China's Cambricon, SenseTime join forces to boost AI push

Chinese chip designer Cambricon Technologies and AI firm SenseTime have teamed up for strategic cooperation to advance AI infrastructure in China.

They will jointly develop AI computing solutions and enterprise services by combining SenseTime’s foundational AI models with Cambricon’s processors.

The partnership aligns with China’s broader push for tech self-reliance under the “AI Plus” strategy announced by the State Council in August.

SenseTime, based in Hong Kong, has faced financial losses and US sanctions, while Cambricon, headquartered in Beijing, reported first-half 2025 revenue of 2.9 billion yuan (US$404.8 million).

SenseTime’s AI models have struggled to compete with local rivals like DeepSeek and Alibaba Cloud’s Qwen.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

Cambricon lacks public MLPerf results to back performance claims

Despite a fivefold stock rise plus a 4,348% year-on-year revenue jump, we found no recent public MLPerf submissions for Cambricon in coverage where NVIDIA leads while AMD advances 1.
MLPerf Inference v5.1 tracks 10 models that span Large Language Models (LLMs), text-to-image, and reasoning workloads 2. Without public MLPerf results for Cambricon, enterprise AI teams lack outside proof that its processors can serve large models at scale on par with established options.
The partnership release promotes ‘indigenous AI development’. It gives no technical specifications, framework compatibility details, or performance metrics to judge if Cambricon chips can handle production workloads for SenseTime’s foundational models.

AI infrastructure operators face tooling gaps if workloads shift to Cambricon

PyTorch (an open-source deep learning framework) users need Cambricon’s CATCH extension 3 to enable MLU support, where MLU refers to Cambricon’s Machine Learning Unit accelerators. CUDA codebases use NVIDIA’s GPU programming platform. Migration requires mapping PyTorch operations that Cambricon supports to standard implementations, which adds friction.
Models built in PyTorch can export to Open Neural Network Exchange (ONNX) format 4 for cross-platform deployment. ONNX Runtime compatibility docs do not mention Cambricon-specific support 5, so MLOps (machine learning operations) teams that run multi-vendor stacks may face model portability and unified serving issues.
Cloud service providers and enterprise AI teams that evaluate Chinese chip options see gaps in standardized toolchains for CUDA-to-Cambricon migration, performance profiling, plus inference tuning. That gap opens room for developer tools and professional services.