DeepSeek discloses that a single GPU can generate $38 in billing revenue per hour, drawing attention to the profitability of AI inference.

Beating AI News Brief: DeepSeek recently open-sourced its inference framework DeepInfer, and disclosed that its cluster can process 155 trillion tokens per day. Converted at DeepSeek API’s off-peak price, the theoretical billing revenue per GPU hour is approximately $38. If operated non-stop throughout the year, the annualized billing revenue would exceed $330,000. DeepSeek boosts inference efficiency via technologies including expert parallelism, speculative decoding, asynchronous scheduling, and KV Cache optimization. However, these figures only represent theoretical billing revenue, and do not equal actual operating income or profit. Actual returns depend on factors such as GPU utilization rate, proportion of paid requests, hardware and power costs, and service latency. The AI industry is grappling with a landscape of growing demand paired with profitability pressure. On one hand, demand for AI applications and agent deployments from small and medium-sized enterprises (SMEs) continues to surge; on the other hand, investments in model training, inference services, and infrastructure are still eating into the profits of some firms. According to reports, CoreWeave’s Q2 revenue rose roughly 112% year-on-year to $2.575 billion, yet it still posted a net loss of $626 million. Analysts note that improved inference efficiency is expected to widen AI service profit margins, but efficient token generation does not automatically translate to commercial profitability. As model call volumes grow, enterprises still need to strike a balance between computing power costs, service pricing, and customer willingness to pay.