Chips, cloud, and compute

AI computing rests on several layers, and organizations in the catalog sit at different ones. Chip designers create accelerators and processors: NVIDIA's annual report describes data center systems that combine its GPUs, CPUs, interconnects, and software stacks. Memory makers supply high-bandwidth memory, which Micron describes as 3D-stacked SDRAM designed for AI and high-performance computing workloads. Network vendors such as Arista sell Ethernet switches for connecting accelerators in AI clusters. Cloud providers such as CoreWeave rent out this hardware from the data centers they operate.1342

Designing a chip is not the same as making it. NVIDIA describes a fabless strategy in which suppliers handle wafer fabrication, assembly, testing, and packaging, and it names foundries such as TSMC and Samsung. Etched, which designs inference hardware, says its A0 silicon came back from TSMC's N4P process.15

A headquarters is not a data-center location. NVIDIA's and CoreWeave's annual reports list principal executive offices in Santa Clara, California, and Livingston, New Jersey. CoreWeave's report separately describes data centers across the United States, Europe, and Canada, located in leased buildings. In this catalog, a headquarters or principal executive office is used to decide eligibility; it says nothing about where an organization's chips are made or where its computing runs.127

This catalog does not measure capacity. It does not publish power-capacity figures, and announced megawatts, GPU counts, and speed claims remain the statements of the companies that make them. For performance, MLPerf Inference from MLCommons measures how fast specific submitted systems process inputs under published rules, with a closed division meant for like-for-like comparison. Those results describe the systems tested, not a company's total capacity.76