AI Servers to Take a Third of Data Center Power as Nvidia Moves to Rack Scale

Power, not silicon, is becoming the headline constraint on AI hardware deployment. TrendForce estimates global data center power demand capacity will reach 161 GW in 2026, up about 31 percent from 122.9 GW in 2025, with AI servers accounting for roughly 33.4 percent of total demand capacity against just under 25 percent a year earlier. That share could exceed 40 percent in 2027, while general servers fall to about 25.6 percent of total demand capacity. Operators are already planning high-voltage direct current distribution architectures to improve end-to-end efficiency as IT and non-IT load rise together.

The supply side of that equation is deteriorating. TrendForce estimates grid deliverability was sufficient to accommodate data center growth through 2025, but that supply and demand begin diverging in 2026 and the gap widens sharply from 2028 onward. By 2030, data center power demand capacity is projected at 490.7 GW against roughly 222.6 GW of grid capacity available to data centers, a nominal gap of 268 GW — though behind-the-meter on-site generation is not counted in that figure, so the effective shortfall is smaller than the headline. In the United States, which hosts the largest concentration of deployments, the gap is projected to exceed 170 GW by 2030.

Hardware roadmaps are being rewritten around rack-scale delivery. DIGITIMES reported that Nvidia will enter a heavy shipment phase for Vera Rubin in the second half of 2026, followed by Vera CPU, Rubin GPU and NVL4, NVL8 and NVL72 platform configurations. That shift moves the procurement unit from the accelerator to the rack, which pulls thermal management, power delivery, high-speed optics and busbar capacity into the critical path alongside GPUs and high-bandwidth memory. For buyers, the practical consequence is longer lead times and a wider set of suppliers that must be qualified in parallel rather than a single vendor scorecard.