More than $500 billion is being committed to AI data centers and next-generation memory.
NVIDIA and South Korea’s SK Group announced an initiative that the companies value at over $500 billion. The plan pairs large-scale AI data center builds with a long-term partnership between NVIDIA and SK hynix to secure and co-develop next-generation high-bandwidth memory, including HBM4. SK Telecom plans a 2-gigawatt AI data center using NVIDIA’s Vera Rubin accelerated computing and SK hynix HBM4, with initial operations targeted for 2027.
The initiative also includes an expansion of NAVER’s NVIDIA DSX deployment at the GAK Sejong site from 55 megawatts to 200 megawatts by 2028. NAVER intends to scale to 1 gigawatt of NVIDIA infrastructure for multi-tenant AI cloud services. NVIDIA plans a $1 billion investment in NAVER, and Brookfield has a non-binding term sheet to provide up to $9 billion of project financing. South Korean briefings describe an even broader package of cooperation, with an aggregate value reported near $950 billion.
This is primarily a compute and infrastructure moment. Multi-gigawatt data centers change the math for what you can train and run. High-bandwidth memory, or HBM, is the fast, wide memory that sits close to AI accelerators and determines how quickly models can access the data they need. A long-term supply and co-development deal for HBM4 reduces a key bottleneck: memory bandwidth and availability for large-scale model training.
The industrial scale matters for three reasons. First, the amount of committed power and chip buildout lowers the friction for operators who need sustained, dense compute. Second, vertical agreements between an accelerator vendor and a memory supplier tilt the supply chain toward integrated stacks tuned for specific AI workloads. Third, the financing signals that infrastructure investors view AI compute as a capital asset class, not just a cloud service.
Risks remain. Brookfield’s financing is non-binding, and multi-year timelines to 2027 and 2028 introduce execution risk. Concentrating cutting-edge memory and accelerator capacity in a few partnerships could raise barriers for smaller developers and shift bargaining leverage toward infrastructure owners.
For business leaders the takeaway is plain: access to frontier compute will increasingly depend on industrial partnerships, not only on software or model design. Companies that want to run the largest models or deliver latency-sensitive AI services will have to align with these new, large-scale supply chains.
If these projects proceed on plan, 2027 and 2028 may mark a step-change in the global availability of frontier AI compute and the commercial arrangements that control it.