Shipments of high-end cloud AI accelerators, 2027

DIGITIMES observed that cloud service providers (CSPs) such as Google and AWS will sharply ramp up self-developed ASICs in 2027, lifting high-end cloud AI accelerator shipments to 25.951 million units, up 64.5% year-over-year. Of this total, ASIC shipments will reach 15.262 million units, up 110.9% year-over-year, and overtake GPU shipments for the first time. Nvidia will remain the largest vendor by market share, with shipments estimated at 9.096 million units, up 18.2% year-over-year, and market share declining to 35.1%. Google shipments will rise to 8.009 million units, up 134% year-over-year, with market share reaching 30.9% and narrowing the gap with Nvidia to 4.2pp. Supply bottlenecks are shifting to key components such as T-Glass, ABF substrates, high-layer-count printed circuit boards (PCBs), and HBM3E, HBM4, and HBM4E.

On the demand side, Microsoft, Amazon, Google, Meta, and Oracle continue to expand AI data centers, power infrastructure, and network infrastructure. Agentic AI extends a single interaction into multi-step planning, tool calls, and repeated reasoning, increasing token consumption and inference demand. OpenAI, Anthropic, neoclouds, and sovereign AI initiatives are also expanding GPU procurement and cloud compute leasing for tensor processing unit (TPU) and Trainium. On the supply side, Nvidia Rubin, Google TPU v8, and Amazon Trainium3 are entering a ramp phase. Next-generation chip package area and layer counts are increasing, and low-CTE glass fabric, high-end ABF substrates, high-layer-count PCBs, and HBM3E, HBM4, and HBM4E may still constrain shipments of high-spec products. Nvidia is also evaluating a tiered high-bandwidth memory (HBM) configuration for Rubin Ultra to reduce memory content per chip.

This report covers the shipment of high-end cloud AI accelerators, the supply-demand factors that affect it, and how shipments differ among suppliers, process technologies, and GPU architectures.