State-owned China Telecom has unveiled what it describes as the country’s first large-scale artificial intelligence models built with the Mixture-of-Experts (MoE) architecture and trained entirely on domestically developed chips from Huawei Technologies, according to SCMP.
The TeleChat3 models, developed by China Telecom’s Institute of Artificial Intelligence (TeleAI), span an unusually wide range of sizes, from 105 billion parameters to models running into the trillions. According to a technical paper published last month, the models were trained exclusively using Huawei’s Ascend 910B chips alongside MindSpore, Huawei’s open-source deep-learning framework.
At a technical level, the use of the Mixture-of-Experts architecture is central to why the announcement matters. MoE models have become the dominant approach for frontier systems because they allow developers to scale models to hundreds of billions or even trillions of parameters without linearly increasing computing costs. Instead of activating the entire network for every task, MoE systems route inputs to smaller, specialized submodels. This efficiency is precisely what makes MoE attractive for Chinese firms that face constrained access to top-tier GPUs.
China Telecom’s researchers say their work demonstrates that Huawei’s Ascend 910B chips, paired with the MindSpore framework, can support this demanding architecture at scale. In practical terms, that means handling complex parallelism, inter-chip communication, and training stability issues that have historically been easier to manage with Nvidia’s CUDA ecosystem.
MoE training is notoriously sensitive to imbalances between experts and prone to instability during fine-tuning, making it a stress test for any AI stack. The fact that TeleAI claims to have run models ranging from 105 billion parameters to the trillion-parameter class on this setup is intended to signal robustness, not just a one-off proof of concept.
Still, the performance results tell a more nuanced story. TeleChat3’s benchmark scores lag behind OpenAI’s GPT-OSS-120B on several standard evaluations. That gap reinforces a point Chinese researchers increasingly acknowledge in public: domestic chips can now support large-scale training, but they still struggle to match the efficiency, maturity, and raw performance of Nvidia’s latest GPUs when it comes to the very top tier of AI capability. In effect, China is narrowing the “can we train at all?” gap faster than the “can we match the best?” gap.
This distinction is important for how Beijing views success. The priority is not necessarily immediate parity with OpenAI or other Western labs, but reducing strategic vulnerability. From that perspective, the ability to train MoE models end-to-end on a fully domestic stack represents a form of resilience. Even if the resulting models are less competitive at the frontier, they can still underpin commercial services, government systems, and industrial applications at a massive scale.
China Telecom’s role also matters. As one of the world’s largest telecom operators, it sits at the intersection of infrastructure, data, and state policy. Its endorsement of Huawei’s AI stack carries political and industrial weight that smaller startups lack. By publishing detailed technical results, the company is effectively validating Huawei’s position as the backbone of China’s AI ambitions at a time when US export controls are designed to slow exactly that outcome.
The broader ecosystem is moving in the same direction, though unevenly. Zhipu AI’s claim that its image-generation model achieved leading results while training entirely on Huawei chips adds momentum to the narrative that domestic hardware is becoming viable for more than just language models. Ant Group’s earlier disclosure about training a 300-billion-parameter MoE model without “premium GPUs” hinted at similar progress, even if it stopped short of full transparency about the hardware used. Together, these announcements suggest a growing willingness among Chinese firms to accept some performance trade-offs in exchange for independence from US technology.
At the same time, Nvidia’s continued relevance underscores the limits of decoupling. The company still positions its GPUs and software tools as the gold standard for MoE training, and for many Chinese developers, access to Nvidia hardware remains the fastest route to state-of-the-art performance.
The recent approval for sales of Nvidia’s H200 chip to China illustrates this tension. While Washington has allowed limited exports, Beijing’s reported stance – approving such purchases only in exceptional cases – signals a deliberate effort to avoid rebuilding dependence just as domestic alternatives are becoming usable.
Beijing has made self-reliance across the AI stack a core objective for the next five years, framing it as both an economic and national security issue. US restrictions have already reshaped investment priorities, pushing capital toward chip design, AI frameworks, and model optimization techniques that squeeze more output from less powerful hardware. MoE architectures fit neatly into that strategy, as they reward software sophistication over brute-force compute.
In that sense, TeleChat3 is as much a policy artefact as a technical one. It demonstrates that Chinese firms can adapt leading AI paradigms to constrained environments, even if the results remain a step behind global leaders. The remaining question is whether incremental improvements in domestic chips and software can eventually close that gap, or whether China will settle into a parallel AI ecosystem that prioritizes scale, deployment, and sovereignty over absolute performance.
