Chinese Companies Turn to Nvidia Gaming Chips for AI Amid Desperation

Nvidia's gaming graphics cards are being repurposed by Chinese companies for AI model applications in a bold attempt to circumvent U.S. export sanctions. According to reports from the Financial Times, these gaming chips are being disassembled, with essential components stripped away and then reassembled onto new circuit boards. While these chips hold substantial power, they fall short in performing the high-precision calculations necessary for training large language models.

To adapt to limitations, Chinese firms are clustering these chips, yet this solution is far from ideal. 86Research analyst Charlie Chai referred to this approach as “desperate,” likening it to using a kitchen knife for fine art—a feasible task, but one that results in subpar outcomes. The challenges stem from the inherent design of gaming chips; although they offer enhanced performance for gaming, they are not optimized for managing large datasets typical in AI applications.

Recent data reveals a significant surge in the demand for Nvidia's gaming GPUs, with one factory reportedly disassembling over 4,000 units in December—more than quadrupling the amount handled in November. This surge follows the Biden administration's restrictions on the export of AI chips to China, which disproportionately affected Nvidia due to its dominant role in this market. Nvidia’s chips are vital for powering AI workloads, particularly in training models.

To address the restrictions, Nvidia developed China-specific hardware, including the 800-series chip, providing an alternative to circumvent the tough U.S. export rules. However, updates to these restrictions last October extended bans on advanced AI chips, ensuring that U.S. chipmakers cannot sell such technology to China or companies in jurisdictions likely to resell to it. Instead, Nvidia and other firms can only export slower systems, such as the three new units introduced last November that adhere to export controls.

Despite ongoing sanctions, the repurposing of Nvidia gaming graphics chips continues. Reports indicate that these repurposed chips are primarily being bought by public enterprises and smaller AI labs in China that could not secure a sufficient supply ahead of the sanctions. Although altering Nvidia’s products infringes on its intellectual property rights, the demand has driven companies to pursue these makeshift alternatives.

Among the most sought-after chips for this purpose are the GeForce RTX 4090 graphics cards. However, the 4090 is now prohibited from export to China, prompting the creation of a slightly less powerful, market-specific variant known as the GeForce RTX 4090 D. Yet, sources suggest that the 4090 D does not have the robust capabilities needed for effective language model training.

Nvidia has expressed its skepticism regarding the viability of disassembling gaming chips to create data center compute clusters for AI, emphasizing that these chips are originally designed for individual consumers and gaming applications rather than high-level data processing required for intensive AI tasks.