XPU
A catch-all name for a custom AI accelerator chip built by a cloud provider or model lab instead of bought off the shelf.
XPU is deliberately vague. The X stands in for whatever letter the chip would otherwise get: GPU, TPU, NPU. In practice it means a processor a company designed itself for its own AI workloads, rather than a general-purpose chip bought from Nvidia or AMD.
The reason so many exist now is cost and leverage. If you run enormous amounts of inference, a chip tuned to exactly your workload can be cheaper per answer, and designing your own is the only real bargaining position against a single dominant supplier. Google’s TPUs are the oldest example, and Amazon, Microsoft, OpenAI and Anthropic have all followed. The hard part is rarely the chip itself; it is everything around it, the memory, the packaging and the networking needed to turn silicon into a working rack.