Refurbished Nvidia GPUs Fuel China’s AI Boom Amid Export Limits

Despite export restrictions and data security rules severing the flow of the most, latest technology from the West, China’s AI industry is discovering methods of continue moving forward. AI companies tapping into this growing demand for inference workloads are turning to refurbished Nvidia GPUs — rather than new accelerators — to meet their needs.

Nvidia GPUs Are the Heart of Chinese AI Clusters, Becoming Clusters of Refurbishments There are export limits against, and regulatory frameworks– both making it hard to source the latest H100 and H20 accelerators– so companies have been buying older A100 and recycled H100 units. These refurbished Nvidia GPUs are rebuilt by vendors and sold at competitive prices, creating an active recycling economy for AI hardware.

While training is well-known for its raw computational power requirements, inference workloads can operate with lower precision than full floating-point and require less raw computation power overall. As such, it causes used-up Nvidia GPUs to be much more worth than gold dust. An example of this is an A100 from 2020 that can still offer up to 80GB of HBM2e memory and ~2TBps of bandwidth, making it ideal to drive chatbots, recommendation engines and other inference-heavy workloads.

Nvidia’s software slack being primarily CUDA means that even secondhand Nvidia GPUs are still useful. For older cards, teams building model deployments have a tool for their applications without all of that expense dropping performance.

Unlike Legacy stuff, the H100 also has HBM3, and orders of magnitude higher throughput. Export-compliant H20 units, meanwhile, do exist but are three to seven times slower than a full H100 and over thirty times lower in double precision. That level of performance difference is leading many Chinese data centers to prefer refurbished Nvidia GPUs over sanctioned options, which of course is a better buy on the black market.

Refurbished Nvidia GPUs are in high demand as they offer a way to scale infrastructure now without having to wait for country approvals. While the reliability issues have raised some eyebrows, the capacity to scale such potential is something that matters more to companies in a race to deliver AI models at scale.

Nvidia itself is caught in the squeeze. As the company disclosed, it recently write-downs $5.5 billion of unsold H20 inventory, due to effects from changing the licensing and export controls. At the same time, the black market for refurbished Nvidia GPU is booming, extending the life of ancient silicon and keeping more modern technologies on the back burner.

Domestically, Chinese authorities are pushing alternatives — namely Huawei’s Ascend chips and Biren’s accelerators. Yet each refurbished Nvidia GPU turned on in domestic data centers erodes the market for indigenous kit. The proliferation of second-hand Nvidia GPUs sap some of the motivation to invest in Chinese chips.

Simply put, the export rules have produced a curious result: refurbished Nvidia GPUs are prolonging the life cycle of bygone hardware while transforming the Chinese AI economy. Less confined globally, refurbished Nvidia GPUs are driving growth on inference clusters based more on repurposed hardware.

FAQ