BREAKING
Zhipu Weighs Custom AI Chip for GLM
0x
token usage in first week
0M
context tokens
Why Build In-House Silicon
1US export controls limit GPUs
2GLM demand surges
3Design ASIC for inference
4Cut reliance, boost efficiency
Promise vs Caution
Strengths
Opus-class substitute in real work
Strong value open weights
Good at agents and coding
Concerns
Underdelivers in agent loops
Trails top commercial models
Verbose reasoning, privacy doubts
Chip Plan Still Early Stage
Vertical Integration in Focus
AI NEWS BLITZ
Chinese AI lab Zhipu is exploring its own chip tuned for GLM models.