TrendForce: AI Server shipments will increase by more than 28% annually in 2026

According to TrendForce’s latest AI industry research, in order to expand the deployment of AI training and inference applications, the five largest cloud service providers (CSPs) in North America have significantly increased their willingness to purchase rack-scale AI Servers in 2026. Not only are they expected to account for more than 60% of global NVIDIA GB/VR demand, they will also simultaneously drive the total AI training computing power of the five major manufacturers to increase by more than 56% annually, and the total AI inference computing power to grow by around 122%.

TrendForce predicts that AI Server shipments will increase by more than 28% annually in 2026, and high-end AI training models will still be the mainstay, accounting for approximately 55%. However, in the medium to long term, the market will be dominated by AI inference machines. The main reason is that CSPs will actively promote AI applications to accelerate the commercialization of AI cloud services. In addition, NVIDIA will also expand more AI inference solutions or usage scenarios, including promoting the GB/VR system, this year’s main AI Server solution, in addition to AI training purposes, with special emphasis on the solution’s ability to support AI inference-related workloads.

According to TrendForce estimates, the combined capital expenditures of Google, Amazon, Microsoft, Meta and Oracle will exceed US$770 billion in 2026, an annual increase of nearly 87%. Analyzing the computing power obtained by purchasing NVIDIA GB/VR series from the five major CSPs in North America, for the AI ​​training part, if FP16/BF16 is used as the basis for estimation, the total computing power of the five major manufacturers has exceeded 9 ExaFLOPS in 2025, and is expected to grow by more than 56% in 2026.

For AI inference, if FP4/NVFP4 computing performance is used as the basis for estimation, the total computing power of the five largest CSPs in North America will exceed 37 ExaFLOPS in 2025, and is expected to grow significantly by nearly 122% in 2026, which is significantly higher than AI training. This reflects that NVIDIA places special emphasis on AI inference performance in this software and hardware system adjustment, and implements it in the new generation of GB300 and VR200 full cabinet solutions.

In addition to GPU solutions, CSP manufacturers are simultaneously promoting self-developed ASIC complete cabinet products, among which Google is the most active. TrendForce predicts that Google's demand for its own TPU chips will increase by nearly 80% year-on-year in 2026, and will gradually upgrade from the v7 generation to the v8 generation starting in the second half of the year. In addition, Amazon’s self-developed ASIC is second only to Google, and it is expected that its Trainium series will account for more than 40% of its own AI Server in 2026.

TrendForce said that the new generation cabinets of NVIDIA, AMD and CSP's self-developed ASICs have integrated liquid cooling systems, which helps reduce the U count (Server rack unit) of AI Server and increase the number of accelerators that can be accommodated in a single cabinet. When the thermal design power consumption (TDP) of a single AI GPU or ASIC increases simultaneously, the power consumption of the AI ​​Server system is structurally amplified.

According to estimates by TrendForce, the total annual increase in server power consumption of the five largest CSPs in North America will jump from 2.8GW in 2023 to 18GW in 2026, with an annual growth rate of 116% from 2025 to 2026. The main reason is that the AI ​​competition is intensifying, and NVIDIA GB300, AMD Helios and CSP self-developed ASIC platforms will increase in volume simultaneously. (Source: TrendForce)

Please indicate the source when reprinting! For more LED information, please pay attention to the official website (www.ledinside.cn) or search the WeChat public account (LEDinside).