How ZhiZi XinYuan’s AI‑Driven Compute Platform Is Disrupting the Industry After Two Funding Rounds

Throughout history, breakthroughs in science have been driven by the ability to compute ever more complex problems, from planetary orbits to molecular interactions and modern AI models. Today, the rapid growth of large‑model, agent‑based, and scientific‑computing workloads pushes compute demand to a new scale, but hardware scaling alone cannot deliver linear efficiency because of process, power, and cost constraints.

In this context, the "AI for Computing" field aims to let AI automatically optimize the compute stack itself. ZhiZi XinYuan, founded in August 2025, adopts a paradigm that combines a large‑model engine, operations‑research optimization, and automated algorithm discovery. The company builds an intelligent agent that can understand a task, explore the full software‑hardware design space, and iteratively validate candidates on real hardware, turning theoretical peak performance into practical, delivered compute power.

The need for such a system stems from three observations: (1) rapid evolution of hardware architectures, compilers, inference frameworks, and networks creates constant adaptation and tuning challenges; (2) expert talent that can simultaneously master algorithms, systems, and hardware is scarce, making manual, iterative optimization inefficient; and (3) modern AI workloads are increasingly dynamic and fragmented, requiring end‑to‑end system efficiency rather than isolated kernel improvements.

ZhiZi XinYuan’s technical pipeline is described in three steps. First, the AI must "see" the compute task by decomposing it into measurable metrics such as latency, throughput, and power, and identifying bottlenecks in memory access, scheduling, or kernel implementation. Second, it performs automatic search and algorithm discovery across a massive implementation space, leveraging large‑model generative capabilities to propose candidate solutions and operations‑research models to schedule resources under complex constraints. Third, the proposed solutions are validated on real chips, allowing hardware feedback to close the loop and transform experience‑based engineering into automated engineering.

These ideas are embodied in the product KernelCAT, an "automatic compute‑acceleration platform" that translates natural‑language user requirements into an executable optimization workflow. KernelCAT follows a four‑stage loop—analysis, code generation, on‑board tuning, and delivery—automatically handling model, operator, graph, workload, target hardware, and performance goals. The platform eliminates repetitive manual steps such as documentation lookup, code rewriting, compilation, profiling, and parameter tuning for each new model, framework, or hardware target.

Within the KernelCAT family, the Kerminal subsystem demonstrates strong automatic acceleration capabilities. It has achieved state‑of‑the‑art results on the KernelBench GPU kernel‑optimization benchmark (top scores in accuracy, average speed‑up, and geometric mean speed‑up) and on CANN‑Bench, where it completed profiling for 50 of 53 tasks, passed 35 fully, and achieved a 95 % pass rate with only one error. Kerminal can also autonomously replace inaccurate implementations with polynomial approximations to meet precision requirements, showing the ability to explore new algorithmic paths without human prompts.

Concrete industry cases illustrate the impact: on the RDK S100 development board, deploying DeepSeek R1 1.5B reduced end‑to‑end time to two hours and improved throughput by 1.5×; in AI‑for‑Science, TorchFold on Ascend chips cut peak memory by 70 % and increased speed by 50 %; and the DSDP molecular‑docking model saw a 138× inference speed‑up after migration from CUDA to the Kunpeng platform. These successes indicate that KernelCAT’s automated pipeline can be reused across platforms and workloads.

The article concludes that as AI penetrates deeper into the acceleration process, the next scarce resource will be the ability to traverse multiple system layers and find optimal implementations for complex business scenarios. ZhiZi XinYuan’s approach, backed by recent angel‑plus financing led by Dingfeng Kechuang, InnoTech Capital, and Shoucheng Capital, positions the company to shape the long‑term value of the AI‑for‑Computing ecosystem and to turn previously unattainable scientific and industrial workloads into practical, scalable solutions.