Nvidia and SK Group formalized the AI industry's most comprehensive hardware partnership on June 7 and June 8, 2026, committing to co-develop next-generation memory across all four of Nvidia's upcoming product lines and to build the first full-stack AI-factory cloud in Korea — a deal that Nvidia CEO Jensen Huang described as the company's first agreement to span multiyear contracts, multiple platforms, and multiple technologies simultaneously. The announcement, made during a press briefing at SK's Seorin headquarters in Seoul on June 8, elevated what had been a product-level memory relationship into a group-wide alliance covering semiconductor design, cloud infrastructure, and autonomous manufacturing.

The scope of the commitment marks a shift in how AI hardware partnerships are structured. Per the official announcement, SK hynix — which holds an estimated 60 to 70 percent of the High Bandwidth Memory allocation for Nvidia's Vera Rubin platform — will now co-develop memory for three additional Nvidia product lines: the Vera CPU, the RTX Spark personal AI computer, and the Jetson Thor robotics platform. This extends a relationship that once focused on a single chip generation into a roadmap covering AI supercomputers, personal computing, and physical AI simultaneously. Samsung Electronics and Micron Technology supply the remaining Vera Rubin HBM4 volume, but neither has a comparable co-development agreement spanning Nvidia's full product portfolio.

SK Group Chairman Chey Tae-won, speaking alongside Huang on June 8, described the deal as an elevation from memory supply to full-group collaboration. "Until now, we have collaborated with Nvidia on SK hynix's memory business," he said. "Now we are elevating our partnership to the entire SK Group." The full SK hynix press release details all three pillars of the agreement.

HBM4 Architecture: Why Stacked Memory Defines AI Factory Throughput

The memory at the center of this agreement — High Bandwidth Memory 4, or HBM4 — is architecturally unlike the DRAM found in consumer devices. Rather than spreading memory chips across a flat circuit board, HBM4 stacks multiple DRAM dies vertically and connects them through microscopic through-silicon vias — vertical wires running through each silicon layer. This three-dimensional stack sits directly beside the GPU on a shared silicon interposer, dramatically shortening the data path and enabling a 2,048-bit interface that delivers bandwidth far beyond what conventional DDR or GDDR memory can reach.

That bandwidth is not incidental to AI performance — it is the constraint. As Nvidia's AI factory concept scales from training large models to continuous, real-time inference across thousands of simultaneous requests, the GPU's processing cores can execute faster than memory can supply them with data. HBM's wide data path addresses that bottleneck directly. Each Vera Rubin server system uses multiple terabytes of HBM4; the entire Vera Rubin platform was built around the assumption that HBM4 — not GPU compute — would be the production variable that determined how quickly AI factories could scale.

SK hynix manufactures its HBM stacks using a bonding process called Mass Reflow Molded Underfill, or MR-MUF, which differs from the Thermal Compression Non-Conductive Film method used by Samsung and Micron. The process choice affects yield, thermal performance, and the ability to scale from the current 12-layer stacks to the 16-layer configurations Nvidia has sought for delivery in the second half of 2026.

Nvidia's AI Software Now Designs the Memory Inside Nvidia's Hardware

The most technically novel element of the partnership is a recursive engineering loop embedded in the agreement's terms: SK hynix is using Nvidia's own GPU-accelerated simulation software to design and manufacture the memory chips that ship inside Nvidia's products.

The specific tools are Nvidia's CUDA-X libraries, its PhysicsNeMo framework, and its cuLitho computational lithography library. At SK hynix, the PhysicsNeMo team has built Graph Neural Network-based surrogate models — including architectures called Graph Network-based Simulator and MeshGraphNet — that predict the outcome of critical manufacturing steps like plasma etching without running the physical process on silicon wafers. TCAD simulations that previously required hours or, in some cases, several weeks of compute time now complete in milliseconds using these AI surrogate models. The reduction changes the economics of chip development by allowing engineers to explore far larger design spaces before committing to physical wafer runs.

Computational lithography carries a similar story. A typical mask set for a semiconductor chip — one of the final pre-production steps in which circuit patterns are rendered onto silicon — can require 30 million or more CPU-hours of computation. Nvidia's cuLitho GPU-accelerated lithography production, running on GPU clusters, collapses that figure: 350 H100 systems can replace what would otherwise require approximately 40,000 CPU systems. SK hynix has extended this GPU-accelerated approach to its own fab workflows, applying it to both TCAD and lithography simulation.

The partnership also includes a third engineering element: SK hynix is building digital twins of its fabrication plants using Nvidia Omniverse and OpenUSD, with the goal of enabling fully autonomous fab operations by 2030. Rather than testing a new process change on a live production line — where a single mistake can halt wafer output and cost chipmakers millions per day — engineers will simulate and verify process changes inside the digital twin first.

SK Telecom, which manages a 60-person physical AI unit that emerged from a former metaverse team, served as the integration partner for SK hynix's initial digital twin deployment. The two companies completed a proof-of-concept in 2025 confirming the platform could handle the complexity of a working semiconductor fab. Phased commercialization is now underway at SK hynix, with plans to extend the platform to other industries once that deployment reaches commercial scale.

What Is the Nvidia DSX Platform: How SK Telecom Builds AI Factories

The SK Telecom side of the June 7 announcement introduced a separate but related agreement: SK Telecom will build AI factories — data centers purpose-built to generate AI outputs at industrial scale, optimized for token throughput per megawatt of power rather than general-purpose computing — using Nvidia's DSX full-stack reference architecture.

