A rack-scale system is a computer designed and sold as a complete data-centre rack rather than as individual servers. Dozens of processors, typically graphics processing units or other accelerators, are joined by a high-bandwidth internal network so that they behave like a single large machine, with shared power delivery and, increasingly, liquid cooling built in. The design suits the training and running of large artificial intelligence models, which need many chips to exchange data very quickly. A prominent example is Nvidia's GB200 NVL72, announced in 2024, which links 72 GPUs in one rack.
For markets, the shift to rack-scale designs concentrates spending on fewer, far more expensive units and raises power density per rack. That affects chip makers, server assemblers, networking and cooling suppliers, and the build-out of the AI data centre by each hyperscaler. Analysts track it within data centre systems spending.