OpenAI has begun using Google’s artificial intelligence chips in a bid to lower operational costs and reduce its dependence on longtime partners Microsoft and Nvidia, according to a Reuters report. The company has recently started renting Google’s tensor processing units (TPUs) through Google Cloud to power applications like ChatGPT.
As first reported by The Information, OpenAI hopes this move will “help bring down the cost of inference,” referring to the process of generating responses from trained AI models—a major cost driver in scaling services like ChatGPT.
This represents OpenAI’s first significant use of non-Nvidia chips, marking a shift from its traditional reliance on Nvidia GPUs for both model training and inference. The collaboration also reflects a surprising turn, as Google, developer of competing AI models such as Gemini, has typically reserved its TPUs for internal projects. Recently, however, it has begun offering TPU access to select external clients like Apple and AI startups including Anthropic and Safe Superintelligence, both founded by former OpenAI leaders.
Still, Google is reportedly keeping its most advanced TPUs off-limits. A Google Cloud employee told The Information, “The company is limiting OpenAI’s use of its highest-end chips,” likely to maintain its competitive advantage.
This strategic shift comes amid growing friction between OpenAI and Microsoft, its largest investor and cloud partner. According to The Wall Street Journal, OpenAI executives have discussed accusing Microsoft of anticompetitive behaviour and seeking regulatory review of their agreement. While both companies publicly expressed confidence in their partnership, unresolved issues reportedly persist, including revenue sharing—OpenAI aims to reduce Microsoft’s cut from 20% to 10% by 2030 and control over future profits.
Tensions also involve OpenAI’s $3 billion acquisition of Windsurf, an AI coding startup. Microsoft is reportedly pushing for access to Windsurf’s IP to boost GitHub Copilot, but OpenAI has resisted.
OpenAI is now seeking to renegotiate exclusivity terms that restrict hosting its models solely on Microsoft’s Azure cloud, signalling a move toward a more diversified and flexible infrastructure strategy.
