Google). This could lead to HBM market consolidation — either Micron will seek a partner, or its share will shrink to a symbolic 5-10%.
3. Huawei will make a countermove. The Chinese manufacturer, cut off from advanced technology, will announce a "breakthrough" in its own HBM-like memory. This will be a propaganda victory, but technically an outdated solution. Still, it will be enough for China's domestic market. Expect loud headlines from Beijing about "overcoming the American blockade."
4. Emergence of new startups in "memory-centric computing." The HBM bottleneck will spur development of alternative architectures — in-memory computing, processors based on Resistive RAM (ReRAM), and other exotic technologies. Investment in such startups will rise 30-40% in the next three months. Watch Israeli and US deep-tech funds — they are already looking for the "next SK Hynix."
Summary for those who read this far: The Nvidia-SK Hynix alliance is an admission that the era of "just GPUs" is over. The future of AI infrastructure is determined not by teraflops, but by terabytes per second. Whoever controls memory controls AI. Today, that control belongs to South Korea. Jensen Huang flew to Seoul not as a victor, but as a supplicant. He asked Chey Tae-won for help. And he got it. But the price of that help is Nvidia's dependence on SK Hynix for years to come. And that may be the smartest bet Korean business has ever made.
Don't watch the GPUs. Watch the memory. Where silicon ends, true power begins.
— Editorial Team
No comments yet.