On July 25, 2026, Nvidia signed away more than $500 billion in future business to a single Korean conglomerate — and industry reports say the memory chip half of that deal goes almost entirely to one company: SK Hynix.
Everyone talks about Nvidia's GPUs. Almost nobody outside the industry talks about the memory sitting right next to them on the chip package, quietly deciding how fast those GPUs can actually work. That's the part of the AI boom Korea has spent the last two years quietly cornering, and the deal signed this July made that dominance official rather than speculative.
The $500 Billion Handshake Nobody Saw Coming
SK Group and Nvidia formalized a sweeping partnership through Letters of Intent on July 25, 2026, spanning both AI infrastructure and next-generation memory. As part of it, SK Telecom will build a 2-gigawatt AI Cloud facility in Korea running on Nvidia's DSX platform and Vera Rubin accelerated computing, with the first facility targeted to launch in 2027. Separately, SK Hynix and Nvidia struck a long-term memory alliance to co-develop and secure supply of next-generation AI memory, HBM4 chief among it — giving Nvidia a locked-in supplier and giving SK Hynix a multi-year growth runway most chipmakers would kill for.
Why a Memory Chip Suddenly Became the Bottleneck
For years the AI conversation was all about compute — how many GPUs, how fast the processors. What's changed is that the processors got so fast that memory became the thing holding them back: a GPU that can't be fed data quickly enough just sits idle no matter how powerful it is. High Bandwidth Memory, or HBM, solves that by stacking memory chips vertically and wiring them directly next to the processor. HBM4 is the next generation of that stack, and it's the specific memory architecture Nvidia's upcoming Vera Rubin platform is built around — which is exactly why whoever supplies it at scale has more leverage in this cycle than at any point in the memory industry's history.
South Korea has all the ingredients to become a global AI powerhouse — world-class networks and data centers, leadership in chip technology, and vast industrial scale.
— Jensen Huang, Nvidia CEOSK Hynix vs. Samsung: The Race Inside the Race
🏭 SK Hynix
- Reached mass-production readiness for HBM4 in September 2025
- Reported to hold roughly 70% of Nvidia's Vera Rubin HBM4 demand
- Large sample volumes reportedly cleared Nvidia's final validation without issues
🏭 Samsung
- Targeting official HBM4 shipments to Nvidia and AMD from February 2026
- Using 6th-generation DRAM on a 10nm process with 4nm foundry logic integration
- Closed a historical pricing gap — HBM3E once sold roughly 30% below SK Hynix
What surprised industry watchers isn't that SK Hynix is ahead — it's by how much the gap grew. Earlier estimates had SK Hynix capturing around half of Nvidia's HBM4 demand; the number that's actually circulating now is closer to 70%. Meanwhile Samsung, despite trailing in share, has closed the price gap entirely: 12-layer HBM4 products are now expected to sell for more than $600 each, at prices roughly matching SK Hynix rather than undercutting it the way Samsung's earlier HBM3E did.
The Timeline Behind the Headline
SK Hynix establishes HBM4 mass-production capability, ahead of most competitors.
Samsung's HBM4 shipments to Nvidia and AMD begin, after completing final qualification testing.
SK Group and Nvidia sign Letters of Intent covering the $500B-plus AI infrastructure and memory partnership.
SK Telecom's first 2GW AI Cloud facility, built on Nvidia's DSX and Vera Rubin platform, is set to go live.
Personal Take: I've followed Korea's chip industry through a few of these cycles now, and what stands out this time isn't the size of the number — it's how directly memory, not logic chips, became the leverage point. Korea didn't need to out-design Nvidia's processors. It just needed to be the company Nvidia couldn't launch its next platform without.
The Bigger Picture: An Export Engine Running Hot
This deal isn't happening in isolation. Korea's exports have been on a tear through September 2026, with the country's trade surplus approaching $23 billion, driven substantially by AI chip demand. Memory chips — the unglamorous half of the semiconductor story — are doing a disproportionate share of that lifting. It's a reminder that the AI boom's biggest financial winners aren't always the companies whose logos are on the GPUs; sometimes they're the ones whose chips are stacked quietly underneath them.
📌 Key Takeaways
- SK Group and Nvidia signed Letters of Intent on July 25, 2026, for a partnership reportedly worth more than $500 billion across AI infrastructure and next-generation memory.
- Industry reports put SK Hynix's share of Nvidia's HBM4 orders for the Vera Rubin platform at roughly 70%, well above earlier 50% estimates, while Samsung has closed the pricing gap even with a smaller share.
- The deal's first visible milestone, a 2-gigawatt AI Cloud facility built by SK Telecom on Nvidia's platform, is targeted to launch in 2027.
What did SK Group and Nvidia actually agree to?
Letters of Intent signed July 25, 2026, covering a partnership reportedly worth more than $500 billion, including a 2GW AI Cloud facility in Korea and a long-term SK Hynix memory supply alliance.
What is HBM4 and why does it matter?
HBM4 is next-generation High Bandwidth Memory, stacked directly next to a GPU to feed it data fast enough to avoid becoming the bottleneck — it's central to Nvidia's upcoming Vera Rubin platform.
How much of Nvidia's HBM4 supply does SK Hynix control?
Industry reports put SK Hynix's share of Nvidia's Vera Rubin HBM4 allocation at around 70%, higher than earlier estimates of roughly 50%.
Is Samsung out of the HBM4 race?
No — Samsung began HBM4 shipments to Nvidia and AMD in February 2026 and has reached price parity with SK Hynix despite a smaller order share.
When does the SK Telecom AI Cloud facility open?
The first 2-gigawatt facility, built on Nvidia's DSX platform and Vera Rubin computing, is targeted to launch in 2027.
What I keep coming back to is how unglamorous the actual product is. Nobody puts a memory chip on a keynote stage the way Nvidia puts a GPU on one. But this July's agreement quietly confirmed something the market had only been guessing at: the company that controls the memory controls a real slice of how fast the entire AI industry can move next.
Tell me in the comments — and let me know which side of this race you're watching more closely.