AI Server Buildout Pushes Data Center Power Demand to 161 GW and Fabless Revenue Up 73%

TrendForce now estimates that global data center power demand capacity will reach 161 gigawatts in 2026, roughly 31% more than a year earlier. AI servers are the primary driver of that growth and are expected to account for about 33.4% of total demand capacity. The firm also warns that the gap between what AI infrastructure wants to draw and what grids can supply will widen in 2028, which shifts the binding constraint on the buildout from chip availability toward grid and generation capacity.

The chip layer is seeing the same pull. Combined revenue of the world's top ten fabless IC design companies surged 73% year on year to more than US$141.45 billion in 2Q26, according to TrendForce, as expanding AI applications drove demand for GPUs, CPUs, ASICs and interconnect products. AMD broke into the top three of the ranking, reflecting how much of the accelerator and server CPU mix is now contested rather than single-sourced.

Rack-level economics are the transmission channel between the two. As power density per rack climbs, spending shifts away from the accelerator itself and toward power delivery, thermal management and high-speed interconnect, where the supply chain has reported some of the sharpest revenue gains. In practice that spreads the benefit across power supply, cabling, thermal and optical vendors rather than concentrating it in the accelerator alone, so a slowdown in GPU units would not immediately translate into a slowdown in total rack spend.

For buyers the planning implication is that 2027 capacity is being decided now by power rather than silicon. A 31% increase to 161 GW implies incremental demand of roughly 38 gigawatts in a single year, and grid interconnect lead times in most major markets run longer than the procurement cycle for servers. Anyone building AI capacity should treat secured power and cooling as the gating item and lock component slots against it.