Verdict: NVIDIA's $500-billion-plus partnership with South Korea's SK Group—announced July 24, 2026—is the largest single AI infrastructure commitment on record, and its real impact on you isn't another data center in Korea. It's a structural lock on the HBM4 memory supply that determines whether GPU rental prices finally fall or stay stubbornly high through 2028. For anyone building AI products on cloud GPUs, the deal signals that memory—not chips—is now the bottleneck, and the companies who control it will shape compute costs for the next three years.
Last verified: 2026-07-30
- SK Group and NVIDIA signed letters of intent for a $500B+ partnership on July 24, 2026. (NVIDIA Newsroom)
- SK Telecom will build a 2-gigawatt AI factory in Korea using NVIDIA Vera Rubin and DSX, with first phase online in 2027.
- NVIDIA and SK Hynix signed a long-term agreement to co-develop HBM4 and future AI memory generations.
- This is a letter of intent, not a binding contract—commercial terms are still pending.
- Separate deals: NVIDIA invested $1B in Naver for cloud expansion; Samsung signed a $200B MOU with Broadcom.
What is the NVIDIA-SK Group $500 billion deal?
The NVIDIA-SK Group partnership is a $500-billion-plus framework covering two pillars: building one of the world's largest AI data centers in South Korea, and securing long-term supply of the high-bandwidth memory (HBM) that NVIDIA's GPUs depend on. The two sides signed letters of intent at the AI Summit in San Francisco on July 24, 2026, formalizing a collaboration that spans AI factory construction, memory co-development, and sovereign AI services across the Asia-Pacific region. (NVIDIA Investor Relations)
The deal is structured as letters of intent, not binding contracts. That means the $500 billion figure represents aggregate planned investment across all partners over multiple years, with firm commercial terms still to be negotiated. The scale is real, but the execution depends on LOI-to-contract conversion in coming quarters. (Reuters)
How does the 2-gigawatt AI factory work?
SK Telecom will build a 2-gigawatt AI factory in Korea running on NVIDIA's DSX full-stack AI factory architecture and powered by the NVIDIA Vera Rubin accelerated computing platform. The facility will use SK Hynix HBM4 memory, with the first phase targeted to come online in 2027. For context, 2 gigawatts is roughly the power consumption of a mid-size city—enough to run thousands of NVL72 rack systems and hundreds of thousands of Rubin GPUs and Vera CPUs. (NVIDIA Newsroom)
The NVIDIA DSX platform integrates accelerated computing, systems, software, and partner technologies into a single data-center-scale platform designed to deliver what NVIDIA calls the "lowest cost per generated AI token." The facility is intended to serve sovereign AI, physical AI, agentic AI, and enterprise AI workloads, with particular focus on Asia-Pacific demand. (Data Center Dynamics)
Why is the SK Hynix memory partnership the part that actually matters for your costs?
The HBM4 memory co-development agreement between NVIDIA and SK Hynix is the load-bearing pillar of this deal for anyone who pays for AI compute. Here's why: HBM (High Bandwidth Memory) is the stacked memory architecture inside AI GPUs that feeds data to compute cores at extreme bandwidth. AI models live in HBM during training and inference. Without enough HBM—and enough HBM bandwidth—the world's most powerful GPUs sit underutilized. (NVIDIA Newsroom, June 2026)
Three suppliers worldwide make HBM at scale: SK Hynix holds approximately 62% global market share, Micron approximately 21%, and Samsung approximately 17%. On NVIDIA's specific HBM4 allocation, SK Hynix has roughly mid-50%, Samsung mid-20%, and Micron about 20%. By deepening its lock on SK Hynix supply, NVIDIA is securing the single most constrained component in the AI hardware stack—but it also concentrates the global AI memory supply chain further. (CNBC)
The critical bottleneck: adding HBM capacity requires new fab construction with 18-to-24-month lead times. HBM demand currently grows roughly twice as fast as HBM supply. That imbalance is what keeps GPU rental prices high, and no partnership—however large—can shortcut the physics of building semiconductor fabs.
| HBM Supplier | Global Share | NVIDIA HBM4 Allocation | Key Detail |
|---|---|---|---|
| SK Hynix | ~62% | ~50-55% | NVIDIA's largest memory partner; HBM4 co-development |
| Micron | ~21% | ~20% | Diversified supplier; not part of SK deal |
| Samsung | ~17% | ~20-25% | Signed separate $200B MOU with Broadcom |
What is NVIDIA Vera Rubin and why does it matter?
