Oracle adds ChatGPT, Codex and Google TPU chips to its cloud

The announcement builds on a relationship that has already produced tangible infrastructure. Thiagarajan said his team observed OpenAI's GPT-6 Astra model get trained at Oracle's Stargate data center campus in Abilene, Texas, and could see the real-time benefits. That training run gives Oracle a concrete reference point as it pitches its cloud to other AI developers.

The chip lineup expansion is notable because it is the first time Google's custom chips will run on another hyperscaler's cloud. Thiagarajan framed the move as a response to chip scarcity, saying customers such as OpenAI can benefit from having different accelerators to choose from, while adding that Nvidia chips remain 'the best, most fungible.' The Cerebras deal is described as material, though Cerebras won't disclose the scale because it is in a quiet period.

Separately, Oracle is managing pushback on its multibillion New Mexico data center, Project Jupiter. Thiagarajan said Jupiter is just one of 30-40 data center projects in development per quarter, and Oracle maintains the project is on schedule despite a force majeure notice delivered to backer Blue Owl Capital last month. Investors are now counting down to Oracle's investor day, where they will want to see if the company can back its fiscal year 2030 revenue targets and whether the margin profile of its cloud deals has changed due to the rising cost of capital. Oracle shares are down more than 30% this year.