Nvidia has tied up the one component it cannot build itself, locking in a high-bandwidth memory supply agreement with SK Hynix that could reach $500 billion over an unspecified number of years. The deal, announced Friday at an AI summit in San Francisco attended by South Korean President Lee Jae Myung, also commits the two companies to building data centers large enough to draw 2 gigawatts of power, a footprint that implies hundreds of thousands of GPUs coming online in 2027. For a company that has turned scarcity into pricing power, securing the HBM bottleneck is the closest thing to a moat Nvidia can still dig.
The memory bottleneck is the real constraint
High-bandwidth memory has become the gating factor for every advanced AI system, and SK Hynix sits at the top of a very short list of suppliers who can produce it at scale. Nvidia enterprise vice president Raj Mirpuri said the agreement includes co-development on next-generation HBM, which is code for ensuring the memory roadmap bends toward Nvidia's architecture rather than a competitor's. The $500 billion figure dwarfs the $200 billion memorandum Samsung signed with Broadcom the same day for memory and foundry collaboration, a reminder that the market still prices Nvidia's ecosystem at a premium to mere component supply.
The buildout is leaving the hyperscalers
The 2-gigawatt target and the parallel $1 billion investment in Korean cloud provider Naver, itself building 200 megawatts of GPU capacity, signal that AI infrastructure is no longer the exclusive province of the usual American cloud giants. SK Telecom, an SK Hynix affiliate, will run a cloud business on Nvidia's upcoming Vera Rubin systems, and SK Hynix's own Nasdaq listing earlier this month was explicitly framed as a financing vehicle for this expansion. Foreign governments and conglomerates are now writing the checks that used to come only from Seattle, Mountain View, and Redmond.
What to watch next
The 2027 timeline for the big data centers is aggressive for projects of this scale, and "could be worth $500 billion" carries the usual flexibility of a framework agreement rather than a purchase order. Samsung's Broadcom pact shows the memory oligopoly is not standing still. But Nvidia has effectively pre-sold a chunk of its own future roadmap to guarantee the silicon it needs to keep that roadmap viable. When the constraint is memory, the company that locks up the memory wins, even if the price tag is half a trillion dollars.