SK Group and NVIDIA have announced a partnership worth more than 500 billion dollars to deploy NVIDIA's next generation Vera Rubin infrastructure, built around SK hynix HBM4 memory. It is one of the largest single commitments the AI buildout has produced, and it says something important: the fight has moved on from who has the best GPU to who can actually feed it.
High bandwidth memory sits stacked right next to the processor die, feeding it data far faster than conventional RAM ever could. Every generation of AI accelerator needs more of it, because a chip that can compute at enormous speed is useless if it is waiting on data to arrive. HBM4 is the next step up from HBM3E, and SK hynix is one of only three companies in the world that can make it at volume, alongside Samsung and Micron.
That scarcity is the whole story. NVIDIA can design the fastest chip on paper, but if it cannot secure enough HBM4 to put around it, the chip does not ship in useful numbers. Locking in SK hynix supply at this scale is NVIDIA buying certainty in a market where certainty is now the expensive thing.
This deal does not sit in isolation. Samsung, AMD, Naver and Hyundai are all pouring money into Korean AI infrastructure at the same time, turning the country into one of the central battlegrounds of the AI hardware race rather than a supplier standing on the sidelines. For SK Group specifically, this pact turns memory manufacturing from a commodity business into a strategic one, with the pricing power that comes with being the thing everyone needs and few can supply.
Regular buyers have already felt an early version of this squeeze. Consumer DDR5 prices have climbed sharply through 2026 as memory makers redirect production capacity toward the AI accelerators that pay far more per gigabyte, our earlier look at the RAM price rise covers the numbers. A 500 billion dollar bet on HBM4 supply is the enterprise scale version of the same pressure, and it is not going to ease that squeeze on ordinary hardware any time soon.
The bigger point is what it tells us about where the AI industry's real bottleneck sits. For two years the story was compute: more GPUs, bigger clusters, faster chips. The quieter and arguably more decisive constraint is memory supply, and whoever controls it gets to set the pace for everyone else building on top. A deal this size from NVIDIA is as close to a public admission of that as you will get.
More hardware on Future Technology: AMD's EPYC Venice and Helios rack systems, NVIDIA's AI PC chips, and why RAM keeps getting more expensive.
Get a plain-English tech briefing in your inbox every morning. Join the free Future Technology newsletter.