[DeepSeek Discloses $38 in Billing Revenue Per GPU Per Hour, AI Inference Profitability Draws Attention] Beating AI News reports that DeepSeek recently open-sourced its inference framework, DeepInfer, and disclosed that its clusters can process 155 trillion tokens per day. Based on the DeepSeek API off-peak pricing, this translates to a theoretical billing revenue of approximately $38 per GPU per hour. If calculated based on continuous 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. However, the above figures represent theoretical billing revenue and do not equate to actual revenue or profit. Actual revenue also depends 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 of simultaneous demand growth and profit pressure. On the one hand, the demand for AI applications and agent deployments from SMEs continues to rise; on the other hand, investments in model training, inference services, and infrastructure are still eroding the profits of some companies. CoreWeave reportedly saw its second-quarter revenue increase by approximately 112% year-over-year to $2.575 billion, but still posted a net loss of $626 million. Analysts believe that improved inference efficiency could expand the profit margins of AI services, but efficient token production does not necessarily translate to commercial profitability. As model usage increases, the company still needs to address the balance between computing costs, service pricing, and customer willingness to pay. [Original Link]