For three years the story of the AI boom has been a story about compute, and compute has meant Nvidia. This week a quieter constraint pushed its way to the front. When Nvidia and SK Group unveiled their half-trillion-dollar Korean venture, the first thing Nvidia said it had secured was not chips but memory: a long-term supply of high-bandwidth memory, or HBM, from SK Hynix. The most valuable company on earth was making sure it would not run short of the one part it does not make itself.
The reason has an evocative name: the memory wall. A modern GPU can perform staggering numbers of calculations, but only if data reaches it fast enough. Ordinary server memory cannot keep up, so the largest AI chips are wrapped in stacks of HBM bonded directly onto the processor package. Without enough of it, as one hardware analysis put it this week, even the most powerful GPUs simply idle, waiting for data to arrive. In an AI hall the expensive silicon is only ever as useful as the memory feeding it.
The boom has turned that memory into the scarcest resource in the industry. SK Hynix, the HBM leader, has spent recent weeks locking in customers; Micron, once a maker of commodity RAM, has been reborn as a specialist supplier whose fortunes now track Nvidia's almost one for one. The clearest tell came from Samsung, which the same week signed a memorandum with Broadcom worth an estimated 200 billion dollars to expand across memory and foundry. Two of the biggest numbers in tech, announced days apart, and both are really contests over who controls the supply of memory.
The re-rating is spreading down the chain. Analysts at Simply Wall St this week pointed investors toward the less glamorous companies a two-gigawatt build implies, such as the makers of power-management chips that convert and marshal the electricity a memory-dense AI hall consumes. As the racks grow hungrier, every watt has to be accounted for, and firms most people have never heard of suddenly look like AI plays. The buildout is turning out to be deep as well as wide.
None of this dethrones Nvidia; its accelerators remain the product everyone is queuing for. But it complicates the tidy picture of a single company owning the AI age. The binding constraint has migrated from the processor to the memory that surrounds it, which is why a chip designer felt it had to spend years, and a headline-grabbing sum, tying up a memory supplier before a rival could. The next phase of the race may be settled less by who designs the fastest chip than by who can guarantee the memory to keep it fed.