DSX is Nvidia's end-to-end blueprint for AI factory construction and operation. It combines Nvidia accelerated computing hardware, the DSX MaxLPS software layer (which maximizes token performance per megawatt), the DSX operating system (which handles lifecycle management, runtime consistency, health automation, and multi-tenant operations), and Nvidia's partner ecosystem for facilities and cooling. An operator building on DSX receives a pre-validated architecture for AI training, inference, and agentic workloads rather than assembling a bespoke system from components. The design criterion is tokens per megawatt — a metric that treats power as the primary production constraint, consistent with Jensen Huang's framing of AI factories as industrial facilities where revenue is limited by available electricity.

The first SK Telecom AI factory is planned to come online in Korea in 2027. This initial facility will not operate at gigawatt scale; SK Telecom clarified on June 8 that the gigawatt target describes a long-term, phased expansion goal as the deployment proves its operating model and spreads across Asia. The partnership builds on a prior agreement between SK Telecom and Nvidia to construct an AI factory powered by more than 50,000 Nvidia GPUs. SK Telecom has also joined Nvidia's global AI-cloud ecosystem program, Nvidia Cloud Partner, giving it access to Nvidia's partner network for enterprise AI cloud services.

Single-Vendor Memory Bet: SK hynix Holds 60–70% of Vera Rubin HBM4 Allocation

Supply-chain analysts have flagged the concentration embedded in this relationship. SK hynix accounts for an estimated 60 to 70 percent of the HBM4 volume for Nvidia's Vera Rubin platform — an allocation that reflects a head start in qualification, manufacturing yield leadership, and SK hynix's MR-MUF bonding advantage, not an exclusive contract. Samsung and Micron have both been certified as Vera Rubin HBM4 suppliers and are competing aggressively for allocation in next-generation configurations.

The co-development partnership announced June 7 deepens SK hynix's embedded position across all four future Nvidia platforms rather than just Vera Rubin, but it does not establish exclusivity. What it does establish is a shared R&D roadmap — meaning SK hynix and Nvidia will align their engineering schedules years in advance, giving SK hynix lead time to develop memory architectures before the product requirements are published to the broader market. That advance alignment is what Samsung and Micron do not currently hold in the same form.

Goldman Sachs projected that SK hynix will maintain more than 50 percent of the total HBM market through at least 2026, supported by its first-mover advantage in HBM4 and its deep customer alignment with Nvidia. The global HBM market outlook shows TrendForce forecasting the market to reach $58 billion in 2026, up from $38 billion in 2025, driven almost entirely by AI infrastructure demand.

Jensen Huang: "Just Standing at the Starting Line"

Huang, speaking at the June 8 press briefing in Seoul, characterized Nvidia's current market position in AI infrastructure as early rather than mature. He compared AI's eventual role in the global economy to the way the internet became universal infrastructure — and noted that South Korea still has relatively little AI infrastructure relative to the expected long-term demand. Asked about the recent decline in chip stocks including SK hynix, he dismissed the concern, telling reporters that lower share prices simply meant the opportunity to buy them had become more accessible, and that the industry was "just standing at the starting line."

That framing is consistent with the structural logic of the SK Group alliance. The partnership is not designed to address today's supply shortfall — SK hynix's entire HBM production capacity through 2026 was already sold out before the announcement was made. It is designed to align the companies' engineering and manufacturing roadmaps for the product generations that will ship in 2027, 2028, and beyond — the years in which AI factories are expected to scale from today's tens of thousands of GPUs per site toward the gigawatt-class installations that both Nvidia and SK Group have described as their long-term infrastructure target.

Frequently Asked Questions

What did Nvidia and SK Group announce on June 7 and June 8, 2026?

Nvidia and SK Group announced a group-level alliance covering three separate commitments: SK hynix will co-develop High Bandwidth Memory 4 across all four of Nvidia's next-generation product lines — Vera Rubin supercomputers, Vera CPUs, RTX Spark personal AI computers, and Jetson Thor robotics platforms; SK hynix will use Nvidia's GPU-accelerated TCAD and computational lithography software to speed its chip design and manufacturing pipeline; and SK Telecom will build AI factories in Korea using Nvidia's DSX platform, with the first facility planned for 2027, eventually scaling toward gigawatt-class capacity across Asia.

How does SK hynix use Nvidia's AI tools to design chips?

SK hynix's TCAD Intelligence team uses Nvidia's PhysicsNeMo framework to build Graph Neural Network surrogate models that simulate critical manufacturing steps — such as plasma etching — without running the physical process on silicon wafers. These AI models reduce simulation times from hours or weeks to milliseconds, allowing engineers to explore far more design configurations before committing to physical production. SK hynix also applies Nvidia's CUDA-X libraries and cuLitho computational lithography software to its manufacturing workflows, using GPU clusters to compress mask-set computation that would otherwise require tens of millions of CPU-hours.

Is SK hynix Nvidia's only HBM supplier?

No. Samsung Electronics and Micron Technology are both qualified and in active production as HBM4 suppliers for Nvidia's Vera Rubin platform, with first shipments scheduled for Q3 2026. Supply-chain analysts estimate SK hynix holds approximately 60 to 70 percent of the Vera Rubin HBM4 allocation, with Samsung and Micron supplying the remainder. The June 7 co-development agreement does not create an exclusive arrangement; it gives SK hynix a shared R&D roadmap with Nvidia across four future product lines, an advantage Samsung and Micron do not currently hold in the same form.

What is the Nvidia DSX platform?

DSX is Nvidia's full-stack reference architecture for building and operating AI factories — data centers optimized to produce AI outputs, measured in tokens, at the lowest cost per megawatt of power. It combines Nvidia hardware, the DSX MaxLPS software layer for token throughput optimization, and the DSX operating system for factory lifecycle management, health monitoring, and multi-tenant operations. SK Telecom will use DSX as the architectural blueprint for its AI factory buildout in Korea beginning in 2027.