NVIDIA and SK Group propose a $500 billion plan: a 2GW AI data center to launch in 2027, targeting HBM4 supply

NVIDIA and Korea’s SK Group threw out an AI plan totaling more than $500B on Friday (7/24). The scope spans large-scale AI data centers and next-generation memory. SK Telecom will build an AI data center with an installed capacity of 2GW, equipped with NVIDIA’s Vera Rubin chips and SK hynix’s HBM4. The first facility is expected to begin operations in 2027. Jensen Huang also used this stop in San Francisco to meet with South Korea’s President Lee Jae-myung. Naver and Brookfield were also pulled into the expansion list.
(Backgrounder: Samsung and SK hynix shell out 2,000 trillion won to build a wafer fab! A bet on memory for the next 10 years—Gwangju becomes the biggest winner)
(Additional context: HBM memory already accounts for 63% of the cost of AI chips; SK hynix, Samsung, and Micron stand to benefit from compute-pricing power)

Table of contents

Toggle

• This contract is buying memory
• SK hynix takes most of Vera Rubin’s HBM4
• The displacement effect has already reached players’ graphics cards

Key takeaways

• NVIDIA and SK Group announce an AI plan totaling more than $500B, covering data centers and next-gen memory
• SK Telecom’s 2GW AI data center features Vera Rubin and HBM4, with the first facility to go live in 2027
• SK hynix estimates it will take 60% to 70% of the HBM4 needed for Vera Rubin, ahead of Samsung and Micron

$500B, 2GW of power, and starting in 2027. This checklist unveiled by NVIDIA (NVIDIA) and Korea’s SK Group on Friday (7/24) doesn’t really look like a typical procurement contract anymore—it looks more like a state-level infrastructure budget. The plan spans two things: first, building large-scale AI data centers in South Korea; second, securing next-gen memory supply in advance.

The division of labor goes like this: SK hynix and NVIDIA sign a long-term partnership to lock in next-gen memory supply for NVIDIA, and they jointly develop HBM (high-bandwidth memory), with clear use cases aimed at three areas: AI training, AI agents, and physical AI. SK Telecom handles the building of the server rooms. The first AI data center, with an installed capacity of 2GW, will be equipped with NVIDIA’s Vera Rubin chips and SK hynix’s HBM4, and is scheduled to launch in 2027. NVIDIA is also bringing in Naver and Canadian asset-management giant Brookfield to expand Naver’s existing AI data centers in Korea.

The announcement took place when Jensen Huang (Jensen Huang) released the figure while meeting with South Korea’s President Lee Jae-myung in San Francisco. That day was the first day of Lee Jae-myung’s two-day investment promotion schedule, and he also announced that he would develop a self-driving version of a Genesis sedan together with the Hyundai Motor Group.

A head of state personally flew to the Silicon Valley stage to put his stamp on it—this level of execution is no longer a business-to-business negotiation.

This contract is buying memory

$50 billion sounds like buying land and building a house, but the real focus is hidden in the memory part. Cointelegraph previously reported that HBM currently accounts for about 63% of the AI chip cost. That means that when NVIDIA sells each flagship chip, more than 60% of the materials cost goes to memory makers. In the past, what NVIDIA feared most was that TSMC’s advanced packaging might not get queued in time; now its biggest worry has shifted to HBM supply not being able to keep up.

So rather than being “an investment,” this long-term deal is more like a supply insurance policy. NVIDIA uses long-term commitments to exchange for SK hynix’s production prioritization, locking in the memory shortfalls for the coming years.

The two sides actually signed a multi-year technology agreement back on June 7. The scope isn’t limited to HBM; it also includes using CUDA-X to accelerate chip-design simulation, and using Omniverse as a digital twin for wafer fabs.

At the time, Jensen Huang put it plainly: “AI factories are the engine of the next industrial revolution, and advanced memory is crucial to their performance.” SK Group chairman Chey Tae-won said the two sides are jointly developing next-gen memory for AI factories while also bringing AI into the design and manufacturing processes of semiconductors.

SK hynix takes most of Vera Rubin’s HBM4

In this memory race, SK hynix is clearly in the lead. Market estimates say it will secure 60% to 70% of the HBM4 supply needed for Vera Rubin, leaving only about one-third for Samsung and Micron. Jensen Huang previously said NVIDIA would be the first customer for HBM4, and the bulk of those orders has already been locked into SK hynix’s production lines.

The company’s profile is changing too. SK hynix listed on Nasdaq in July, raising $26.5 billion, the second-largest equity offering in history. When Chey Tae-won was asked about the stock price after listing, he gave a very confident answer. He said memory demand will always be there; as time horizons stretch, the stock price will naturally move upward, advising investors not to trade short-term.

The displacement effect has already burned through to players’ graphics cards

When capacity is tied up by AI, the cost will show up elsewhere. NVIDIA only this week issued a new round of price increase notices to board makers. With GDDR6 and GDDR7 memory switching to an allocation system, the market price of graphics cards has been pushed up another notch. Samsung has also raised DRAM quotes multiple times over the past year. When it gets more expensive to build a gaming PC, the root cause is the capacity allocations locked into those multi-hundred-billion-dollar contracts.

SK Telecom’s appetite isn’t limited to 2GW. It already disclosed a 15GW long-term blueprint in early July. Phase one is building clusters of more than 2GW centered on Ulsan and deployed in the Gyeongsang region. Phase two adds another 1GW in the Jeolla region. Domestically, the 5GW buildout will be phased in starting in 2029. A single 1GW AI data center has engineering costs of roughly 70 trillion won on its own. This scale is something no single company can carry alone. SKT also lists its funding sources in four parts: in-house investment, strategic partners, long-term contract customers, and project financing.

Frequently asked questions

Why did NVIDIA sign a long-term memory contract with SK hynix?

HBM already accounts for about 63% of the cost of AI chips, and it’s also the easiest point in the Vera Rubin mass-production process to get bottlenecked. NVIDIA uses long-term commitments to obtain SK hynix’s production prioritization—effectively buying supply insurance for the next few years in advance, preventing memory shortages from slowing down shipment schedules.

When will SK Telecom’s 2GW AI data center be completed?

The first facility is expected to begin operations in 2027. That is phase one of SK Telecom’s 15GW blueprint, deploying clusters of more than 2GW centered on Ulsan in the Gyeongsang region. Phase two is planned to build another 1GW in the Jeolla region, and the domestic 5GW will be rolled out in phases starting in 2029.

This article references Reuters, NVIDIA official press releases, and SK Telecom announcements, and is compiled and reported by Dongqu Dongqu.