Alibaba CEO: next Qwen model is being trained on the company’s own T-Head chips
Alibaba is already training its next-generation Qwen model on chips from T-Head, its own chip unit, chief executive Eddie Wu said. Wu expects most of the company’s new AI chips to come from T-Head after 2027.
Wu spoke on the company’s quarterly earnings call on Thursday 19 November, as reported by Nikkei Asia and the news agency AFP, among others. T-Head is Alibaba’s in-house chip design unit. All the reports are based on Wu’s remarks. Alibaba did not give the model’s name, say when it will be completed or released, or say which chip it is using for training.
The heart of the story is the difference between training and inference. In training, a model is built from a large dataset, which takes the most computing power. In inference, the finished model responds to users’ requests.
Tencent’s management draws the line at exactly this point. Chief strategy officer James Mitchell said the company does not see the restrictions affecting its cloud business, because the service now uses its computing mainly for inference, Fortune reported. President Martin Lau, however, admitted that training the next Hunyuan model will now take longer. Alibaba also acknowledged that the restrictions could slow the development of new models.
Wu said that current T-Head chips are already fully capable of training even ultra-large models, Fortune reported. He added, however, that it was ‘too early at this stage’ to judge how the restrictions would affect Qwen’s development. Training is still under way. The Economy, citing Nikkei, reports that Alibaba plans to run the entire training process, from pre-training to post-training, on domestic chips next year. The Economy says the chips are Zhenwu chips, but other sources do not name them.
NeoTeo, citing AFP, reports that Wu expects most of Alibaba’s new AI chips after 2027 to be T-Head chips. This is an expectation. For its own inference, the company also plans to buy chips from other domestic manufacturers. TechNode, citing the Chinese news service CLS, reports that Alibaba’s Bailian model platform already provides inference on Chinese chips.
The background is an interim rule that the US Department of Commerce published on Monday. Under the rule, advanced chips can now be sold to data centres in Malaysia, Singapore, Thailand, Indonesia and Brazil only with an export licence. Large data centres also need a licence if they offer substantial computing power to Chinese customers. Wu estimated that cloud revenue growth could slow to 30–45 per cent and return to at least 50 per cent by 2028. In the third quarter, Alibaba’s AI and cloud revenue grew by 53 per cent.
Read more: Models