OpenAI has begun renting Google Cloud’s Tensor Processing Units to power its flagship ChatGPT product and other AI services—marking its first major use of non‑Nvidia chips and a strategic shift away from its historic reliance on Microsoft Azure and Nvidia GPUs. OpenAI is now utilizing Google’s TPUs for model serving, a move aimed primarily at reducing the high costs associated with inference computation—the phase where trained AI models respond to user input. While Google is not offering its top-tier TPUs, this marks a milestone in OpenAI’s diversification strategy.
This arrangement reflects a growing recognition in the AI industry that supply-chain resilience and cost efficiency matter. By expanding beyond its traditional Nvidia‑Azure setup, OpenAI aims to introduce competition into its computing stack. For Google, landing OpenAI as a customer—alongside other AI players like Apple and Anthropic, validates its move to commercialize internal TPU hardware.
Nvidia currently dominates AI chip supply, with an estimated 90% market share in high-performance AI hardware. Google’s push to externalize TPUs signals an intent to challenge that dominance and offer a compelling alternative for compute-hungry AI services. OpenAI expects Google TPUs to help lower the costs of inference workloads, which have ballooned with ChatGPT’s user growth. Reports estimate OpenAI’s chip spending could rise from $4 billion in the prior year to nearly $14 billion by 2025. Renting TPUs also gives OpenAI a safety net as it pursues its own custom chip development efforts.
Despite the competitive backdrop—OpenAI’s ChatGPT rivals Google’s Gemini—the deal underscores a pragmatic turn in AI infrastructure provisioning. Even fierce rivals are forging infrastructure deals to meet skyrocketing compute demands. In essence, OpenAI’s move to Google TPUs puts pressure on Nvidia, boosts Google Cloud’s AI credentials, and highlights a strategic rethink among AI leaders: diversifying and optimizing chip supply chains is becoming essential in scaling cutting-edge services.
