China is losing the LLM race but can still win in AI, ex‑Tencent AI lead says

The assessment

A former head of AI at Tencent (腾讯) has warned that China is falling behind Western firms on large language models (LLMs), but argues the country can still "win" in broader artificial intelligence by focusing on practical deployment and industrial applications. It has been reported that the former executive pointed to gaps in compute, access to cutting‑edge chips and open research as key reasons China’s biggest models lag those from the US. Short on the largest training clusters and constrained by export controls, China’s LLM landscape is being reshaped by forces beyond pure engineering.

Why China can still compete

Why does that matter? Because dominance in raw model size is not the only path to economic impact. China’s tech giants — including Baidu (百度), Alibaba (阿里巴巴), ByteDance (字节跳动) and Tencent (腾讯) — have unrivalled scale in users, data and distribution channels. Reportedly, the ex‑executive argued that this ecosystem advantage, combined with state support for AI deployment and a focus on customised vertical solutions, could translate into industrial leadership even if frontier LLM research trails.

Geopolitics and the technology gap

The geopolitical context is central. US export controls on advanced semiconductors and other trade policies have limited Chinese firms’ access to the largest GPUs and AI chips, raising the bar for training ever‑bigger models. At the same time, Chinese policymakers have pushed for self‑reliance in chips and AI, accelerating investment in domestic hardware and research. It has been reported that the trade restrictions are forcing Chinese teams to prioritise efficiency and application‑level innovation rather than competing purely on model scale.

The outlook

The former Tencent AI lead’s view underscores a bifurcated race: one track for headline‑grabbing LLM breakthroughs, and another for real‑world, sector‑specific AI that changes industries. Who wins will depend on more than architecture papers and model parameters. It will be decided by supply chains, chip access, regulatory environment and the ability to translate models into products people and businesses actually use.