Other
2026-07-17
Baidu Kunlun Chip Supernode Mass Production; Shishi Tech Unveils Domestic Token Optimization Factory
Summary
Key Takeaways
At WAIC 2026, Baidu announced the mass production delivery of its Kunlun Chip Supernode, supporting trillion-parameter model training per node, marking a milestone for domestic AI chips moving from design to scale deployment. Baidu's DuMate was recognized as one of the top ten 'treasures of the museum', the only general agent product showcasing natural language task decomposition and cross-application closure.
Shishi Tech debuted at WAIC with the Domestic Token Optimization Factory, positioning itself as 'taming chips' rather than 'making chips'. It focuses on software stack optimization, scheduling framework tuning, model adaptation, and operator-level fine-tuning to industrialize large model inference. The booth displayed real-time metrics including average TPM and TTFT, allowing comparison of inference speed before and after optimization across multiple mainstream models.
This signal indicates Baidu's dual push in computing and applications, with Kunlun Chip Supernode providing hardware foundation, while Shishi Tech's 'taming' approach solves the last-mile problem from chip to usable computing power via software optimization.
Why It Matters
Beneath the surface of domestic chip mass production and software optimization lies a strategic move to counter NVIDIA in the Chinese market. Baidu's Kunlun Chip Supernode deliberately omits key performance metrics (e.g., peak TFLOPS, HBM bandwidth), suggesting it still lags behind H100. Shishi Tech's Token Optimization Factory, while improving inference via operator-level tuning, risks toolchain lock-in: once users adapt to its optimization stack, migration to other chips becomes costly. Moreover, the optimization is model-specific and may not generalize to evolving workloads. Enterprises must scrutinize the communication bottlenecks and stability issues in large-scale distributed training that software optimization alone cannot resolve.
PRO Decision
For Vendors (competitors like Huawei Ascend, Cambricon, NVIDIA): Huawei should exploit Baidu's opaque performance metrics by publishing standard benchmarks (e.g., MLPerf) and offering more open optimization stacks. NVIDIA should reinforce CUDA ecosystem moats and create compatibility layers for domestic optimization tools.
For Enterprises: CIOs must conduct independent benchmarking of Kunlun Chip Supernode, focusing on linear scaling efficiency and interconnect bandwidth. For Shishi Tech's Token Factory, adopt a modular integration approach to avoid lock-in. Evaluate long-term supply stability and software ecosystem maturity, and maintain multi-cloud, multi-chip redundancy.
For Investors: Look beyond PR hype; track actual shipment volumes and customer adoption of Kunlun Chip. Shishi Tech's 'taming' model is innovative but its sustainability depends on customer stickiness and technical moats. The domestic AI chip race will ultimately favor vendors with superior performance, ecosystem, and openness.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)