- Chinese AI centers are dismantling 4090D GPUs and selling them off
- Reselling idle GPUs yields faster profits than waiting on rental returns
- Some say it’s overcapacity, others claim preparation for newer tech
Some Chinese AI data centers are dismantling and reselling China-specific Nvidia RTX 4090D GPUs, reports have claimed.
A report from DigiTimes Asia says these 48GB cards, designed to circumvent U.S. export restrictions on the flagship AD102 gaming GPU, and initially deployed as part of China’s AI infrastructure push, are now being pulled from racks, refurbished, and sold on the open market.
Data center operators are reportedly finding that this offers a faster and more profitable return than waiting three to five years to recover their investment through GPU rental.
An alternative theory
Each RTX 4090D sells for between CNY20,000 and CNY40,000 (around US$2,735 to US$5,470), and even if lightly used, the cards require modification for consumer resale.
Typically this means converting them from fan-style to blower-style coolers, which are better suited to dense server environments but less effective for single-GPU use.
DigiTimes Asia says this move reflects deeper financial pressure, with many AI data centers struggling to keep their heads above water amid low demand.
According to the report, data centers need utilization rates of at least 70% to break even – but current rates are often below 20%. That leaves expensive infrastructure sitting idle while loan repayments loom.
This isn’t an isolated case. As we reported recently, China’s rapid AI infrastructure expansion – encouraged by state policy – has led to overbuilding.
Hundreds of data center projects were launched across the country in 2023 and 2024, but actual usage has lagged behind expectations. It’s no surprise therefore that developers are now offloading hardware to reduce losses.
Although overcapacity is likely at least part of the reason behind the selloff, some operators may simply be clearing space for newer technology.
With interest moving from large-scale model training to real-time inference, older training-focused systems may no longer be as relevant as they once were.
Combined with the latest US export restrictions, which affect chips like Nvidia’s H20, Chinese data centers will be looking at offloading old hardware and shifting to inference-ready setups.
Leave a comment