Analysis
Nvidia and South Korea's SK Group unveiled a more than $500 billion AI initiative on Friday, bundling large-scale AI data centers with a next-generation memory partnership in one of the largest single commitments to AI infrastructure ever made public. The centerpiece is a 2-gigawatt AI data center that SK Telecom will build, powered by Nvidia's next-generation Vera Rubin chip platform and SK Hynix's HBM4 high-bandwidth memory, with the first facility due to come online in 2027. Nvidia separately confirmed it will invest $1 billion directly into Naver, and that Nvidia, Naver and Brookfield plan to jointly expand Naver's existing AI data center footprint in South Korea.
The announcement lands just one day after Samsung and SK Hynix disclosed a related but far vaguer set of megadeals -- large, undisclosed-value supply and investment agreements for high-bandwidth memory, timed to South Korean President Yoon Suk-yeol's Silicon Valley AI summit visit that drew Nvidia CEO Jensen Huang and the CEOs of OpenAI, Anthropic and Broadcom into the same room. Friday's Nvidia-SK Group announcement is the concrete follow-through: where Wednesday's reporting described "very large" agreements without dollar figures, this initiative puts an actual number -- more than $500 billion -- on the combined commitment, alongside specific chip platforms, memory generations, and a facility timeline.
The deal formalizes South Korea's position at the center of the AI memory supply chain. SK Hynix and Samsung together control the overwhelming majority of global HBM production, the specialized memory stacks that feed data to GPUs fast enough to keep pace with AI training and inference. HBM has become one of the most supply-constrained components in the AI buildout over the past year, alongside advanced packaging capacity at TSMC and Amkor and raw GPU output from Nvidia itself -- meaning memory is now as much a bottleneck on AI infrastructure scaling as compute.
“The deal formalizes South Korea's position at the center of the AI memory supply chain.”
The Naver piece of the deal is notable for a different reason: it's Nvidia taking a direct equity stake in a cloud and AI platform company rather than simply selling it chips, a strategy Nvidia has increasingly used with AI labs and infrastructure players (echoing its investments in xAI, CoreWeave, and others) to lock in demand for its own hardware roadmap years out. Naver, sometimes described as South Korea's Google, has been building out GPU-anchored data centers domestically; Brookfield's involvement adds an institutional infrastructure investor with deep experience financing capital-intensive, multi-decade projects.
The timing compounds an extraordinary week for AI infrastructure commitments globally. It follows OpenAI's disclosure of a self-built, $30 billion-plus, 3.2-gigawatt Georgia data center campus, Nvidia's own $1.5 billion prepayment to Amkor for US advanced packaging capacity, and Intel's decision to raise 2026 capex guidance past $20 billion -- a stretch of days in which compute, power, packaging and memory were all locked down simultaneously by different players across the AI stack, rather than any single bottleneck getting solved before the next one emerged.
For venture investors, the read-through is that memory and packaging are becoming durable, investable categories in their own right, not commodity inputs -- SK Hynix and Samsung's newfound pricing leverage on HBM should show up as a rising cost line in every AI infrastructure buyer's model, from hyperscalers down to GPU-cloud startups. The bear case: a headline "$500 billion-plus" figure spanning data centers, chip supply, and equity investment across multiple entities and years is not a single check being written tomorrow, and Nvidia's SK Group partnerships (like its OpenAI and xAI commitments) tend to be structured as long-horizon capacity and supply commitments rather than immediate cash outlays.
Watch whether SK Telecom's 2027 online date holds as advanced packaging and power constraints continue to bite across the industry, whether Samsung strikes a comparable Nvidia partnership of its own to avoid ceding ground to SK Hynix, and whether Naver's data center buildout becomes a template other regional cloud players -- in Japan, India or the Gulf -- try to replicate with their own Nvidia equity partnerships.
