Google Cloud

Full briefing

Google LLC operates Google Cloud Platform (GCP), the world's third-largest cloud by revenue after AWS and Azure. Google occupies a unique position in the AI chip supply chain as both a hyperscaler buyer of AI compute and a chip designer in its own right — through its Tensor Processing Unit (TPU) program, now in its sixth generation. Google's TPU program began in 2013 as an internal project to accelerate neural network inference workloads, driven by the realization that CPUs and GPUs were inefficient for matrix multiplication — the dominant operation in deep learning. The first-generation TPU was deployed internally in 2015 and disclosed publicly in 2017. Subsequent generations (TPU v2 through TPU v5e/v5p) have evolved from inference-only to full training+inference capability at datacenter scale. The TPU supply chain is: (1) Google's hardware team defines the chip architecture and microarchitecture; (2) Broadcom serves as the ASIC design partner, providing physical design and tape-out expertise; (3) TSMC fabricates the chips on advanced nodes (TPU v4 on N7, TPU v5e on N4); (4) custom rack-scale systems are assembled by ODMs including Quanta Computer. This end-to-end custom stack allows Google to optimize system efficiency for transformer workloads in ways that off-the-shelf NVIDIA hardware cannot match. For workloads where TPUs are not optimal — particularly third-party AI model serving where customers bring their own NVIDIA-trained models — Google Cloud offers A2, A3, and A3 Mega instances based on NVIDIA A100, H100, and H200 GPUs respectively. Google is thus a significant NVIDIA GPU customer alongside being a TPU developer. Google's deep integration with Broadcom (for networking ASICs, switch silicon, and TPU co-design) makes the Google-Broadcom-TSMC triangle one of the most strategically important supply chain relationships in AI infrastructure.