How NVIDIA Is Locking in Its Memory Chip Suppliers
Two Deals, One Strategic Throughline
NVIDIA has announced separate partnerships with SK Hynix and Samsung — the two biggest makers of high-bandwidth memory (HBM), the specialized chips that feed data to AI accelerators at high speed. Unlike typical supplier contracts, these deals involve the Korean chipmakers working directly with NVIDIA on product design, manufacturing improvements, and chip architecture.
Both partnerships are significant. With SK Hynix, NVIDIA is co-developing memory for four of its platforms: Vera Rubin (a data-center AI computer), the Vera CPU, RTX Spark PCs, and Jetson Thor (a robot computing platform). With Samsung, the focus is different: NVIDIA is helping Samsung build an "AI Megafactory" — a manufacturing facility that will use more than 50,000 NVIDIA GPUs to run everything from chip design to quality checks across Samsung's entire semiconductor operation. Source: NVIDIA Newsroom | Source: Samsung Newsroom
SK Hynix: From Vendor to Partner
Historically, NVIDIA would specify what it needed in memory chips, and suppliers would deliver. The SK Hynix deal changes that: the two companies are now designing memory together from the start.
When a customer sits down with a supplier at the design table — not after design is locked — the resulting chip fits together much better. SK Hynix engineers working on Vera Rubin's memory architecture can optimize for speed and responsiveness alongside NVIDIA's system designers, rather than trying to patch things together afterward. That matters because memory speed and latency (the time it takes to retrieve data) drive how fast AI systems run.
SK Hynix is also using NVIDIA's GPU-powered software tools to speed up its own chip simulation work. Two areas stand out: TCAD (a process that models how electricity moves through a transistor before it's actually manufactured) and computational lithography (the math that corrects photomask patterns so they print correctly on silicon). Simulating these steps at advanced nodes can take months and cost hundreds of millions of dollars per step. Running the simulation on GPUs instead of traditional computers could compress that timeline significantly. Source: NVIDIA Newsroom
SK Hynix got a head start: it began mass production of its latest-generation HBM chips in March 2024, giving it early advantage as NVIDIA ramped production of its H100 and H200 chips. The new deal formalizes what was already a productive relationship. Source: Reuters
Samsung: AI on the Factory Floor
Samsung's deal is a different animal. Rather than co-developing memory architecture, Samsung is using NVIDIA's technology to rethink how it manufactures chips.
Installing 50,000 GPUs across Samsung's fabs (semiconductor factories) is a scale of commitment usually seen when cloud giants like Amazon or Google build data centers, not at chip makers. Samsung is pairing the GPUs with NVIDIA Omniverse — software that creates a digital twin, or accurate virtual model, of the factory floor. This virtual factory can simulate equipment, process flows, and logistics in real time, letting engineers test changes in software before running them on actual machines. Source: Samsung Newsroom
Separately, Samsung is pursuing its own memory innovation: HBM-PIM, or processing-in-memory. The idea is simple but powerful: instead of shuttling data back and forth between memory and processor (which costs energy), embed computing power directly in the memory stack. Samsung is also working on CXL-PNM (a standard for connecting memory and processors more efficiently) to build what it calls an HBM-PIM Cluster — a memory system optimized for running large AI inference (using trained models to make predictions). Source: Samsung Semiconductor
Samsung planned to begin producing HBM4 (the next generation after HBM3E) in February 2026, targeting NVIDIA. This is a milestone because Samsung lagged SK Hynix in qualifying HBM3E — meaning it took longer to prove its chips worked at scale. Source: Reuters
Why Both Deals at Once
NVIDIA is not doing these partnerships by coincidence. Every Blackwell GPU NVIDIA ships includes HBM3E memory, and the upcoming Vera Rubin will require HBM4 at large scale. Depending on one supplier for that much memory — particularly when both SK Hynix and Samsung are based in South Korea — creates two risks: a business risk (a single company's problems disrupt you) and a geopolitical risk (political tensions could limit access).
NVIDIA learned from how Apple approached this problem in the early 2010s. Apple deliberately bought DRAM and NAND flash memory from Samsung, SK Hynix, and Micron — spreading orders across multiple suppliers. When shortages hit the broader market, Apple had buffer stock and other channels. NVIDIA is building a similar hedge, but with a wrinkle: it is embedding its own tools — CUDA-X, Omniverse, GPU-accelerated simulation software — into both SK Hynix and Samsung's R&D and manufacturing. The longer both companies use NVIDIA's software, the harder it becomes to switch away. At the same time, NVIDIA gains inside knowledge of how fast these companies can develop new chips and scale production.
Memory Is Becoming Active, Not Passive
The bigger picture: the industry is rethinking what memory does in an AI system. For years, high-bandwidth memory was treated mainly as a data pipeline — stack more layers, push more gigabytes per second to the chip, call it a day. Both the SK Hynix codevelopment and Samsung's processing-in-memory work signal a shift. Memory is becoming an active partner in computing.
When you embed compute logic in memory — as HBM-PIM does — you can run certain common operations (like summing values or applying activation functions in AI models) right there in the memory stack. The payoff: less data has to move to the main processor and back, which saves enormous amounts of power. In massive data centers running hundreds of AI systems, even a small reduction in data-movement energy compounds across thousands of chips.
What This Means for Competitors
For Micron — the only large non-Korean HBM maker — the picture gets tougher. Shipping memory that meets NVIDIA's specs will be table stakes, but may not be enough if SK Hynix and Samsung are baked into NVIDIA's platform design from day one.
For foundry and packaging companies like TSMC, Samsung's AI Megafactory sends a message: using AI to control your manufacturing process is not just a cost-saving tactic anymore. If Samsung can prove it shrinks production time or improves yield through AI-driven intelligence, it becomes a selling point to customers evaluating which fab to use.
For the broader AI chip supply chain, these announcements mark a turning point. NVIDIA is not simply telling memory suppliers what to build. It is working alongside them to design it, and handing them the tools to manufacture it better.
