On August 8, 2026, Zhuhai Xile Interactive Entertainment Network Technology Co., Ltd. launched its HeteroFlow v2 inference service, an open-source heterogeneous computing scheduling platform that unifies management of GPUs from nine manufacturers, including NVIDIA, Huawei Ascend, and Hygon DCU, along with five inference engines, through a single OpenAI-compatible API. The platform supports hybrid deployment, automatic failover, model hot-loading, and rolling updates, aiming to boost GPU resource utilization from 30% to above 80%. Its release follows DeepSeek’s August 6 announcement of plans to raise API pricing across the board, as China’s large model industry shifts from low-price subsidies to cost recovery. According to analysis reported by NetEase, agent scenarios have driven single-task token consumption to dozens of times that of traditional conversational AI, creating computing cost pressures that can no longer be absorbed through financing and subsidies. HeteroFlow v2 is fully compatible with the official OpenAI API protocol, requiring developers only to modify the request URL and API key, with existing Python, JavaScript, Go, or Java code unchanged, enabling rapid migration to heterogeneous computing clusters and avoiding vendor lock-in. The platform targets enterprises seeking cost reduction, efficiency gains, and a transition to domestic Chinese hardware alternatives, positioning it to benefit from computing prices returning to fair value.