花旗：NVIDIA_ GPU 与定制 ASIC – 两者将共存

Investment Rating - The report assigns a Buy rating to NVIDIA Corp (NVDA) with a target price of US$150.00, implying an expected share price return of 11.3% [6] Core Viewpoints - GPUs and ASICs will coexist in the AI/ML infrastructure buildout, with ASICs used for specialized models and GPUs for training and inference of larger, more complex models [4] - NVIDIA GPUs are expected to dominate the AI accelerator market, with 75% share of the $380B AI accelerator TAM by 2028 [4][22] - NVIDIA holds a 67% share of the AI accelerator installed base among the top four US hyperscalers, driven by its performance lead, scaling capabilities, and large installed base [5] GPU and ASIC Market Dynamics - NVIDIA GPU compute sales are forecasted to grow 118% YoY in 2024E and 84% YoY in 2025E, with hyperscaler GPU sales growing 100% in 2024E and 67% in 2025E [4] - Custom ASICs are expected to grow to 8% of US cloud provider capex in 2024E, with sales ramping by 200%+ in 2024E before normalizing to 16% YoY growth in 2025E [4] - By 2028, the AI accelerator market is projected to reach $500B, with GPUs accounting for 76% and ASICs for 25% of the market [26] Hyperscaler AI Accelerator Landscape - Google and Microsoft lead in AI accelerator installed base, with ~2M units each, while Amazon has ~1.5M and Meta has ~800K [26] - Microsoft relies heavily on NVIDIA GPUs (95%+ of installed base), while Google leans on its TPU infrastructure (65% of installed base) but is increasing investments in NVIDIA GPUs [26] - Amazon balances its own Inferentia and Trainium chips with NVIDIA offerings (60% of installed base), while Meta is aggressively investing in NVIDIA products (90% of installed base) [26] Financial Projections - NVIDIA's sales revenue is projected to grow from $26.97B in 2023 to $262.28B in 2027E, with a 73.4% gross margin in 2027E [10] - EBITDA is expected to increase from $13.34B in 2023 to $181.33B in 2027E, with an EBITDA margin of 69.1% in 2027E [10] - Core EPS is forecasted to grow from $0.33 in 2023 to $5.51 in 2027E, with a P/E ratio of 24.5x in 2027E [10] ASIC Design and Development - Full-custom ASIC design offers reduced size and higher capabilities but is more complex and costly, while semi-custom ASIC design reduces initial design efforts and limits non-recurring engineering costs [17] - Programmable ASICs provide flexibility by allowing hardware-level programming after manufacturing, making them suitable for specialized workloads [18]