曙光8000超算发布会及投资者反馈

Hygon Information Technology Co., Ltd.| AsiaPacific Morgan Stanley Asia Limited+ Daisy Dai, CFAEquity AnalystDaisy.Dai@morganstanley.comMorgan Stanley Taiwan Limited+ Sugon 8000 launch andinvestor feedback Charlie ChanEquity AnalystCharlie.Chan@morganstanley.com Daniel Yen, CFAEquity AnalystDaniel.Yen@morganstanley.comMorgan Stanley Asia Limited+ China launched its first 100,000-card AI supercluster usingHygon's CPUs, GPUs and switch chips. We also addressquestions about Hygon's differentiation, valuation and theagentic AI opportunity. Ethan JiaResearch AssociateEthan.Jia@morganstanley.com Henry ZhaoResearch AssociateHenry.Zhao@morganstanley.com Hygon's CPUs and GPUs used in China's first 100,000-card AI supercluster: Sugon launched Sugon 8000, a 100,000-accelerator AI supercluster, in Zhengzhoulast Friday. All six core chip types in the cluster were developed domestically by theSugon/Hygon ecosystem, including the Hygon C86 CPU, Hygon DCU (GPGPU AIaccelerator), HySW scale-up switch ASIC (interconnecting GPUs within the same theboard), PCIe 5.0 switch chip (interconnecting CPUs and GPUs), 400/800G switchASIC, 400G NIC ASIC. For more see Key Takeaways from HIECO Conference. Hygon Information Technology Co., Ltd. (688041.SS, 688041 CH) Greater China Technology Semiconductors|China First, the differentiation of Hygon vs. other GPU companies. Hygon’s integrated“CPU + GPU” model differentiates it from most local peers that focus primarily onGPUs, allowing Hygon to address both general-purpose server workloads and AItraining/inference simultaneously. Also, its largest shareholder Sugon (27.96%shareholding) is not only a supercomputer manufacturer but a key player in China AIinfrastructure, spanning chips, computing, storage, networking, cooling, applicationsand services. We think Hygon could leverage its largest shareholder's ecosystem andgain more share in China's CPU and GPU markets. Second, why does Hygon deserve a high valuation like other GPU companies? Some investors are reluctant to give a high valuation to Hygon's CPU business, asmost CPUs are not restricted from import and CPU TAM growth seems slower thanGPUs. We argue that China's server CPU TAM growth, including agentic AI demand,will be higher at a 31% CAGR ( Exhibit 2 ) than AI GPU TAM growth at 23% ( Exhibit4 ). Also, the CPU localization drive is coming from policy and procurementmandates, and supply-chain security. Therefore, we don’t think Hygon’s CPUbusiness should be valued at lower multiples than other GPU companies Morgan Stanley does and seeks to do business withcompanies covered in Morgan Stanley Research. As a result,investors should be aware that the firm may have a conflict ofinterest that could affect the objectivity of Morgan StanleyResearch. Investors should consider Morgan StanleyResearch as only a single factor in making their investmentdecision. Third, how should we think about Hygon's positioning in the agentic AI era?AsAI systems evolve toward multi-agent architectures, CPUs are increasinglyresponsible for workflow control and system-level coordination. We estimateChina's agentic AI CPU TAM to be US$17bn by 2030, or 21% of global demand. Weexpect Hygon to address 3-5% of that demand as its CPUs' presence in security-sensitive verticals like government and finance aligns well with emergingrequirements for data sovereignty, compute isolation, and trusted infrastructure. Weexpect Hygon’s next-generation CPUs to deliver higher clock speeds, strongerperformance, and enhanced security to meet the demands of agentic AI. For analyst certification and other important disclosures,refer to the Disclosure Section, located at the end of thisreport. += Analysts employed by non-U.S. affiliates are not registeredwith FINRA, may not be associated persons of the memberand may not be subject to FINRA restrictions oncommunications with a subject company, public appearancesand trading securities held by a research analyst account. Key Takeaways from HIECO Conference Core chips in China's first homegrown 100,000 card AI supercluster are localized: Sugon launched Sugon 8000, a 100,000-accelerator AI supercluster whose deploymentwas completed in June. All six types of core chip in the cluster were developeddomestically by the Sugon/Hygon ecosystem, including the Hygon C86 CPU, Hygon DCU(GPGPU AI accelerator), HySW scale-up switch ASIC (interconnecting GPUs within thesame the board), PCIe 5.0 switch chip (interconnecting CPUs and GPUs), 400/800Gswitch ASIC, 400G NIC ASIC. Sugon's scaleFabric 400 scale-out networking platform isbuilt around the network switch and NIC ASIC, with proprietary 112G SerDes IP. The AI supercluster enables highly efficient training and inference for massive models.The system supports 64-bit, 32-bit, 16-bit, and 8-bit (incl. FP8 and INT8) formats, as wellas mixed-precision execution, with vendor-reported benchmark results indicatingcompetitive performance across these precision formats. ParaStor, Sugon's d