AI
China
TLDR: BEIJING—China Mobile and China Telecom will link a nationwide AI data center grid by 2028, with a plan to source 80 percent of AI chips domestically. The timeline faces bottlenecks from SMIC production limits and scarce domestic HBM, risking idle capacity.
Key Takeaways:
- The National Development and Reform Commission is designing a five year buildout using state carriers and debt funding.
- The blueprint targets 80 percent domestic AI chips, effectively sidelining Nvidia and AMD while relying on Huawei Ascend supply.
- SMIC N plus 2 output and limited domestic HBM could cap growth, with officials warning construction may outpace usable silicon.
- If domestic chips cover only about 76 percent of demand by 2030, the compute grid may need workarounds or face shortages.
Beijing wants a nationwide compute web, but the real bottleneck is not money. It is whether domestic chips and HBM can arrive fast enough to keep the servers fed, not just built.
Beijing wants a nationwide compute web, but the real bottleneck is not money. It is whether domestic chips and HBM can arrive fast enough to keep the servers fed, not just built.
Q&A
What happens if data center construction finishes faster than SMIC can supply advanced wafers?
Operators may idle racks, delay training workloads, or shift to lighter inference tasks until sufficient accelerators and memory are available.
Why does high bandwidth memory matter more than raw chip counts for AI accelerators?
Even with enough compute cores, limited HBM can throttle throughput for bandwidth heavy training and large batch inference, reducing practical performance.
How could the government’s 80 percent domestic chip rule change procurement strategy inside China’s state projects?
It can force planned substitution across hardware generations, prioritize certain accelerator models, and increase reliance on domestic memory and packaging rather than performance optimized imports.
What does Huawei’s reported chip shipping pace imply about sustaining an all domestic training stack?
If supply chain strain already shows up at scale, a larger national grid could amplify the pressure on packaging capacity, yields, and upstream components beyond compute dies.
What precedent does the DeepSeek switch from Huawei toward Nvidia suggest for the next generation of models?
It hints that frontier training workloads may still hunt for the most capable accelerators, so domestic hardware may be strongest for inference or staged training.
No comments yet. Be the first to share your thoughts!