NVIDIA Vera Rubin is the successor to the Blackwell platform, announced at CES 2026 and detailed at GTC 2026. The Vera Rubin NVL72 rack contains 72 Rubin GPUs and 36 Vera CPUs connected through NVLink 6, delivering 3.6 EFLOPS of NVFP4 inference and 2.5 EFLOPS of training compute per rack. Each Rubin GPU has 288 GB of HBM4 with 22 TB/s bandwidth and 50 PFLOPS of NVFP4 inference performance—roughly 5x Blackwell's per-GPU inference throughput. NVIDIA claims up to 10x lower cost per token versus Blackwell at the rack level. (NVIDIA Product Page, Hashrate Index)
For the SK Group deal, this matters because the 2-gigawatt factory is one of the first large-scale deployments of Vera Rubin. The facility's thousands of NVL72 rack systems will run on HBM4 supplied by SK Hynix—the same memory that the co-development partnership locks down. The platforms are designed and the memory is sourced from the same ecosystem, which is the structural point of the deal.
| Spec | Blackwell NVL72 | Vera Rubin NVL72 | Improvement |
|---|---|---|---|
| Inference (NVFP4, per GPU) | 10 PFLOPS | 50 PFLOPS | 5x |
| Training (NVFP4, per GPU) | 10 PFLOPS | 35 PFLOPS | 3.5x |
| Per-GPU memory | 192 GB HBM3e | 288 GB HBM4 | 1.5x capacity |
| Memory bandwidth | 8 TB/s | 22 TB/s | 2.75x |
| NVLink bandwidth per GPU | 1.8 TB/s | 3.6 TB/s | 2x |
| Cost per token (inference) | Baseline | 1/10 | 10x lower (NVIDIA claim) |
Will this deal make GPU rentals cheaper?
Probably not before 2027, and possibly not through 2028. The partnership reduces shortage risk and stabilizes supply at the high end, but the underlying imbalance between HBM demand growth (80-100% per year) and supply growth (50-60%) keeps GPU rental prices elevated. The deal makes more GPU capacity available to the Asia-Pacific market, but it doesn't add new HBM capacity any faster—the fabs are already under construction and their timelines are fixed by physics, not by contract announcements.
For builders and small businesses using cloud GPUs, the practical takeaway is to anticipate stable-to-high compute costs through late 2027, with potential relief in 2028 as new HBM4 fabs ramp and Vera Rubin volumes scale. Enterprises that budgeted on the assumption of falling GPU prices through late 2026 should revisit those assumptions. The Vera Rubin platform's claim of 10x lower cost per token is a per-GPU efficiency figure, not a market price figure—NVIDIA sets the architecture, the market sets the rental rate, and the market is still supply-constrained.
What does this mean alongside NVIDIA's other Korea deals?
The SK Group announcement is not an isolated deal—it's the anchor of a coordinated push to make South Korea a global AI infrastructure hub. On the same announcement day, NVIDIA separately revealed a $1 billion investment in Korean cloud provider Naver to expand its data center capacity. Naver will issue approximately 7.24 million new shares to NVIDIA, giving the chipmaker a ~4.5% stake and making it Naver's third-largest shareholder. The Naver-Brookfield-NVIDIA partnership targets a 200-megawatt AI factory at Naver's GAK Sejong data center, scaling to gigawatt-level infrastructure. (UPI, NVIDIA Investor Relations)
Samsung Electronics separately signed a $200 billion memorandum of understanding with Broadcom covering memory chips and foundry services through 2030. Combined with the SK Group deal, South Korean presidential adviser Kim Yong-beom confirmed that U.S. companies account for 80-90% of the demand underpinning South Korea's semiconductor expansion. The signal: Korea's entire semiconductor and cloud ecosystem is being pulled into the global AI build-out, and NVIDIA is the central customer. (Chosun, SE Daily)
| Deal | Parties | Value | Focus |
|---|---|---|---|
| SK Group-NVIDIA | SK Telecom, SK Hynix, NVIDIA | $500B+ | AI factory + HBM4 co-development |
| Naver-NVIDIA | NVIDIA, Naver, Brookfield | ~$10B | 200 MW AI factory + $1B equity stake |
| Samsung-Broadcom | Samsung, Broadcom | $200B | Memory + foundry + advanced packaging |
What this means for you
If you build AI products on cloud GPUs: Expect high but more predictable compute costs through 2027. The SK deal stabilizes the memory supply chain, but doesn't add HBM capacity faster. Plan for the Vera Rubin generation (H2 2026 cloud availability from AWS, Google Cloud, Microsoft, OCI, CoreWeave) and watch whether LOI-to-contract conversion in Q3-Q4 2026 filings delivers disclosed volumes. For a deeper look at how the rack-scale competition is reshaping the chip landscape, see our AMD vs NVIDIA AI Chips comparison: Helios vs Vera Rubin.
