On June 7, 2026, in a move that tightens the semiconductor industry's grip on the future of artificial intelligence, NVIDIA and SK hynix announced a multiyear technology partnership to co-develop next-generation memory for AI factories. The agreement, struck in Seoul, promises to customize high-bandwidth memory (HBM) for NVIDIA's most ambitious computing platforms yet — including Vera Rubin systems, Vera CPUs, and RTX Spark PCs. The deal effectively locks SK hynix into the role of premier memory architect for NVIDIA's sprawling AI infrastructure vision, one that stretches from hyperscale data centers to the edge and personal computing.

For an industry already grappling with an insatiable hunger for memory bandwidth, the partnership sends a clear signal: the brute-force scaling of HBM stacks is no longer enough. Instead, NVIDIA and SK hynix intend to co-engineer memory from the ground up, baking in optimizations that align with the specific dataflow patterns of next-generation GPUs, CPUs, and AI-accelerated PCs. The result could be a new class of HBM designed not just for speed, but for the massive concurrency and low-latency demands of trillion-parameter models running across thousands of nodes.

The AI Factory Needs a New Memory Blueprint

AI factories — a term NVIDIA CEO Jensen Huang has used to describe the purpose-built data centers that will generate intelligence as today's factories produce goods — are fundamentally memory-bound. Training a single large language model can require terabytes of HBM, and inference at scale demands both capacity and throughput that outpace current DRAM technologies. HBM, with its 3D-stacked design and wide interface, has become the default high-speed buffer between compute and storage. Yet every new GPU generation pushes the limits: HBM3e, the current industry standard, already delivers over 1.2 TB/s per stack, but Blackwell Ultra and its planned successors will demand even more.

SK hynix, the world's leading HBM supplier, has been at the center of this escalation. Its HBM3e memory is inside NVIDIA's H200 and B200 GPUs, and the company has already demonstrated HBM4 prototypes with data rates exceeding 6.4 Gbps per pin. But the June 7 announcement signals a departure from the traditional vendor-buyer relationship. Instead of simply shipping standard HBM modules, SK hynix will collaborate with NVIDIA's architects during the silicon definition phase, much as TSMC does for process technology. The co-development model aims to eliminate the last-mile inefficiencies that arise when memory designers and processor designers work in isolation.

Vera Rubin: The First Co-Developed Platform

The partnership explicitly names Vera Rubin systems — a previously rumored codename for NVIDIA's post-Blackwell GPU architecture — as a primary target. Although NVIDIA has not yet formally unveiled Vera Rubin, industry analysts widely expect it to power the company's next-generation DGX and HGX platforms, likely arriving around 2027. By involving SK hynix this early, NVIDIA ensures that Vera Rubin's memory subsystem will be tailored to the GPU's compute partitions, cache hierarchies, and NVLink topologies.

This matters because the gap between compute throughput and memory bandwidth has become the defining bottleneck of AI scaling. Even with HBM3e, GPUs often sit idle waiting for data. Custom HBM for Vera Rubin could mean wider I/O, higher capacities per stack, or even on-package memory that blurs the line between cache and main memory. The announcement also references
Vera CPUs, suggesting that NVIDIA's Arm-based processor line — currently Grace — will receive the same co-design treatment. A CPU with HBM-class bandwidth could dramatically accelerate vector databases, recommendation systems, and in-memory analytics, all critical for AI factories.

RTX Spark PCs: Bringing AI Factories to the Desktop

Perhaps the most intriguing flank of the deal is its mention of RTX Spark PCs. While NVIDIA has been steadily infusing its client GPUs with AI capabilities — from DLSS to Tensor cores — a co-developed HBM memory solution for desktop systems hints at a much more radical shift. An RTX Spark PC equipped with custom HBM could act as a local AI factory, training or fine-tuning models without the latency and cost of cloud round-trips.

Such a machine would represent a new category: a workstation-class device that merges the real-time rendering prowess of GeForce with the memory bandwidth of a data center GPU. For developers, researchers, and creators, it could shrink the feedback loop on AI experimentation from hours to minutes. The partnership with SK hynix suggests this is not a distant concept; memory co-development timelines typically run three to four years, so RTX Spark PCs may already be in advanced prototyping.

Redrawing the HBM Supply Chain

From a supply-chain perspective, the agreement cements SK hynix's dominance over Samsung and Micron in the AI memory segment. SK hynix already commands over 50% of the HBM market, and by locking in a multiyear roadmap with the largest AI chipmaker, it gains visibility and volume that competitors cannot easily match. The deal also likely includes capacity reservations and joint investment in advanced packaging facilities, crucial for the 2.5D and 3D integration that HBM requires.

For NVIDIA, the partnership reduces the risk of memory shortages that have historically constrained GPU availability. Custom HBM further erects a barrier to competitors seeking to replicate NVIDIA's AI-factory stack. While AMD and Intel also use HBM, they rely on more generic memory products; a custom SK hynix-NVIDIA design would make it harder for rivals to achieve parity through off-the-shelf parts.

Implications for the Windows Ecosystem

Although the announcement focuses on AI factories, the inclusion of RTX Spark PCs holds particular significance for the Windows ecosystem. Microsoft has been aggressively embedding AI into Windows through Copilot and the Windows AI Studio, but the hardware has lagged behind the cloud. A PC with HBM-class bandwidth and a deeply integrated AI stack — spanning the Vera CPU, a powerful GPU, and co-developed memory — could make local AI assistants, generative design, and real-time translation truly seamless.

It also raises questions about form factors. HBM generates significant heat and requires complex interposers; putting such a package into a traditional desktop chassis will challenge thermal engineers. But NVIDIA's quiet mention of the platform suggests the company has already solved many of these problems, possibly through advanced liquid cooling or new motherboard designs that treat memory and compute as a single, tightly coupled module.

The Road Ahead

The multiyear nature of the deal implies that the first fruits of the collaboration will not appear until 2027 at the earliest, aligning with the expected launch of Vera Rubin systems. In the meantime, both companies must navigate geopolitical headwinds: SK hynix's advanced fabs are in South Korea, while export controls on AI technology continue to tighten. Co-developing memory may also require joint R&D centers and personnel exchanges, which could draw regulatory scrutiny.

Still, for an industry that measures progress in percentage points of bandwidth improvement, the promise of from-scratch memory co-design is transformative. It acknowledges that the next leap in AI performance will not come from transistor shrinks alone, but from rethinking the entire compute-memory stack. As AI factories become the engine rooms of the global economy, the NVIDIA–SK hynix partnership may well be remembered as the moment when memory stopped being a commodity and became a core strategic weapon.