TPU

Google has launched its seventh-generation Tensor Processing Unit (TPU), Ironwood, now generally available to cloud customers. This custom-designed AI accelerator aims to rival Nvidia GPUs by offering specialized, high-performance computing power for AI workloads like training LLMs and real-time inference. Ironwood boasts interconnectivity of up to 9,216 TPUs in a single pod. AI startup Anthropic has pledged to utilize up to 1 million Ironwood TPUs. The move underscores Google’s commitment to AI infrastructure amid growing cloud revenue and increased capital expenditure.

OpenAI is diversifying its AI infrastructure by utilizing Google’s TPUs alongside NVIDIA GPUs, as part of its drive for computational power, including for ChatGPT. This move, motivated by cost management and infrastructure diversification, could boost the profile of Google’s TPUs as an alternative to NVIDIA’s GPUs. The partnership, however, shows competitive dynamics, with Google limiting access to its most powerful TPUs.