Allegations of Nvidia Chip Usage in Chinese AI Models
Reports circulating on social media platforms have surfaced allegations that several prominent Chinese artificial intelligence models, including DeepSeek, Kimi, MiniMax, and Ling, were trained using Nvidia hardware. According to claims shared by a former OpenAI researcher on Reddit, these models utilized Nvidia chips despite ongoing U.S. export restrictions designed to limit China’s access to advanced semiconductor technology.
While these claims lack independent verification from official regulatory bodies, they have prompted discussions regarding the efficacy of current export controls. Some observers suggest the timing of these revelations may coincide with legislative efforts in the U.S. Congress to tighten chip control measures.
Investigations into Supply Chain Loopholes
The discourse surrounding these allegations has focused on potential workarounds used to bypass trade restrictions. Unverified reports suggest that Megaspeed, a company established in 2023 by Alice Huang, may have played a role in the procurement of advanced hardware. Originally incorporated in 2015 as 7Road International—an overseas spin-off of a Chinese online gaming company with ties to state-backed investors—the company was restructured under Huang Le Limited.
According to discussions on Reddit, Megaspeed became a significant customer for Nvidia in Southeast Asia, reportedly purchasing approximately 240,000 chips. Furthermore, unconfirmed reports suggest that investors tied to Huang may have participated in DeepSeek’s second funding round through special purpose vehicles, with alleged promises to supply the company with banned Nvidia Blackwell chips.
DeepSeek’s Hardware Diversification
Despite the allegations regarding Nvidia hardware, there are conflicting reports regarding DeepSeek’s technical trajectory. Unverified claims on social media suggest that DeepSeek is currently training a 2-trillion-parameter “V5” model using Huawei Ascend 950 chips rather than Nvidia GPUs.
This reported shift would utilize an “ultra-sparse Mixture-of-Experts” (MoE) architecture, a design intended to reduce compute requirements by activating only a small fraction of parameters at any given time. While this method may lower compute costs, it reportedly requires significant video random-access memory (VRAM) to maintain the full model. However, some technical observers noted that certain iterations, such as v4.1 flash, may already be operating on Ascend hardware, suggesting a bifurcated approach to hardware utilization within the company.
Regulatory Implications
The emergence of these reports has led to speculation regarding a potential “round two” of U.S. restrictions. Analysts on social media have suggested that the focus of future export controls may expand beyond hardware to include open-weights models and increased scrutiny of Chinese researchers working within U.S.-based AI firms. As of now, no official government action has been confirmed in response to these specific allegations.
