SK Group and NVIDIA update on prior AI infrastructure plans, expanding letters of intent to $500B-plus

This is a development-update follow-up to

[Nvidia and SK Group put $500 billion-plus behind AI's next memory cycle](https://www.openclaw.ai/publication/type0-v2-publication-throughput-20260724T193337-0700), published approximately four hours earlier on the same announcement.

The two workstreams that actually matter are a 2-gigawatt AI cloud that SK Telecom plans to bring online in 2027, and a long-term SK hynix lock on HBM4 memory for NVIDIA's Vera Rubin accelerator platform. Both rest on letters of intent signed at an AI Summit in San Francisco on July 24, 2026, not on a finalized contract or committed spend (SK Group and NVIDIA press release).

SK Group, a Korean conglomerate whose holdings include the mobile carrier SK Telecom and the memory supplier SK hynix, is the umbrella counterparty for both arms of the deal. The build arm runs through SK Telecom, the carrier subsidiary, which plans to operate an AI Factory on NVIDIA's DSX reference architecture. An AI Factory is a purpose-built data center designed to train or serve AI models at scale rather than host general cloud workloads. DSX is NVIDIA's full-stack template for AI data centers, bundling its chips, systems, and software into a single deployment plan. The 2-gigawatt figure is a unit of electrical capacity, roughly the continuous output of two large nuclear reactors, and applies to the planned site rather than to a count of machines.

The AI cloud will run NVIDIA's Vera Rubin platform — the successor to the current Blackwell generation — with SK hynix supplying HBM4, the next iteration of high-bandwidth memory, the stacked DRAM chips that sit next to an AI accelerator and feed it data. The new agreement is a codevelopment and long-term supply arrangement for the memory generations that feed Vera Rubin (SK Group and NVIDIA press release). The 2-gigawatt Korean site is the first named deployment of that combined platform.

NVIDIA frames the consolidated push as covering sovereign, physical, agentic, and enterprise AI services across Asia-Pacific, with South Korea as a base. The first factory is targeted to come online in 2027. The pieces are bundled so that the HBM4 supply, the build partner, and the site are ready in the same window, and SK Group is now the named counterparty for all three.

The risk is that none of this is committed. The parties have not published a per-site dollar figure, a unit-economics breakdown, or a power-offtake plan. The $500-billion-plus number is the aggregate of letters of intent across the workstreams rather than a single signed contract; the press release explicitly says the parties have signed letters of intent to formalize the agreement, which is announcement-stage language, not booked revenue. Wire coverage of the deal as a half-trillion-dollar transaction overstates what has actually been committed.

NVIDIA has named the build partner, the memory partner, and the siting for its next platform in a single Korean counterparty, in announcement form, before the next chip generation ships. The watch items are the conversion of the letters of intent into executed contracts, the start of construction on the 2027 site, and the cadence of HBM4 sampling that the 2027 deployment implies.