Citi Ranks Broadcom No.2 as Chinese AI Models Cut HBM Demand
AVGO•Chinese AI labs deploying MoE, MLA and FP8 mixed-precision models like GLM-5.2 are reducing per-query computation and HBM demand, potentially undermining Broadcom’s custom AI silicon revenues. Meanwhile, Citi has ranked Broadcom second among AI compute semiconductor stocks, citing strong memory supply agreements through 2030.
1. Algorithmic Efficiency Gains
Chinese labs’ GLM-5.2 and other MoE models activate only necessary expert subnetworks, while MLA cuts long-context memory footprint and FP8 lowers compute requirements, together slashing inference costs to roughly one-eighth of conventional architectures.
2. Implications for Broadcom’s AI Silicon
Broadcom, a leading supplier of custom AI silicon and HBM memory, faces potential volume pressure as per-token compute and bandwidth needs decline, challenging its revenue assumptions built on high-bandwidth memory demand.
3. Citi’s AI Compute Semiconductor Ranking
Citi analysts placed Broadcom second among AI compute semiconductor stocks, assigning top weight to memory supply allocation and highlighting strategic customer agreements extending through 2030 that secure its HBM capacity.
4. Strategic Outlook for Broadcom
To mitigate risk from algorithmic efficiency trends, Broadcom may diversify its silicon portfolio, pursue advanced memory technologies or deepen hyperscale partnerships to sustain growth amid shifting compute and bandwidth dynamics.
