Open source AI models could turn out to be bullish, not bearish, for memory chipmakers, according to a point BofA analyst Vivek Arya raised that Jukan is amplifying. The intuition many investors have is that cheaper open weight models mean less spending on expensive memory hardware. Arya argues the opposite. Closed models like GPT or Gemini pool global demand into a small number of massive shared data centers, so the memory footprint per user actually shrinks as usage scales. Open models work the other way: every company that self hosts the same open model has to replicate the full set of weights and build its own separate memory pool, plus a dedicated KV cache, the temporary memory that holds a model's short term notes while it is generating an answer. With context windows of 128,000 to 1 million tokens becoming normal in 2026, that cache alone can top 40 gigabytes per active session. Layer on cheap Chinese open source APIs, which are pushing more companies to self host instead of renting from a big cloud, and you get more separate deployments, not fewer. Each one needs its own high bandwidth memory. So a wave of adoption that looks like it should be memory light instead multiplies memory sockets across the industry, a dynamic that plays directly into demand for chips like SK hynix's HBM stacks.
Why it matters · If this thesis holds, the bear case that cheap open models will shrink AI infrastructure spending gets flipped on its head, meaning memory demand for HBM makers could keep compounding even as model costs fall.
Worth asking · Does the self-hosting math actually multiply memory demand as open models spread, or will efficiency gains in inference eventually offset the extra footprint?