NVIDIA’s simultaneous push to lock in massive memory supplies from South Korea while leading a coalition against restrictive AI policies reveals a calculated effort to shape both the technical and regulatory foundations of the coming agentic era. At stake is whether frontier capabilities remain distributed across open-weight systems or consolidate behind a handful of closed providers, a contest sharpened by China’s rapid gains in benchmark performance.
The July announcements underscore how hardware leadership, supply-chain control, and policy influence now function as interdependent levers. NVIDIA is no longer content to supply GPUs; it is assembling the full stack—compute, memory, software, research partnerships, and even military-grade deployments—that will determine who trains, fine-tunes, and operates the next generation of AI agents.
Coalition Warns Policymakers Against Premature Limits on Open Models
A letter released July 24 by NVIDIA, Microsoft, Meta, Palantir and more than twenty other firms urged U.S. officials to refrain from restricting open-weight models, arguing that such measures would “stifle competition or drive innovation overseas.” The signatories emphasized that closed systems can be breached or misused without external scrutiny, while concentration behind a few providers amplifies systemic risk. NVIDIA CEO Jensen Huang and Microsoft CEO Satya Nadella amplified the letter on social media, and Elon Musk publicly endorsed it.
The timing is deliberate. Moonshot AI’s Kimi K3 model recently surpassed leading U.S. offerings on several benchmarks, prompting Treasury Secretary Scott Bessent to signal potential sanctions over alleged intellectual-property theft. By framing open weights as a competitive and security advantage rather than a liability, the coalition seeks to preempt export-style controls that could hand Chinese labs an uncontested domestic market while slowing Western experimentation.
South Korea Emerges as NVIDIA’s Strategic Anchor in Asia
Parallel announcements at the AI Summit in San Francisco cemented a multi-billion-dollar expansion of NVIDIA’s Korean footprint. SK Group and NVIDIA outlined a partnership potentially exceeding $500 billion, encompassing AI factories powered by Vera Rubin systems and SK hynix HBM4 memory, with SK Telecom constructing cloud capacity targeting two gigawatts of demand. Separately, NVIDIA committed $1 billion to NAVER to scale its GAK Sejong data center to 200 megawatts—roughly 100,000 GPUs—on the Vera Rubin platform.
At the same summit, NVIDIA and KAIST launched the first joint AI research laboratory between a Korean university and a global technology company. The lab will focus on agentic models optimized for Korean language and industrial use cases, leveraging NVIDIA Nemotron open models and local cloud infrastructure while creating internship and full-time pathways for Korean talent. These moves position South Korea as both a major consumer and co-developer of the next wave of physical and agentic AI systems.
Hardware Roadmap Extends Control Across the Full AI Stack
NVIDIA’s Vera Rubin architecture, successor to Grace Blackwell, integrates one Vera CPU for every two Rubin GPUs, delivering a 36-to-72 ratio inside each NVL72 rack. Executives described the system as markedly more “plug-and-play” than prior generations, with OpenAI already operating an early rack. The design reflects a deliberate strategy to supply not only accelerators but the orchestration CPUs required for complex agent workflows that move data across networking, storage, and inference pipelines.
Technical validation arrived earlier in the week when the GB300 NVL72 set a world record for mixture-of-experts pre-training, achieving 1,648 TFLOPs per GPU on DeepSeek-V3’s 671-billion-parameter model. Because MoE architectures shift the bottleneck from compute to all-to-all communication, the result demonstrates that NVIDIA’s tightly coupled scale-up fabric continues to outpace alternatives even as model architectures evolve. The company is simultaneously shipping the Vera CPU as a standalone product, signaling intent to capture sockets traditionally held by x86 vendors inside AI data centers.
Military and Research Institutions Adopt On-Premises Frontier Compute
On July 23, Jensen Huang commissioned a DGX GB300 system at the Naval Postgraduate School in Monterey, California. The installation, anchored by an NVIDIA AI Technology Center, gives more than 1,500 resident students and 600 faculty on-site access to large-scale training and inference for weather prediction, cybersecurity, and disaster-response planning. Admiral Samuel Paparo, commander of U.S. Pacific Command, framed the deployment as essential modernization of leadership education, noting that officers must understand both the opportunities and responsibilities of these technologies.
The move illustrates how national-security customers are moving beyond cloud-only models toward sovereign, on-premises infrastructure capable of classified workloads. It also extends NVIDIA’s pattern of embedding reference architectures inside institutions that shape future doctrine and procurement.
Implications for Competition, Supply Chains, and Governance
Taken together, the week’s developments show NVIDIA converting its current hardware dominance into durable advantages across three domains: regulatory influence that preserves open-weight optionality, deep industrial partnerships that secure memory and power capacity, and direct relationships with military and academic customers that embed its platforms in high-stakes environments. Rivals face a narrowing window in which to match both the performance trajectory of GB300-class systems and the ecosystem density NVIDIA is constructing in Korea and elsewhere.
The open-weight coalition’s success or failure will largely determine whether future agentic systems can be inspected, modified, and redeployed by a broad set of actors or whether they remain gated behind proprietary APIs. Meanwhile, the scale of the SK Group and NAVER commitments—hundreds of thousands of GPUs and gigawatts of power—suggests that the next phase of AI infrastructure will be financed and operated by nation-scale entities rather than hyperscalers alone. How these parallel tracks intersect will shape both the technical character and the geopolitical distribution of advanced AI capabilities through the remainder of the decade.
