The hardware behind AI.

Despite Thursday's selloff, Nvidia reported strong earnings this week, driven by consistent demand for its AI chips. Their GPUs are general-purpose processors that handle a wide range of workloads, which is why most companies train and deploy models on them. The demand has pushed orders several months into the future, and the company continues to sell every batch it produces.

Google takes a different approach. Its TPUs are built specifically for the matrix math that powers large models, and the hardware is connected across large clusters to support complete training runs. The design prioritizes speed and efficiency when models need to process billions of examples, and the system performs best within Google's cloud, where the infrastructure is optimized for it.

Both methods have their advantages. GPUs offer flexibility and a broad ecosystem, while TPUs focus on scale and predictable performance. As new models are developed, the type of hardware they use will influence cost, capability, and how reliably they can train at the sizes current frontier models demand.