Tencent President Discusses Significant Stockpiling of AI GPUs – Open to Future Adoption of Native Designs

Martin Lau, President of Tencent, has divulged that his company has accumulated a “pretty strong stockpile” of NVIDIA AI chips. In a mid-week earnings call, the Chinese executive reckoned that this surplus will come in handy—upon the company unleashing its full-on upcoming “AI strategy.” Lau was responding to a question regarding ripples caused by a recent introduction of revised licensing requirements for “high-end GPUs.” His lengthy reply seems to align with “leaked April time” information; when industry analysts theorized a massive $16 billion spend—reportedly, big Chinese tech firms had splurged out with swift acquisitions of NVIDIA H20 GPUs. Lau commented on present day conditions: “it’s actually a very dynamic situation right now. Since the last earnings call, we have seen an H20 ban, and then after that there was the BIS new guidelines that just came in overnight…If you look at the allocation of the usage of these chips, obviously they’ll be used for the applications that will generate immediate returns for us. For example, in the advertising business as well as content recommendation product, where we actually would be using a lot of these GPUs to generate results and generate returns for us. Secondly, in terms of the training of our large language models, they will be of the next priority and the training actually requires higher-end chips.”

Team Green’s engineering team has likely been strong-armed into designing further compromised hardware; as “exclusive” sanction-conforming options for important enterprise customers in China. Tencent seems to have enough pre-ban specimens to tide things over, for a while. The firm’s president envisioned a comfortable position, for the foreseeable future: “over the past few months, we (started) to move off the concept or the belief of American tech companies—which they call ‘the scaling law’—which required continuous expansion of the training cluster. And now we can see even with a smaller cluster you can actually achieve very good training results. And there’s a lot of potential that we can get on the post-training side which do not necessarily meet very large clusters. We should have enough high-end chips to continue our training of models for a few more generations going forward.” Huawei’s controversial Ascend 910C AI accelerator seems to be the top alternative contender; tech watchdogs believe that this design’s fortunes will be closely tied to the rising dominance of DeepSeek. Fairly recent leaks have indicated impressive progress being made within China’s domestic AI accelerator infrastructure.

Lau only alluded to alternate hardware avenues, with his usage of very general terms—be it through local channels, or slightly further afield: “we actually can potentially make use of other chips; compliant chips available in China or available for us to be imported, as well as ASICs and GPUs in some cases for smaller models inferences. So I think there are a lot of ways to which we can fulfill the expanding and growing inference needs, and we just need to keep exploring these venues, and spend probably more time on the software side, rather than just force a buying of GPUs.”