【英文】瑞信报告：中国CPU国产替代与人工智能加速（84页）

英文研究报告 2020年08月13日 06:24 管理员

IC design service companies are benefiting from the trend of rising IC design difficulty as technology migrates, with design cost surging from US$26 mn on 28nm to US$52 mn on 16nm and US$130 mn on 7nm. To ensure successful tape-out and secure foundries' capacity, design service companies could fill the gap, supplying design support and foundry design-in to system companies and start-ups designing chipsets on the advanced nodes. The trend of IC customisation and higher design cost has accelerated the growth for major Taiwan IC design service companies at a 9% CAGR from 2013-19, in line with global fabless, following the slow period during 2007-13 (1% CAGR vs global fabless' 9%), when the market was dominated by the standardised chipsets (e.g., main processors in PC and smartphone). Design service leveraged to China's CPU localisation Design service companies could support China’s more aggressive stance in recent years to develop its own processors to capture more value in the tech chain and have better control of the IT ecosystem amid its ambition of ‘Made in China 2025’ and the growing trade tension with the US. We believe the PC/NB CPU used in the government applications and SOE is the first wave of chipsets China aims to replace (5 mn unit demand annually), and higher-end server CPU should be next (3 mn unit annual demand). We expect Phytium, a China ARM CPU fabless, to lead its local peers on the robust roadmap, migrating to 16nm for PC/NB and 7nm for server, benefiting its design service partner, Alchip supplying a US$720 mn TAM (vs its US$144 mn sales in 2019). Alchip also has exposure to China’s supercomputing system which should see content per system growing from US$30 mn on 16nm to US$50 mn on 7nm.

Growing AI computing drives ASIC demand Design service companies are also seeing new growth from AI chipset innovation accelerating to support demand for higher compute performance and data analytics required advanced semiconductor manufacturing technology. Gartner expects global AI semiconductor revenue to grow from US$12.3 bn in 2019 to US$43.9 bn in 2024E, at a CAGR of 29%. Although GPU and FPGA are the mainstream chipsets for AI computing, the demand for customised ASIC is rising as it can provide more efficient computing for training and inference, with AI chipset projects mostly on 16/12nm and 7nm, and driving growth for Alchip/GUC in the coming years. Initiate coverage on Alchip and Global Unichip We initiate coverage on Alchip with an OUTPERFORM rating and a TP of NT$630.00, implying 31% upside, based on 30x our 2021E EPS, factoring in our expectation of 30% sales CAGR from 2019-22E. We believe the upper half of the long-term valuation should be supported by the opportunity in China PC CPU replacement demand, in addition to its AI exposure. We also initiate coverage on Global Unichip with a NEUTRAL rating and a TP of NT$250.00, reflecting 28x our 2021E EPS. We believe the company’s share is fairly valued at the upper half of its range, factoring in its opportunity in 5G and AI. Key risks: (1) Slower local CPU/AI penetration in China or the US restrictions, (2) Consolidation or failure of the emerging AI start-ups and system company projects, (3) the US further expands its ban on China companies’ access to IP/EDA tools, (4) Customers growth and shift to a foundry direct business model, and (5) and intensifying competition from Verisilicon in China or build-up of more in-house teams.