🧠 Food for thought

Implications, context, and why it matters.

China Mobile chip plan runs into HBM limits

• China Mobile plans to use only homegrown chips by 2028 and is exploring a 100,000 GPU cluster, but high bandwidth memory (HBM) supply is a bottleneck. China stockpiled about 13 million HBM stacks before US curbs, enough for roughly 1.6 million Huawei Ascend 910C packages 1.
• CXMT will make about 2 million stacks in 2026, which can support around 250,000 to 300,000 Ascend chips 1. Without foreign HBM, China cannot build 1 million AI chips in 2025 1, so a 100,000 GPU cluster looks possible while the 100 EFLOPS plan by 2028 stays uncertain.
• SemiAnalysis says SMIC would need at most 20,000 wafers per month to ship millions of Ascend dies 1. Capacity may rise from 45,000 per month by end 2025 to 80,000 by 2027 1, yet memory still caps scale.

Suppliers stand to benefit from liquid cooling buildout

• Racks packed with Ascend 910B or 910C chips that draw about 310 W each push power density beyond 15 to 30 kW per rack 2, which makes air cooling fall short.
• China’s liquid cooled server market hit $2.37 billion in 2024 with 67% growth, and it is projected to reach $16.2 billion by 2029 on a 46.8% CAGR 2. Cold plate designs hold 95% share due to lower retrofit costs 2, which opens near term orders for cooling distribution units, heat exchangers, and specialized piping.
• Vendors also see openings in high speed Ethernet or RDMA over Converged Ethernet (RoCE) gear for large clusters, as Huawei’s UnifiedBus seeks TB per second bandwidth with about 2.1 microsecond latency 3. Power systems face rising demand as AI pushes Tier 4 data center designs, the highest Uptime Institute tier 4.