According to the latest AI server industry research from TrendForce, global cloud service providers (CSPs) are intensifying investments in AI servers and related infrastructure to accelerate the adoption and upgrading of AI applications. The combined capital expenditure of the eight leading CSPs is projected to surpass $710 billion by 2026, representing an annual growth rate of approximately 61%. In addition to continuing procurement of GPU solutions from NVIDIA and AMD, these providers are expanding the adoption of ASIC infrastructure to ensure the suitability of AI application services and improve the cost efficiency of data center construction. The eight major CSPs include U.S.-based Google, AWS, Meta, Microsoft, and Oracle, as well as China-based Tencent, Alibaba, and Baidu.
TrendForce estimates that Alphabet, Google's parent company, could see its capital expenditure exceed $178.3 billion in 2026, surging 95% year-over-year. Google has invested earlier than other CSPs in developing its own ASIC solutions, accumulating significant R&D advantages. The company is expected to transition its primary TPU to the new v8 platform this year. Driven by AI applications such as Google Cloud Platform and Gemini, TPU shipments are projected to account for nearly 78% of Google’s AI server output by 2026, widening the gap with GPU-based AI servers. Google is the only CSP where ASIC-based server shipments exceed those of GPU-based models.
Amazon recently increased its procurement scale for NVIDIA’s GB300 and V200 full-rack systems, reflecting its accelerated adoption of higher-power, higher-density GPU platforms to meet expanding cloud AI training and inference service demands. It is estimated that GPU-based models will constitute nearly 60% of Amazon’s AI server shipments by 2026. For its in-house ASIC development, the next-generation Trainium 3 is expected to begin volume production in the second quarter of 2026, replacing Trainium 2/2.5. However, due to factors such as software maturity and product validation, significant shipment momentum may not emerge until the second half of the year.
TrendForce forecasts that Meta’s capital expenditure will exceed $124.5 billion in 2026, growing 77% annually. Meta’s AI servers primarily rely on NVIDIA and AMD solutions, with GPU-based models expected to maintain over 80% of its AI server mix. The company is also advancing its own ASIC development to reduce per-unit computing costs and diversify reliance on single suppliers. However, supply chain assessments indicate that its MTIA initiative may face delays due to prolonged hardware-software integration efforts, potentially falling short of initial shipment expectations.
Microsoft remains optimistic about long-term demand for large-scale model training and inference, primarily deploying NVIDIA’s full-rack solutions to support its AI server shipments. The company recently introduced its in-house Maia 200 chip, targeting high-efficiency AI inference applications. Oracle continues to invest in GPU rack-scale solutions to support expanding AI data center projects, including collaborations with Stargate and OpenAI.
Analyzing Chinese CSP trends, although ByteDance has not publicly disclosed its 2026 capital expenditure details, TrendForce estimates that over half of its spending will be allocated to AI chip procurement. NVIDIA’s H200 is expected to become a key solution for ByteDance’s AI servers, though this depends on subsequent U.S.-China regulatory reviews. ByteDance is also expanding the use of domestic AI chips, primarily from suppliers such as Cambricon.
Tencent continues to rely on externally sourced GPU solutions from NVIDIA and others to support cloud and generative AI demands, while also partnering with local suppliers to develop proprietary ASIC solutions. These efforts focus on networking applications, data center infrastructure, and online AI services, aiming to diversify computing sources and enhance system integration flexibility.
Both Alibaba and Baidu are actively developing in-house ASIC AI chips. Alibaba, through subsidiaries such as T-Head and Alibaba Cloud, provides AI application infrastructure for public cloud and other online services. It also develops the Qwen large language model and related applications, serving its own cloud, enterprise, and consumer users. Baidu plans to gradually introduce its new Kunlun solutions after 2026, targeting large-scale AI training and inference applications. The company is also developing its Tianchi series of AI server clusters, capable of interconnecting hundreds of AI chips to strengthen overall AI system computing power.
Comments