With the increasing need for AI training and inference models, the leading cloud service providers (CSPs) in North America are expected to significantly increase their procurement of rack-scale AI servers in 2026, according to new data from TrendForce.
These top 5 companies from North America will account for more than 60% of global demand for Nvidia’s GB and VR series of GPUs. TrendForce forecasts that infrastructure expansion will boost these CSPs combined AI training compute power by more than 56% year-over-year.
AI server growth
Not surprisingly, global AI server shipments are also expected to produce year-over-year growth to the tune of more than 28% with high-end training servers leading the market with 55% of these total shipments.
AI inference servers are projected to overtake training servers in the medium-to-long term, TrendForce said. This is due to CSPs looking to commercialize AI cloud servers. Nvidia is set to expand its AI inference solutions and use cases to support this trend, the market research firm added.
The combined capital expenditures on rack-scale AI servers will exceed $770 billion in 2026 from Google, Amazon, Microsoft, Meta and Oracle. This will be an increase of 87% in spending with purchases of Nvidia GB and VR series GPUs as the main thrust of this spending.
As a result, combined computing power is projected to exceed 9 ExaFLOPS in 2025 and grow by more than 56% in 2026.
What else are CSPs buying?
GPUs are not just a hot buy for CSPs but these companies are steadily advancing their own in-house application specific integrated circuits (ASICs) with Google leading the way in this sector.
TrendForce forecasts that Google’s TPU chips will grow in volume by nearly 80% year-over-year in 2026 with a transition from the version 7 to the version 8 generation in the second part of the year. Amazon will be second in its procurement of its own ASICs with its Trainium series accounting for more than 40% of Amazon’s own AI server shipments in 2026.
Liquid cooling rising, thermal issues remain
Server racks with Nvidia and AMD GPUs along with these in-house ASICs are rapidly adopting liquid cooling systems. This will not only reduce the server height in these racks, but will also allow a single AI server rack to accommodate more accelerators.
However, this will not solve the thermal challenges as these chips continue to demand more power and therefore more AI server power as the architecture becomes more powerful, TrendForce said.
The total power consumption of all servers operated by the top 5 North American CSPs grew by 2.8 GW in 2023 compared to the previous year. In 2026, total server power consumption is projected to surge by 18 GW compared to 2025, or year-over-year growth of a staggering 116%, TrendForce data said.
And it will likely continue to increase in the years ahead.
