"We're targeting the AI inference market – that's where the real trillion-dollar opportunity lies. Training is massive in scale, but inference will be used in every scenario. People are running out of available inference computing capacity." This statement from Oracle founder Larry Ellison during an earnings call, supported by the surge in remaining performance obligations (RPO) to $455 billion, helped him briefly claim the title of world's richest person overnight.
As early as 2024 when AI applications exploded, Larry Ellison had repeatedly predicted that "inference will be used for everything." AI training demand is typically cyclical and high-intensity resource consumption, usually occurring on massive GPU clusters either one-time or periodically. Only AI inference demand represents the continuous invocation after models are "productized and servicized" – triggered constantly every day, every second, across millions of users or hundreds of automated systems. This normalized, large-scale resource utilization will drive sustained growth in AI cloud services.
Recently, QINIU (02567) announced in its financial report that AI-related revenue reached 184 million yuan, contributing 22.2% of total revenue. Related business executives indicated that the company's AI-related revenue is primarily concentrated in AI inference services and computing resources. Building on the foundation of AI-related users exceeding 10,000 in early August, QINIU's AI-related user base has recently reached 15,000, benefiting from access to over 50 callable large models covering LLM inference models, tool invocation, AI programming, and inference interface support including Claude CodeAI functions.
Handling AI inference demand is not easy, requiring continuous reduction of end-to-end latency for model inference requests in production environments, improving throughput rates, and rapidly responding to user or business system requests. Under high request pressure (QPS – queries per second) and TPM throughput requirements, inference computing demand far exceeds training needs.
Furthermore, since inference models need to provide useful, actionable answers, they require high-quality, connectable enterprise data. Enterprise-level and vertical industry-level massive structured data assets are key resources for entering the "inference era."
Benefiting from QINIU's 14-year accumulation in audio/video cloud services, the company's low-latency, high-throughput global real-time nodes, massive storage capabilities, and technologies that safely "expose" private audio/video heterogeneous data to inference models through vectorization and Private LLM access, these capabilities will all contribute to QINIU's AI cloud services as a second growth curve. By occupying dual positions in the inference computing value chain – upstream data provision and midstream computing infrastructure – the company is positioned to capture substantial long-term inference computing revenue and service fees, achieving dual growth in both service volume and pricing.
Comments