Infrastructure September 21, 2026 mixed ⇧ 503 pts across 1 thread

HBM4 supply surge and the AI memory cost war

Samsung is expected to more than double output of HBM4 and HBM4E DRAM, which sounds like good news for AI infrastructure costs. But commenters quickly noted the likely effect on consumer DRAM: prices will probably stay high or get worse, because Samsung's capacity is being redirected to the AI supply chain rather than the consumer market. One commenter noted they bought DDR4 or DDR5 a couple years ago and the price has since tripled.

The pattern here is that the AI infrastructure buildout is creating a two-tier memory market. High-bandwidth memory for AI accelerators is getting massive investment, while commodity consumer RAM sits in a weird supply-demand state where the same manufacturers are prioritizing the higher-margin AI contracts. The question of whether HBM4 supply will actually be enough for AI's appetite remains genuinely open, and multiple commenters treated it as rhetorical.

For builders watching infrastructure costs, this matters because HBM availability is a direct constraint on how many GPU hours the major cloud providers can actually sell.


So what?

If your business model depends on cheap GPU compute, the HBM supply chain is the constraint to watch. More Samsung HBM4 production helps the hyperscalers but doesn't directly translate to cheaper cloud GPU pricing in the short term. Plan your compute budgets conservatively and prioritize workloads where inference efficiency matters, because raw compute is not about to get dramatically cheaper.

Read these