Before Groq’s LPU rode Nvidia’s spotlight to fame, Yuanchuan Micro founder and CEO Yang Bin was already sketching an inference chip. LeiFeng Net has learned exclusively that the LPU start-up has closed a several-hundred-million-yuan Pre-A round led by IDG Capital, with Futong Capital, JK Venture, Shangqi Capital, Shunxi Fund and Xuhui Sci-Tech Innovation Investment joining. It is the company’s fourth raise in 2026, and its next round is already underway.
Yuanchuan was founded in September 2025. An angel round earlier this year, also several hundred million yuan, drew top VCs, government funds and supply-chain backers. The continued capital reflects a new chip race in the inference era, stretching from GPUs to purpose-built architectures. On chip form alone, spatiotemporal compiler, hardwired pipeline, high-bandwidth memory, Yuanchuan looks like a “Chinese Groq”.
A Huawei veteran reasoning backward from AI’s endgame
Yang rejects the simple label. “If our only goal were to imitate an LPU, the room to expand would be small,” he says. His thinking is shaped by more than twenty years at Huawei, where he built the processor team in the US from scratch, then led wireless baseband algorithm and chip work in China, owning large-scale hardware scheduling and data architecture. “Starting up is cashing in every resource you accumulated,” he says.
The core team is the bedrock: CTO Dr Sun has studied GPGPU, GPU and DSP architectures for over thirty years; chief software engineer Dr Will brings more than twenty years in compilers, EDA and heterogeneous systems. Rather than debate “domestic substitution” or “value for money”, this crew of veterans asks what inference compute should be reorganised around once AI enters the agent era.
In early 2025 Yang and his CTO saw that model capability matters less than the value created after deployment. DeepSeek R1 trained near-top performance at far lower cost, Manus opened the general-agent imagination, and Huang Renxun made inference and agents the centre of GTC with a full solution for the inference era. A hundred-trillion-token study showed inference models already handle about half of OpenRouter’s tokens, and agentic inference is the fastest-growing interaction.
LPU+, not a copy of Groq
An agent task chains planning, tool calls and verification, dozens to hundreds of inferences, and any single delay compounds down the chain, like a butterfly effect. Yang argues the classic GPU, built on von Neumann, never matched today’s compute pattern, while full ASIC loses the flexibility models keep demanding. The new architecture must decide what to harden and what to keep soft, maximising the combined efficiency of compute, memory and bandwidth.
Yuanchuan names its architecture LPU+, designed from the start for multimodal models and optimised for the dynamic scheduling of Mixture-of-Experts models. It is not a language-only unit. Yang credits Groq’s real lesson, its persistent hardwired pipeline, which avoids resource contention and idle data waits and enables large on-chip SRAM. He compares the spatiotemporal compiler to drawing a high-speed rail timetable: every train’s departure, stop and dwell is planned in advance, so at run time the chip just executes. The three cores, compiler, pipeline, memory, are inseparable; the first two set theoretical performance, the compiler sets how much is actually realised.
At WAIC 2026 the architecture debuted as an “Agentic AI Token factory” supplying bottom-layer compute for enterprise inference. For Yang, the Huawei lesson was never one technique but a systems mindset that derives technology choice from industrial value and business loop. Every step since has tested it.
Editor’s note: This is an adapted translation of the original LeiFeng Net report. It has been trimmed and restructured for readability for an international business audience.
