Accelerator vendors buy their way into memory — the HBM supply chain starts vertical-integrating

As a leading indicator, the reading is that the RAM shortage created by AI demand is now being addressed by AI-silicon players themselves acquiring memory technology in-house. The prior structure — Samsung, SK hynix and Micron independently supplying HBM — is starting to give way to internalization from the accelerator side. From a capex lens, it's early evidence of a vertical integration push spanning GPU/accelerator, memory and interconnect.

The fork is whether (a) AMD's Mext buy is a one-off technology add or (b) the opening move in a broader vertical-integration trend across AI silicon. In the first read, the independence of Samsung and SK hynix is preserved; in the second, Nvidia likely follows with similar M&A and the structure of memory supply rewrites itself. The Marvell hire looks like a second-order wave in which technology assets get reshuffled across the industry.

Watch three things: (1) Nvidia's own memory-related M&A or partnership moves, (2) how concentrated HBM sales are among Samsung, SK hynix and Micron, and (3) whether custom-silicon players like Marvell and Broadcom continue absorbing AMD-origin talent. If (2) shifts away from hyperscaler-driven concentration, memory vendors move from independent suppliers to embedded components in vertically integrated stacks — a meaningful reshaping of the AI capex value chain.