半导体：从比伦科技参观中汲取的启示
Takeaway fromBirenTechvisit WevisitedBirenTech's managementlast week. Biren (6082 HK, NR) is aGPGPU designer of training-and-inference accelerators,operating a fablessmodelwith localized supply chain.Three key takeaways from our visitinclude:1)Biren’s2nd-genBR20X product series,with pre-silicon designcompleted in 2025 and commercial launch expected in 2026;2) a broadeningcustomer baseacross AI computing platforms,telecom operators, commercialAIDC providers, and AI / LLM companies;and3)FY25 revenue of RMB1.03bnwith 2H25 contributing ~RMB0.98bnat GPMof55.2% (vs. 1H25's 31.9%), China Semiconductors Sector Kevin ZHANG(852) 3761 8727kevinzhang@cmbi.com.hk Aaron GUO(852) 3916 3715aaronguo@cmbi.com.hk Company profile:Biren designs general-purpose GPU chips andGPGPU-based intelligent computing solutions for AI training andinference.Biren reportedrevenue of RMB1,035mnin FY25(+207% YoY), Product portfolio: The launched BR10X portfoliocomprises:1)BR106(single-die, four form-factor variants 106L/M/B/C),2)BR166(2.5D chipletintegrating 2x BR106 dies plus 4xDRAM,three variants 166L/M/C), and3)BR110(edge-inference SKU sharing the BR10X architecture).BR166L/Mentered mass production in August 2025 and BR166C in December2025.2nd-genBR20X has completed pre-silicon design in 2025 and iscurrently in the process of physical design and verification for tape-out Read-across on China GPU localisation: The broader industry backdropisconsistent with the direction of travel on domestic AI-computelocalisation. According to Reuters citing IDC, domestic vendors collectivelyshipped 1.65mn AI accelerator cards in China in 2025, capturing ~41% ofthe ~4.0mn-unit market. NVIDIA retained market leadership with ~2.2mn Company Overview Potential2nd-genBR20Xlaunch as a key catalyst Biren completed BR20X pre-silicon design in 2025, with commercial launch expected in2026. The1st-GenBR10X line was in volume shipment through late 2025, with BR20X's2026 launch setting up the1st-gen/2nd-genhand-off. 1.Architecture roadmapprogression and software ecosystem depth: Firstgen architecture:The BR10X base architecture combines four disclosedinnovations: 1) SIMT (Single Instruction Multiple Thread) for general-purpose flexibility;2) a specialized T-core tensor engine with data recycle to reduce DRAM bandwidth Secondgenarchitecture: BR20X extends along four vectors: 1) native FP8/FP4precision; 2) larger/faster memory; 3) higher interconnect bandwidth via proprietary Thirdgenoutlook:BR30X (cloud) and BR31X (edge) target 2028 commercial launch,andarein the process of conducting feasibility analysis and preliminary R&D, per Software stack: BIRENSUPA native support spans PyTorch/vLLM/SGLang, extendedto Triton/TileLang, with Day-0 adaptation for DeepSeek V3/R1, MiniMax M2, ZhipuGLM, Alibaba Qwen, StepFun Step, and Tencent Hunyuan, per mgmt.. Source:Companydata,CMBIGMNote:”1+1+N+X”:“1”GPU architecture+“1”unified software platform,deriving”multiple”chips and a 2.Customer base broadening across AIDCs,telecom operators and LLMcompanies Biren's customer base now spans national-level AI computing platforms, telecomoperators, commercial AIDC providers, AI/LLM companies, and enterprises, with AI FY25 top-five customerscontributed RMB738.1m combined (~71.3% of revenue),withtop customer concentration down to ~19% from ~54.5% in FY24. 3.Super-node delivery track record from 1,024-GPU to 2,048-card Management highlighted two super-node deployments that validate capability at scale. The Nanjing 1,024-GPU cluster (Sep-24, RMB180m contract, 1,024 BILI 106M across128 servers on 8-GPU BLink full-mesh UBB) delivered 95% linear scalability for large- In 2025, Biren launched the "LightSphere X" SuperPod built on the BR10X series withecosystem partners, using optical interconnect and distributed optical circuit switching(dOCS); a 2,048-card GPU SuperPod of this design has been deployed at a national- 4.Increase inR&D intensity;stablegross marginthrough scale-up FY25 R&D expense of RMB1,476m (+78.5% YoY)was~143% of total revenue,reflecting sustained investment in next-generation product and technology iterationahead of the2nd-gencycle. FY25 gross margin landed at 53.8%, +63bps YoY, on Disclosures& Disclaimers Analyst Certification The research analyst who is primary responsible for the content of thisresearch report, in whole or in part, certifies that with respect to the securities or issuerthat the analyst covered in this report: (1) all of the views expressed accurately reflect his or her personal views about the subject securities or issuer; and (2)no part of his or her compensation was, is, or will be, directly or indirectly, related to the specific views expressed by that analyst in this report.Besides, the analyst confirms that neither the analyst nor his/her associates (as defined in the codeof conduct issued by The Hong Kong Securities and FuturesCommission) (1) have dealt in or traded in the stock(s) covered in this research report within
