Tencent Hunyuan Pre-Training Lead Joins Thinking Machines Lab, September 2026
Yao Xingcheng, who led pre-training for Tencent's Hunyuan LLM, has reportedly joined North American startup Thinking Machines Lab at a compensation package above his tens-of-millions-of-RMB annual salary at Tencent, arriving as Tencent-backed Enflame drew 6,109 times oversubscription in its Shanghai debut — exposing a growing divide between Tencent's hardware momentum and its frontier talent retention.
The departure of Yao Xingcheng, who led pre-training for Tencent's Hunyuan large language model, to North American startup Thinking Machines Lab is, by itself, a single data point. But the reported compensation — described as meaningfully above the tens of millions of RMB he earned annually at Tencent — transforms it into a market signal: the global competition for frontier AI researchers has reached deep enough into China's hyperscalers to pull out the architects of their own foundation models. The pattern is not unique to any one company; elite ML engineers everywhere weigh compensation, publication freedom, and compute access against institutional constraints. What makes the timing of the departure notable is that Yao led the pre-training function — the foundational layer of any large language model.
The departure lands against a backdrop of accelerating hardware investment. Under China's Eastern Data, Western Computing programme, Tencent and Huawei have been building large-scale AI data-centre clusters in Guizhou and other rural provinces, leveraging surplus renewable energy to reduce operating costs at scale. Separately, Tencent and ByteDance received early deliveries of Nvidia H200 units following a relaxation of import restrictions, though most licensed chips reportedly had to remain in Hong Kong due to power-supply constraints on the mainland. Whatever the staging challenges, the chip-access picture for China's hyperscalers has materially improved since the export-control tightening of 2022 to 2023, and Chinese cloud giants including Tencent were, as of late 2024, already among the largest global buyers of Nvidia flagship AI chips outside Microsoft.
On domestic semiconductor supply, Tencent has pursued deliberate diversification. The company ordered AI accelerators designed internally by Baidu — a transaction analysts described as a formal sign of decoupling within China's historically closed internet ecosystem — while also backing Enflame, a Shanghai-based AI chip maker. Enflame's September IPO drew 6,109 times online demand and its Shanghai debut surged 206%; Tencent, Xiaomi, and GigaDevice each participated in the strategic placement tranche. The picture is not uniformly clean: Yunbao Intelligent, a domestic DPU startup whose prospectus disclosed that over 90% of revenue comes from Tencent-related parties, is itself preparing an IPO — a degree of concentration that exposes both Tencent's purchasing dominance in the domestic chip ecosystem and the concentration risk embedded in its early-stage vendor base.
At the model layer, pressures are more structural. Comparative analysis reported in September 2026 showed a 20-fold price spread between Tencent's Hy3, GLM-5.3-Flash, and Kimi K3 inference endpoints — a differential that signals aggressive commoditisation across China's LLM market. In that environment, losing the engineer who built the pre-training stack matters: model quality and inference efficiency depend heavily on architectural choices made at that layer, and those choices compound over successive training runs. At the market level, CXMT, the domestic DRAM and HBM manufacturer, overtook Tencent to become China's most valuable listed company just 17 days after its own IPO, reaching a reported $524 billion valuation — a reordering that places semiconductor infrastructure ahead of software platform incumbents in Chinese investor priorities, and reflects how far the centre of gravity in Chinese tech has shifted toward the hardware layer Tencent is now actively working to secure.
On the strategic front, Tencent has reportedly been in discussions to acquire a majority stake in AI-agent startup Manus at the $2 billion valuation at which Meta had earlier agreed a deal before Beijing forced Meta to unwind it. Whether that acquisition closes remains unconfirmed. Tencent's stock was at HK$426, down 1.7%, at the latest available session; no published report attributed that move to any specific development.
The honest outlook is one of genuine tension rather than a clean verdict. Tencent's hardware diversification — domestic chip orders, Enflame backing, H200 deliveries, rural campus buildout — is substantive and reflects a company that learned tangible lessons from the supply shocks of 2022 to 2024. Its planned Malaysia cloud region extends the footprint into a fast-growing Southeast Asian market. But the talent signal is harder to dismiss: infrastructure without the researchers to fill it with competitive models is a necessary but not sufficient condition. Three things are worth tracking: whether the compensation gap between well-funded North American AI startups and Chinese hyperscalers continues to widen; whether Enflame and Yunbao can mature into genuine alternatives that meaningfully reduce Tencent's chip-supply concentration; and whether the Manus deal, if it closes, gives Tencent a credible platform in agentic AI to complement its LLM capacity — or turns out to be an acquisition whose delivery falls short of what the valuation implies.