“AI compute is not currently fungible, as workloads cannot be easily moved between different types of chips, even from the same manufacturer like NVIDIA's H100 and GB200.”