Nvidia and SK Group put $500 billion-plus behind AI's next memory cycle

Nvidia and South Korea's SK Group announced more than $500 billion in AI data centers and a long-term high-bandwidth memory deal on Friday. The piece that locks in first is the memory supply, not the compute.

SK Hynix is the only supplier that has completed development of HBM4, the next generation of the stacked DRAM that feeds Nvidia's Vera Rubin accelerators. A 2-gigawatt SK Telecom facility, due online in 2027, will run on HBM4 from the start and serve as the first concrete artifact of the bet. "Gigawatt" is the unit hyperscalers now use for next-generation AI campuses; a single 2-gigawatt site can draw as much power as a mid-size American city at peak load, and it is the order of magnitude OpenAI's Stargate and similar U.S. programs have been quoting.

The Reuters wire carrying the announcement frames the package as "more than $500 billion" across AI data centers and next-generation memory, with SK Telecom's data center as the flagship site. The figure is the aggregate program scale Nvidia and SK Group have signaled, not a single signed check or a confirmed 2026 capex line. Read it as a headline-scale anchor for the industrial bet, not a wire-ready itemized spend.

The SK Telecom facility is the deal's most concrete artifact. Per Korea JoongAng Daily, the buildout is phased: a 55-megawatt pilot in the first half of 2027, expansion to 100 megawatts within 2027, 200 megawatts by 2028, and a path to 2-gigawatt ultimate scale. The 55-megawatt H1 2027 milestone is the first date worth tracking; if it slips, the 2-gigawatt endpoint slips with it.

The chip side is the second leg. SK Telecom's facility will run Nvidia's Vera Rubin platform, the successor to the current Blackwell generation. Nvidia's newsroom release describes the partnership as covering AI training, AI agents, and "physical AI" applications. The last category covers robotics and autonomous systems, where Nvidia's Jetson Thor platform is positioned. The joint Hynix-Nvidia work also extends to the Vera CPU and the RTX Spark desktop AI PC, per SK Hynix's multi-year partnership release. Memory is the through-line: every product on that list depends on HBM or next-generation stacked DRAM.

The memory piece is also the part of the deal that is most concretely delivered. SK Hynix announced earlier this month that it had completed the world's first HBM4 development and was readying mass production. HBM4 roughly doubles the bandwidth of HBM3E, and Nvidia's Vera Rubin platform is designed around that bandwidth envelope. Locking in Hynix capacity now is the kind of vertical supply move AI labs and hyperscalers have been making for accelerators; this one is for the memory underneath them.

A second track sits beside the SK Telecom project. Nvidia, Naver, and Brookfield plan to expand Naver's existing AI data center in South Korea, per Channel News Asia. Naver is South Korea's largest internet company and runs its own foundation-model effort; Brookfield is the Canadian asset manager that has been one of the most aggressive financial backers of new AI compute infrastructure globally. The two deals (SK Telecom as the greenfield flagship, Naver as the brownfield expansion) give Nvidia two on-ramps into the Korean market and give SK Group's affiliates a multi-year demand floor for HBM4.

Power, permitting, and grid interconnection in South Korea are not solved problems at gigawatt scale. And the more-than-$500-billion figure is a public commitment, not a signed capex schedule. Investors and the Korean government will be reading the next SK Hynix and SK Telecom earnings for the line items that back the headline.

The 55-megawatt pilot phase is targeted for the first half of 2027. The HBM4 supply behind it is being locked in well before then.