元川微 (ArcheFlowX)
Category: AI Chips / Semiconductors
Chinese LPU (Language Processing Unit) inference chip startup building native architecture for Agentic AI real-time inference 元川微 (ArcheFlowX) was founded in 2025. The company is led by 杨滨 (Yang Bin). Based in Hangzhou, China. Team size: 101-500. Total funding raised: $100.0M. Latest round: Pre-Series A. Key investors include IDG Capital, 孚腾资本, 九坤创投, 尚颀资本, 顺禧基金, 徐汇科创投, 东方嘉富, 元禾原点, 峰瑞资本, 源来资本, 中芯聚源, 深创投.
- Founded
- 2025
- Headquarters
- Hangzhou, China
- Team size
- 101-500
- Total funding
- $100.0M
Value proposition
Ultra-low latency AI inference chips based on proprietary LPU+ architecture, purpose-built for Agentic AI workloads across cloud, edge, and endpoint devices
Products and solutions
LPU+ architecture inference chips (Mountain series for data center/high-end edge, River series for endpoint/edge), Agentic AI Token Factory (unveiled at WAIC 2026), Spatiotemporal compiler, Hard dataflow architecture, Large-bandwidth on-chip SRAM
Unique value
First native Chinese LPU architecture designed from the ground up for Transformer/LLM/Agentic AI era (vs Groq's CNN-era LPU); native MoE support; multi-modal optimization; cloud-edge-endpoint unified architecture without redesign
Target customer
Cloud service providers, data center operators, AI application developers, edge computing companies, robotics and autonomous driving firms
Industries served
AI infrastructure, Cloud computing, Edge computing, Autonomous driving, Robotics, Smart manufacturing, Smart home
Technology advantage
Proprietary LPU+ architecture with three core technologies: spatiotemporal compiler (static scheduling), hard dataflow architecture (ASIC-ized deterministic execution), large-bandwidth on-chip SRAM; company/investor materials cite ~6x H100 token speed / 75% lower token cost / ~1/3 energy as target or FPGA-validation benchmarks (public JWIPC IR also attributes 6x/75%/1/3 figures to Groq ISSCC data with 元川微 FPGA results said to be comparable); team with deep Huawei wireless baseband chip expertise (dataflow architecture DNA)
How they differentiate
Unlike Groq (designed for CNN era, 2016), LPU+ is natively designed for Transformer/LLM/MoE/multi-modal models; supports cloud-edge-endpoint scaling without chip redesign; Chinese domestic supply chain; industry investors doubling as early customers
Main competitors
Groq (US; ~$20B NVIDIA non-exclusive LPU tech license + acquihire Dec 2025; Groq remains independent), 迈特芯 (MetaChip / szmetachip, China edge LPU), 深明奥思 (Semiotics AI, China LPU)
Key partnerships
IDG Capital (lead Pre-A), 星宸科技 (Sigmastar, 301536) - strategic investor & customer, 智微智能 (JWIPC, 001339) - strategic investor & customer, 中芯聚源 (CIC Pinnacle Capital), 深创投 (SCGC), 东方嘉富, 元禾原点, 峰瑞资本 (FreeS Fund), 浙江省科创投, 杭州润苗基金
Notable customers
星宸科技 (Sigmastar, 301536) - strategic customer, 智微智能 (JWIPC, 001339) - strategic customer, Multiple cloud service providers and industrial customers in product validation phase
Major milestones
2025-09: Company founded, 2026-01: First Angel round closed, 2026-03: Strategic investment from JWIPC, 2026-04: Angel series round (hundreds of millions RMB), 2026-07: Pre-A round led by IDG Capital, 2026-07: Agentic AI Token Factory unveiled at WAIC 2026, Planning first chip tape-out in H1 2027
Growth metrics
Completed FPGA verification of LPU+ architecture (all performance metrics met design targets); planning first tape-out in H1 2027; completed 4 funding rounds in 2026; unveiled Agentic AI Token Factory at WAIC 2026
Market positioning
China's first and leading native LPU inference chip startup; positioned as the "Chinese Groq" but with next-gen architecture purpose-built for Agentic AI; backed by top-tier VCs and strategic industry partners
Geographic focus
China (primary), with potential expansion to global markets as LPU architecture gains industry adoption
Patents and IP
Proprietary LPU+ architecture with spatiotemporal compiler, hard dataflow pipeline, and large-bandwidth SRAM integration; FPGA verification completed with all performance metrics meeting design targets
About 杨滨 (Yang Bin)
Ex-Huawei 22-year veteran; 2008 founded Huawei's US processor team from scratch; 2012 returned to China to lead Huawei's wireless baseband algorithm and chip division; Graduated from Xidian University (西安电子科技大学)
