Qingmao Intelligence Unveils Self-Evolving AI Infrastructure for Domestic Chips

The report highlights a broader shift in AI competition toward recursive self-improvement (RSI), where AI systems optimize themselves. OpenAI's GPT-5.6 has used RSI-related capabilities to cut end-to-end service costs by 20% and boost token generation efficiency by over 15%, according to the report. Anthropic disclosed that more than 80% of the code in its codebase was autonomously generated by Claude. Jeff Dean, after leaving Google, founded Discovery Loop to focus on RSI; latest market reports value the venture at around $10 billion.

Qingmao's approach stems from a mismatch: while upper-layer models are moving toward L4-style autonomous execution, the underlying compute layer still relies heavily on manual engineering. In China, restricted access to advanced foreign chips has increased reliance on domestic processors. But domestic chips require time-consuming operator adaptation, compilation optimization, memory scheduling, framework adaptation, and inference tuning. A single complex operator can take weeks for a senior engineer to develop and tune, only to be outdated by fast-moving model and chip iterations. The company says this manual bottleneck leaves much of domestic compute capacity underutilized.

To address this, Qingmao builds on research from the laboratory of Zhu Wenwu, a professor at Tsinghua University's Department of Computer Science and Technology. In 2022, the lab formally proposed the concept of self-driven machine learning, extending machines' autonomy to tasks, data, models, optimization strategies, and evaluation criteria. Guan Chaoyu, a student of Zhu and winner of Tsinghua's top award, founded Qingmao and led teams to championships in international competitions such as NeurIPS and AAAI, defeating teams from Harvard, MIT, and Stanford in AutoML and MetaDL challenges.

At the engineering level, Qingmao decomposes its system into two core abilities: autonomous optimization and autonomous evolution. The former automatically locates bottlenecks and tries different options based on real runtime status; the latter accumulates experience from each optimization cycle and reuses it on new models, chips, and environments. This creates a closed loop of optimize-accumulate-reoptimize, making the system increasingly better at optimizing itself.

The technology has been validated in production environments using domestic chips. Qingmao says a complex operator that once took a senior expert about two weeks to develop and tune can now be handled by its autonomous optimization system in about one hour, a roughly 336-fold improvement. Full-chain enablement of a domestic chip has been cut from about one month to between one day and one week, about 30 times faster. Some representative operators achieved up to 30x acceleration over baseline versions, reaching an industry-leading level. To date, Qingmao has adapted and optimized dozens of domestic chip models, covering more than half of mainstream domestic compute chips in China, making it one of the domestic AI Infra vendors with the most chip adaptations and performance optimizations.

Qingmao's broader goal is a trillion-level token factory based on domestic chips. In 2026, the boom in AI agents has driven token consumption sharply; some complex agents use hundreds of times more tokens per task than traditional conversational models, and daily calls in many business scenarios are approaching trillions. Qingmao believes future video models, multimodal systems, AI for science, embodied intelligence, and physical models will push token demand even higher. The company offers two delivery models: selling high-cost-performance tokens outright, or licensing its self-evolving token production platform. It also leases large-scale domestic compute capacity and provides deeply integrated private deployment or managed operations, aiming to build a domestic compute ecosystem spanning production, adaptation, scheduling, distribution, and applications.

The company stresses that the key change is that adaptation itself becomes a continuous process. When a new model arrives, the system automatically learns; a new chip triggers automatic tuning; a new business scenario does not start from zero. Optimized models and agents, in turn, feed back into the compute base, raising the ceiling of automated infrastructure. As Qingmao puts it, AI optimizes infrastructure, and infrastructure empowers AI.