Nvidia and South Korean memory manufacturer SK Hynix have signed a multi-year technology agreement that goes far beyond conventional supply contracts. The two companies will jointly develop the next generations of memory for so-called AI factories—massive data centers built exclusively for artificial intelligence. And it does not end there. SK Hynix will deploy Nvidia’s tools directly in its factories, where artificial intelligence is expected to help design and manufacture the chips themselves.
Nvidia CEO Jensen Huang announced the agreement during a visit to South Korea, where he met with executives from SK Group and other local companies. “SK Hynix has been Nvidia’s largest memory partner. And it will remain our largest partner,” he said after meeting SK Group Chairman Chey Tae-won. The agreement runs for more than two years, with the option of further extensions. And how large is the deal? Huang said Nvidia already buys billions of dollars’ worth of memory from SK Hynix each year, and that volume is expected to grow substantially.
Memory is crucial for AI
You may wonder why a company that makes the world’s most powerful AI chips is so intensely focused on memory. The answer lies in the fundamental physics of computers. A graphics chip can perform billions of operations per second, but if data does not reach it quickly enough, it simply sits idle and waits. This problem is known as the memory wall, and experts first described it in 1994. Since then, the gap between computing speed and memory speed has only continued to widen.
The issue is even more pronounced with large language models. When generating each new word, the model must reload data about all the preceding words in the context. In long conversations, this means reading gigabytes of data for a single generated word. A faster chip with the same memory therefore will not help. The bottleneck is not computation, but data transfer.
The solution is HBM, or high-bandwidth memory. Manufacturers stack it in layers directly beside the graphics chip and connect it using thousands of tiny interconnects. While a standard computer memory channel is 64 bits wide, a single HBM4 memory stack offers 2,048 bits. The result is bandwidth measured in terabytes per second. But manufacturing such memory is extremely complex. It requires roughly twenty additional production steps compared with conventional memory, and each gigabyte consumes about three times as many silicon wafers.
Manufacturing from PCs through AI to robots
The agreement covers Nvidia’s entire future product portfolio. SK Hynix will develop custom memory for Vera Rubin supercomputers, Vera processors, RTX Spark personal AI computers, and Jetson Thor robotics platforms. The collaboration therefore spans data centers, conventional computers, and robots.
The other side of the agreement is also noteworthy. SK Hynix will begin using Nvidia’s software tools in its factories. The CUDA-X libraries and PhysicsNeMo system are intended to accelerate semiconductor simulations and the design of new chips. The Korean company is also building digital twins of its factories using the Omniverse platform. The goal is fully autonomous operation of production facilities, where artificial intelligence controls robot movement, optimizes manufacturing, and makes decisions based on operational data. This creates an unusual loop: AI helps manufacture the memory that then powers more AI.
“AI factories are the engines of the next industrial era, and advanced memory is essential to their performance,” Huang said in an official statement. Chairman Chey added that the two companies had been working toward such cooperation for years and that the agreement merely reflects its depth.
Korea is expanding, but even that is not enough
SK Hynix currently produces roughly 550,000 memory wafers per month and wants to double its capacity to one million by 2030. This is being driven by enormous demand and investment. The M15X factory in Cheongju will ramp up to 40,000 wafers per month in the second half of 2026 and double that volume by 2027. The company has also ordered approximately $8.6 billion worth of EUV lithography machines from the Dutch company ASML. In the long term, its backbone will be the Yongin complex, whose first building is expected to add 360,000 wafers per month by the first half of 2030.
And Huang’s reaction to these plans? In his view, even doubling capacity will not be enough to meet growing demand for AI. That says a great deal about the scale of the world’s coming appetite for computing power.
Unfortunately, ordinary users will feel the impact as well. Because HBM production consumes three times as many wafers per gigabyte and manufacturers are prioritizing it, less capacity remains for RAM used in phones, laptops, and servers. Prices for conventional memory are therefore rising, and according to available analyses, manufacturers will retain their pricing power for the rest of the decade.
The competition is not standing still
SK Hynix is estimated to account for 60 to 70 percent of HBM4 memory supplies for the Vera Rubin platform. But it must defend its position. Samsung manufactures the memory stack’s control chip directly on an advanced 4-nanometer process and achieves a speed of 11.7 gigabits per second, 46 percent above the standard. Meanwhile, SK Hynix depends on Taiwan’s TSMC to manufacture its logic chips, and TSMC is already operating at the limit of its capacity. On top of that, according to market reports, Nvidia is developing its own base dies for memory stacks, which would reduce its dependence on both companies.
The agreement with SK Hynix was only one part of Huang’s trip to Korea. SK Telecom will use Nvidia technologies to build a gigawatt-scale cloud AI center, with the first phase launching in 2027. Internet company Naver and the Doosan Group will also build data centers. Nvidia will collaborate with LG on humanoid robots and with automaker Hyundai on autonomous driving, robotics, and AI-driven manufacturing. The Korean government also plans to purchase more than 9,700 graphics chips for a state AI project, including over two thousand Vera Rubin units.
Nevertheless, the news failed to impress investors. South Korea’s Kospi index, which had doubled in value over six months thanks to the wave of interest in AI, fell 8.3 percent on the day of the announcement. SK Hynix shares lost 7.7 percent, while Samsung fell by more than ten percent. However, the agreement itself was not to blame, but rather strong U.S. labor market data that raised concerns about interest rate hikes. Huang commented on it himself: anyone who wanted the shares can now buy them more cheaply.
Sources: medium.com, tomshardware.com and bloomberg.com
