JPM analysts are observing a potential rise in Custom HBM, specifically starting with the HBM4e generation, which could disrupt the standard HBM market. They believe this architecture could become dominant and a key competitive differentiator for chipmakers and CSPs. JPM also projects CSP capital expenditure upside in 2026, driven by robust AI-related compute, storage, and networking despite concerns about negative free cash flow for smaller CSPs like Oracle and Coreweave. They anticipate HBM's value share in data center and AI revenue will continue to increase, potentially exceeding 20% of the DC/AI semiconductor revenue pool by 2028 due to content increases. This, combined with server memory, is expected to further boost memory's value share in AI cycles, supporting a multi-year bullish outlook for the memory market. JPM's HBM projections do not fully account for potential in-house demand from frontier AI model companies, representing further upside risk.