Analysis
Samsung Electronics and SK Hynix moved to lock in major new supply and investment agreements with U.S. technology companies on July 24, focused on long-term contracts for high-bandwidth memory (HBM) -- the specialized chip stacks that feed data to GPUs fast enough to keep pace with AI training and inference. The deal values were not disclosed ahead of the companies' announcements, but reporting describes them as "very large" and structured as a mix of supply contracts, strategic investments, and memorandums of understanding.
The timing is deliberate: the announcements landed during South Korean President Yoon Suk-yeol's visit to Silicon Valley for a high-profile AI summit that drew Samsung chairman Lee Jae-yong, SK Group chairman Chey Tae-won, Nvidia CEO Jensen Huang, and the CEOs of OpenAI, Anthropic, and Broadcom into the same room -- as clear a signal as the industry gets that memory supply has become a matter of state-level industrial diplomacy, not just corporate procurement.
The megadeals build on an already-disclosed roughly $880 billion joint capacity-expansion plan between Samsung and SK Hynix, the world's two largest memory chipmakers, to scale HBM production through 2026 and beyond. HBM has become one of the most supply-constrained components in the AI buildout, alongside advanced packaging capacity at firms like TSMC and Amkor and raw GPU output from Nvidia -- meaning memory is now as much a bottleneck on AI infrastructure scaling as compute itself.
“The deals land the same week Nvidia prepaid Amkor $1.5 billion for U.S.”
The deals land the same week Nvidia prepaid Amkor $1.5 billion for U.S. packaging capacity and OpenAI disclosed its $30 billion, self-built Georgia data center campus. Together, the three moves show compute, power, and memory all being locked down simultaneously by different players across the AI stack, rather than any single company cornering the whole supply chain.
For investors, the read-through is that memory pricing power has shifted decisively toward Samsung and SK Hynix after years of commoditized DRAM cycles -- HBM's technical complexity and AI-driven demand give both companies leverage they haven't had in a standard memory cycle. That pricing power flows directly into higher input costs for every AI infrastructure buyer, from hyperscalers to GPU-cloud startups.
The bear case: undisclosed deal values make it hard to assess whether these agreements represent genuinely new demand or simply formalize capacity Samsung and SK Hynix had already committed to the broader $880 billion plan, and memory markets have a long history of overbuilding capacity in boom years only to face gluts once demand normalizes.
Watch for the specific dollar figures to surface in follow-up disclosures, whether Micron secures comparable HBM agreements of its own, and whether HBM pricing shows up as a new line-item cost pressure in hyperscalers' next capex disclosures.
