DeepSeek Throughput Implies $38 in Theoretical Billings per GPU-Hour

Throughput disclosed for DeepSeek’s inference cluster implies about $38 in theoretical API billings per GPU-hour at off-peak prices. The estimate does not show how much money the service collects or whether it is profitable.

DeepSeek disclosed that the cluster processes 155 trillion tokens a day. Applying the hourly estimate to uninterrupted operation for a year yields more than $330,000 per GPU. Actual returns depend on GPU utilization, the share of requests that are paid, hardware and electricity costs, and service latency.

DeepSeek has open-sourced its DeepInfer inference framework. Techniques associated with improving inference efficiency include expert parallelism, speculative decoding, asynchronous scheduling and KV Cache optimization. Higher token output per GPU alone does not establish that inference services are profitable.