VCs eye $12B opportunity as China’s AI chip upstarts crack training

The quest to replace Nvidia in China’s AI ecosystem is entering a new phase, as domestic labs begin shifting model training — not just inference — onto homegrown silicon. The South China Morning Post article examines five AI models trained on Chinese chips, revealing both significant progress and a stubborn compute gap. While Chinese AI models are increasingly competitive with US counterparts, none of the country’s top-tier large language models (LLMs) have been pre-trained entirely on local hardware. Instead, adoption has been concentrated in the less demanding inference stage. This disparity traces back to the three-tiered AI development pipeline: pre-training (the most computationally intensive phase requiring thousands of GPUs running for months), post-training (fine-tuning with less compute), and inference (deploying the model). Driven by Washington’s escalating export controls and Beijing’s push for technological self-sufficiency, Chinese AI labs are now experimenting with early-stage training on indigenous accelerators. The article highlights Zhipu AI’s GLM-Image, an image generation model open-sourced in January 2026, as the first state-of-the-art multimodal model trained entirely on domestic hardware — specifically Huawei’s Ascend Atlas 800T A2 servers powered by the Ascend 910 AI accelerator and the MindSpore deep learning framework. However, training image models demands significantly less compute than LLMs, and Zhipu has yet to achieve full LLM pre-training on domestic chips. Natixis economist Gary Ng notes that while dependence on indigenous suppliers may slow development efficiency, China is constructing an entire domestic AI supply chain — a rare global feat. The long-term strategic play is clear: if Chinese labs can crack high-end training on local silicon, it would fundamentally alter the global AI hardware market, which Nvidia currently dominates with an estimated 80%–90% share in the data center GPU segment. The implications are profound for investors and startups alike. For Nvidia, the growth story in China — which contributed 20–25% of its data center revenue before recent restrictions — faces structural headwinds. For the startup ecosystem, the shift opens a massive opportunity for chip design firms, AI optimization software, and model-training-as-a-service platforms tailored to domestic hardware ecosystems. The transition will be gradual; current domestic chips lack the performance and software maturity of Nvidia’s CUDA ecosystem, but the forced decoupling is accelerating investment. The article identifies five models that mark the early inroads: GLM-Image, and presumably others detailed in the full report. As of mid-2026, no Chinese LLM has been fully pre-trained on local silicon, but the technical trajectory suggests that milestone could be reached within 12–18 months. Once that happens, expect a wave of updates from cloud providers like Alibaba Cloud and Huawei Cloud offering domestic training clusters, and a re-rating of Chinese AI chip startups. For global markets, a successful decoupling would reduce China’s reliance on Nvidia, potentially softening its pricing power and reshaping geopolitical risk premiums.