NVIDIA and SK Group unveil $500B-plus AI data center and memory initiative

The SK Group tie-up is one of the largest single AI-infrastructure commitments announced to date, spanning data-center buildout and next-generation memory supply through SK Hynix — a critical partner given how tightly high-bandwidth memory constrains AI accelerator production. The scale signals NVIDIA's strategy of locking in both compute demand (data centers) and its own upstream supply chain (memory) in a single geopolitical bloc.

It lands alongside strong operational proof points for NVIDIA's roadmap. CoreWeave became the first cloud provider to fully validate NVIDIA's Vera Rubin NVL72, scaling to 1,152 Rubin GPUs and roughly 60 exaflops of FP8 across five racks, with up to 10x better inference-per-watt than Blackwell. That 'tokens-per-watt' figure drew celebration on Hacker News for reframing power efficiency — not raw FLOPS — as the binding constraint on AI deployments. NVIDIA also launched its Halos robotics safety stack and GB300 Blackwell Ultra systems supporting Anthropic's Claude models.

Competitively, the announcements reinforce NVIDIA's dominance even as customers hedge: Anthropic's AMD deal and Google's 'Frozen v2' custom-chip plans show buyers diversifying, but the sheer scale of the SK commitment and the Vera Rubin efficiency gains keep NVIDIA at the center of AI infrastructure economics. The near-record stock reflects that demand story. The risk to watch is concentration — half-trillion-dollar commitments assume AI demand keeps compounding, and any deceleration would expose the capital intensity underpinning these deals.