OpenAI tests Google's TPUs, but sticks with Nvidia

OpenAI isn't ditching its green machine just yet.

Despite whispers that the ChatGPT-maker was cozying up to Google's hardware to survive the ongoing compute crunch, OpenAI has stepped up to clear the air. A spokesperson confirmed to Reuters that while the team is conducting early testing with some of Google's Tensor Processing Units, there are currently no plans to deploy them at scale.

Kicking tires, but staying loyal

The rumor kicked off after The Information reported that OpenAI was exploring Google TPUs to meet the endless thirst for AI compute. Testing rival silicon is standard practice in Silicon Valley, but ripping out infrastructure to run production models at scale is a totally different beast with massive software hurdles.

For now, OpenAI remains deeply tethered to Nvidia GPUs, with AMD picking up secondary scraps. Even though OpenAI recently surprised the industry by tapping Google Cloud capacity, the vast majority of its heavy lifting still runs on Nvidia servers hosted by specialized cloud provider CoreWeave.

Meanwhile, Google has been aggressively pitching its TPUs to the rest of the world. External heavyweights like Apple are already on board, along with OpenAI's fiercest rivals, Anthropic and Ilya Sutskever's Safe Superintelligence.

OpenAI's real escape plan isn't Google at all. The company is developing its own in-house AI processor with Broadcom and TSMC, aiming to hit the critical tape-out milestone later this year, with hardware potentially ready by late 2025 according to Taiwan's Commercial Times.

Why it matters

We think this clarification highlights the sheer impossibility of escaping Nvidia's gravity right now.

Every AI lab wants to cut costs and avoid single-vendor lock-in, but switching away from Nvidia's battle-tested software ecosystem is fraught with friction. Testing Google's TPUs is a smart way for OpenAI to gather intelligence on rival tech and gain modest leverage in supply negotiations, but running flagship models on an arch-rival's proprietary silicon was always going to be a tough pill to swallow.

For end users, sticking with Nvidia ensures that model reliability and rollout schedules won't be derailed by messy architectural shifts.

Long-term, OpenAI's true hardware gamble looks focused on its custom silicon slated for late 2025. Until those in-house chips arrive from TSMC, OpenAI has little choice but to pay the Nvidia tax and rely on cloud partners like CoreWeave to keep the lights on.