- Chinese AI centers are demolished from the GPU 4090D and sell them
- Inactive GPU resemption gives faster benefits than waiting for rental yields
- Some say it is overcapacity, others claim preparing for technology
Some Chinese AI data centers demonstrated and thank the NVIDIA RTX 4090D GPUs specific to China, according to reports.
A report of Digitimes Asia Said that these 48 GB cards, designed to bypass the American export restrictions on the GPU of play AD102 flagship, and initially deployed as part of the push of the IA infrastructure in China, are now taken from racks, renovated and sold on the free market.
The operators of the data center would have noted that this offers faster and more profitable yield than waiting three to five years to recover their investment via GPU rental.
An alternative theory
Each RTX 4090D sells between 20,000 CNY and 40,000 CNY (approximately 2,735 US at US $ 5,470), and even if they are slightly used, cards require a change in consumer resale.
As a rule, this means converting them with fan -style coolers to the blowing style cooler, which are better suited to dense but less effective server environments for single GPU use.
Digitimes Asia Said that this decision reflects a deeper financial pressure, many AI data centers find it difficult to keep the head out of the water in the middle of a low request.
According to the report, data centers require use rates of at least 70% to break – but current rates are often less than 20%. This leaves an expensive inactive sitting infrastructure while loan repayments are looming.
This is not an isolated case. As we have reported recently, the rapid expansion of the rapid IA infrastructure of China – encouraged by state policy – has led to a deaf.
Hundreds of data centers projects were launched across the country in 2023 and 2024, but real use has lagged behind. It is therefore not surprising that developers are now unloading equipment to reduce losses.
Although overcapacity is probably at least part of the reason behind the sale, some operators can simply clean up space for more recent technology.
With the interest passing from the formation of large -scale models to real -time inference, older training systems can no longer be as relevant as they were in the past.
Combined with the latest American export restrictions, which affect chips like the Nvidia H20, Chinese data centers will plan to unload old equipment and move to configurations ready for inference.
