研报 | 预估2026年全球八大CSP合计资本支出将破7,100亿美元，谷歌TPU引领ASIC布局

Core Insights - The global cloud service providers (CSPs) are significantly increasing their capital expenditures on AI servers and related infrastructure, with a projected total exceeding $71 billion in 2026, reflecting a year-on-year growth rate of approximately 61% [2][5][6]. Group 1: Major CSPs and Their Investments - The eight major CSPs include Google, AWS, Meta, Microsoft, Oracle, Tencent, Alibaba, and Baidu [6]. - Google is expected to have a capital expenditure exceeding $1.783 billion in 2026, with a year-on-year increase of 95%. Google has a significant lead in ASIC development, with its TPU shipments projected to account for nearly 78% of its AI server output [6][7]. - Amazon is increasing its procurement of NVIDIA GPU systems, with its GPU models expected to represent nearly 60% of its AI server offerings in 2026. The new generation of its self-developed ASIC, Trainium 3, is anticipated to launch in the second quarter of 2026 [7]. - Meta's capital expenditure is projected to exceed $1.245 billion in 2026, with a year-on-year growth of 77%. Its AI servers will primarily utilize NVIDIA and AMD solutions, with GPU models expected to account for over 80% [8]. Group 2: ASIC Development and Market Dynamics - Microsoft is focusing on long-term demand for large model training and inference, primarily acquiring NVIDIA solutions for its AI servers. The company has released its self-developed chip, Maia 200, aimed at high-efficiency AI inference applications [8]. - Oracle is expanding its GPU solutions in response to AI data center projects, while ByteDance is expected to allocate over half of its capital expenditure to AI chip procurement, with NVIDIA's H200 being a key solution [9]. - Tencent is sourcing NVIDIA GPUs to support cloud and generative AI demands while collaborating with local firms to develop its own ASIC solutions [9]. - Both Alibaba and Baidu are actively developing their own ASIC AI chips, with Alibaba providing AI infrastructure through its subsidiaries and Baidu planning to introduce its Kunlun solutions for large-scale AI training and inference applications [9].