If you're a small business exploring AI: The practical move is to design around the available, not the promised. Current-generation GPUs (H100, H200, B200) are available on cloud platforms now; Vera Rubin arrives in H2 2026 with broad cloud availability. The deal means more capacity is coming to Asia-Pacific, but the cost trajectory depends on memory supply. For a broader view of the AI infrastructure investment cycle and what it means for builders, see our analysis of the AI infrastructure supercycle.
If you're budgeting for AI infrastructure: The single most important number to track isn't the $500 billion headline—it's SK Hynix's HBM4 yield rate and capacity ramp, which determines when memory supply catches up with demand. Until that gap closes, GPU rental prices stay high regardless of how many data centers get built. For the larger pattern of hyperscaler infrastructure bets and how they intersect with your compute budget, see our breakdown of OpenAI's $750 billion infrastructure plan through 2030.
If you're watching the geopolitics: South Korea is positioning itself as a sovereign AI infrastructure node that reduces the world's dependence on a handful of U.S. hyperscalers. NVIDIA's Jensen Huang explicitly named Korea's "networks, data centers, chip technology, and industrial scale" as the ingredients for becoming a global AI powerhouse. The deal is as much about geopolitical supply chain alignment as it is about compute capacity.
FAQ
Q: What is the NVIDIA SK Group $500 billion deal?
A: Announced July 24, 2026, it's a $500-billion-plus partnership between NVIDIA and South Korea's SK Group spanning two pillars: SK Telecom building a 2-gigawatt AI factory in Korea using NVIDIA Vera Rubin and DSX, and NVIDIA-SK Hynix co-developing HBM4 and future AI memory. The parties signed letters of intent, not binding contracts.
Q: Will the deal make cloud GPUs cheaper?
A: Probably not before 2027. The deal stabilizes HBM4 memory supply but doesn't add new fab capacity faster—fabs have 18-24 month lead times. HBM demand grows roughly 2x faster than supply, keeping GPU rental prices elevated. Relief may come in 2028 as new HBM4 fabs ramp and Vera Rubin volumes scale.
Q: What is HBM4 and why is it the AI memory bottleneck?
A: HBM4 (High Bandwidth Memory generation 4) is the stacked memory inside AI GPUs that feeds data to compute cores at extreme bandwidth. Without enough HBM, GPUs sit underutilized. Three companies make HBM at scale: SK Hynix (~62% share), Micron (~21%), and Samsung (~17%). New HBM capacity requires new fab construction with 18-24 month lead times.
Q: What is NVIDIA Vera Rubin and when does it ship?
A: Vera Rubin is NVIDIA's successor to Blackwell, announced at CES 2026 and in production as of Q1 2026. The NVL72 rack contains 72 Rubin GPUs and 36 Vera CPUs, delivering 3.6 EFLOPS NVFP4 inference. Each GPU has 288 GB HBM4 at 22 TB/s bandwidth. Cloud availability from AWS, Google Cloud, Microsoft, OCI, and CoreWeave begins H2 2026.
Q: Is the $500 billion figure confirmed or just a plan?
A: It's a plan expressed through letters of intent. The $500 billion represents aggregate planned investment across all partners over multiple years—not money that has changed hands, and only a portion accrues as NVIDIA revenue. Firm commercial terms depend on LOI-to-contract conversion in Q3-Q4 2026.
Q: How does this affect builders using AI APIs and cloud GPUs?
A: The deal improves supply chain predictability but doesn't lower prices short-term. Builders should plan for stable-to-high compute costs through 2027, design around currently-available GPUs (H100, H200, B200), and watch for Vera Rubin cloud instances in H2 2026 from major providers. The broader trend is more compute capacity in Asia-Pacific, helping sovereign AI customers.
Discussion
0 comments