If you had to sell a GPU today, how much would it actually be worth? Certainly less than the price you paid for it. Nvidia releases or advances new product architectures every quarter, from Hopper to Blackwell to Rubin. Even considering price, GPU operators and tenants are likely to prefer phasing out older chips as their energy efficiency declines relative to newer generations. The rental market has already demonstrated this. An H100 chip launched at around $7 to $10 per GPU-hour, but within two years that price dropped to $2 to $4.
Larger capacity, higher energy efficiency, and faster speeds are all accelerating the price decline. This pressure almost guarantees that a chip purchased today through financing could be a whole generation behind before its corresponding loan is even amortized by half. To be fair, it's too early to judge the residual value of these chips and the demand that will arise over time. CoreWeave has reported that its H100 chips still command over 95% of their original rental price. This situation could also be an anomaly caused by chip scarcity.
Due to chip scarcity, previous-generation accelerators could very well continue operating efficiently in hyperscale cloud computing clusters for over five years. GPUs are tools, but their true value is determined by their use cases. In reality, their practical applications can be remarkably resilient.
The problem isn't that GPUs shouldn't be lent out, but rather that they are being lent out without understanding the residual value risk. We see market participants assuming a GPU's residual value after three years ranges from 10% to 60% of its original price. In other words, nobody really knows the answer. The risk lies not only in how quickly the asset depreciates, but also in the fact that chip suppliers themselves control the pace of technological obsolescence.
