Rather than getting cheaper, older NVIDIA graphics cards are getting more expensive, which speaks volumes about the state of the AI ​​market.

The price of used NVIDIA accelerators is suddenly an indicator of the health of the AI ​​business. According to Silicon Data, demand was so high that even aging GPUs retained their value, with the B200 selling for more than its original price after a year.

The NVIDIA B200 stands out the most. Its residual value is now about 158% of the starting value, which means the accelerator is actually trading at a premium of 58%. Silicon Data attributes this to shortages and continued demand, which continues unabated even a year after GPUs became widely available.

It’s a similar story with older A100s and H100s, with residual values ​​much higher than what would be expected with normal straight-line depreciation over three or five years. This is a rather unusual situation for the device – unlike the usual price drops, demand continues to support the cost of the accelerator.

The rental market confirms this trend. In January, B200 rental prices were below $5 an hour, while in August, prices rose to $5.50 to $5.80. One reason remains the B200’s effectiveness in inference tasks: when paired with DeepSeek R1, the processing cost is approximately $0.20 per million tokens, according to SemiAnaanalysis. The artificial intelligence craze appears to show no signs of slowing down.

Other publications on this topic