Graphics processing unit (GPU)

GPUs contain thousands of small cores that handle matrix and vector operations simultaneously, which suits the math at the heart of neural networks. Data center GPUs pair this with high-bandwidth memory and fast interconnects that link many chips into a single cluster. Software ecosystems that make GPUs programmable have been as important to adoption as the hardware itself.

GPUs are the largest component of AI data center spending, and their supply, pricing and generational upgrades shape the economics of AI providers and cloud companies. Alternatives include custom accelerators designed by hyperscalers and specialized chips for inference. Key metrics include performance per watt, memory capacity, interconnect bandwidth and total cost of ownership. Rapid product cycles raise questions about useful life and depreciation schedules.

For allocators, GPU demand links semiconductor, cloud, power and real estate investment themes. Example: a neocloud borrows against its GPU fleet to buy the next generation of chips, with the loan secured by contracts from AI labs renting the capacity.