- Chinese centers of AI are dismantling the 4090d GPUs and sell them
- The inactive GPU revolution obtains faster profits than waiting in rental returns
- Some say it is overcapacity, others claim the preparation for the newest technology
Some Chinese AI data centers are dismantling and reluctanting the NVIDIA RTX 4090D Specific China, according to reports.
A report from Digitimes Asia He says these 48 GB cards, designed to avoid the export restrictions of the United States at the Ad102 flagship GPUs, and initially deployed as part of the Chinese infrastructure impulse of AI, are now being removed from the shelves, restored and sold in the open market.
According to reports, data center operators discover that this offers faster and profitable performance than waiting for three to five years to recover their investment through the GPU rental.
An alternative theory
Each RTX 4090D is sold between CNY20,000 and CNY40,000 (around US $ 2,735 to US $ 5,470), and even if used slightly, cards require modification for consumer’s resale.
Usually, this means converting them from ventilator fan -style refrigerators, which are more suitable for dense but less effective server environments for the use of a single GPU.
Digitimes Asia He says that this movement reflects a deeper financial pressure, with many AI data centers that fight to keep their heads above the water in low demand.
According to the report, data centers need at least 70% use rates to reach the equilibrium point, but current rates are often less than 20%. That leaves an inactively expensive infrastructure while loan payments are closed.
This is not an isolated case. As we recently reported, the rapid expansion of China’s infrastructure, encouraged by state policy, has led to overcompilation.
Hundreds of data centers throughout the country were launched in 2023 and 2024, but the real use has been left behind expectations. It is not surprising, therefore, developers are now downloading hardware to reduce losses.
Although overcapacity is likely to be at least part of the reason behind the sale of the sale, some operators can simply be cleaning the new technology space.
With the interest of moving from large -scale model training to real -time inference, the oldest training systems can no longer be as relevant as before.
Combined with the latest US export restrictions, which affect chips such as NVIDIA H20, Chinese data centers will seek to download old hardware and change to configurations ready for inference.
