Tencent says it has “fully adapted” its AI computing infrastructure to support Chinese-designed processors, in a move that shifts one of the country’s biggest buyers of Nvidia chips closer to home-grown hardware, as reported by SCMP. The announcement came at the company’s Global Digital Ecosystem Summit on September 16, where Tencent Cloud president Qiu Yuepeng confirmed the firm is now using “mainstream domestic chips” and building infrastructure around them.
While Tencent stopped short of naming the specific silicon in use, the phrasing suggests that production deployments are involved, not just experimentation. Senior executive vice-president Dowson Tong Tao-sang added that the company is working with “multiple domestic chip companies” to apply “the most suitable hardware” to each scenario, and that long-term strategic investment will focus on optimizing hardware-software co-design to lower the cost of compute.
Tencent’s announcement comes just a day after China’s State Administration for Market Regulation said Nvidia had violated antitrust rules and the terms of approval for its 2019 acquisition of Mellanox Technologies. The regulator did not elaborate but confirmed the investigation remains active. This adds another layer of uncertainty for U.S. firms selling into China’s cloud and AI sectors, which are already under tight export restrictions from Washington.
For Tencent, the company now has to factor in both geopolitics and supply continuity into its decision-making. Company president Martin Lau Chi-ping said in August that Tencent already has enough training chips in stock and “many options” for inference, suggesting that the firm has already diversified procurement. But adapting software to support non-Nvidia architectures is a deeper shift that Tencent appears to be leaning into, mirroring earlier signals from AI start-up DeepSeek, which said in August its V3.1 model was tuned for the next wave of domestic accelerators.
The most likely candidate for those deployments is Huawei’s Ascend platform, which has already been adopted at scale by ByteDance and is supported by an increasingly mature stack built around the MindSpore framework. But whether Ascend or other domestic chips can sustain large-scale training remains an open question, especially as U.S. officials estimate that Huawei will only be able to produce around 200,000 AI chips next year.
