DeepSeek discloses that a single GPU can generate $38 per hour in billable revenue, sparking attention on AI inference profitability.

Beating AI News Flash: DeepSeek recently open-sourced its inference framework DeepInfer and disclosed that its cluster can process 155 trillion tokens per day. Converted at DeepSeek API off-peak pricing, each GPU hour corresponds to approximately $38 in theoretical billing revenue; if calculated based on uninterrupted operation throughout the year, the annualized billing revenue exceeds $330,000.

DeepSeek improves inference efficiency through technologies such as expert parallelism, speculative decoding, asynchronous scheduling, and KV Cache optimization, but the above figures are theoretical billing revenue and do not equal actual revenue or profit. Actual returns also depend on factors such as GPU utilization, the proportion of paid requests, hardware and electricity costs, and service latency.

The AI industry is facing a situation where demand growth and profitability pressure coexist. On the one hand, demand for AI applications and Agent deployment among small and medium-sized enterprises continues to heat up; on the other hand, model training, inference services, and infrastructure investment are still eroding the profits of some companies. According to reports, CoreWeave's second-quarter revenue increased by about 112% year over year to $2.575 billion, but it still had a net loss of $626 million.

Analysts believe that improved inference efficiency is expected to expand the profit margins of AI services, but efficiently producing tokens does not necessarily mean commercial profitability. As model invocation volume grows, companies still need to address the balance among computing power costs, service pricing, and customers' willingness to pay.