OpenAI in talks for massive Google TPU deal as AI labs race to lock up every chip available
Key Points
- OpenAI is in talks to commit 1 gigawatt of Google TPU capacity, expanding its compute footprint beyond a 1.8 gigawatt NVIDIA deal in Ohio.
- Anthropic signed a 1.7 gigawatt commitment this quarter, including a 360 megawatt facility in Queensland—the largest overseas AI infrastructure lease to date.
- AI labs are now diversifying compute across NVIDIA, Google, AMD, and custom providers simultaneously, reducing single-vendor risk while enabling optimal chip configurations per workload.
Summary
OpenAI locks in massive Google TPU deal as AI labs race to secure every chip
OpenAI is in talks to commit 1 gigawatt of Google TPU capacity, according to SemiAnalysis analyst Dylan Patel. The deal would sit alongside OpenAI's existing 1.8 gigawatt multi-phase commitment in Ohio with NVIDIA, which Patel calls "the largest single leasing commitment we have ever seen."
The TPU arrangement is part of a broader shift: AI labs are now diversifying compute across every major chip maker and cloud provider available. OpenAI is also in discussions with NeoClouds for both GPU and TPU as a service.
Anthropic is moving faster into new geographies. The company signed a 1.7 gigawatt commitment this quarter, including a 360 megawatt facility in Queensland, Australia—the largest overseas lease by an AI lab to date. Operations start in Q1 2028. Google is backing 1 gigawatt of that commitment, though credit support for the remaining capacity is still undecided.
The strategic shift is structural. OpenAI was founded partly on the premise that Google was a competitive threat. Now the two companies are in direct infrastructure partnerships. The same pattern holds across the sector: AI labs are using "every chip and every data center and every hyperscaler"—signing with NVIDIA, Google, AMD, and custom cloud providers simultaneously.
The practical reasoning is straightforward: different chips and configurations suit different models and inference workloads. Even if one processor is faster at a specific task, there's often an optimal way to run a model on slower hardware. Anthropic's rapid porting of its implementation to AMD infrastructure during the Lisa Su interview is one recent example. The result is a rising-tide-lifts-all-boats dynamic for semiconductor vendors and neo-clouds, rather than a winner-take-all outcome.
There is still heavy NVIDIA dependence overall, but the diversification reduces the risk that a single architectural misstep becomes catastrophic.
Every deal, every interview. 5 minutes.
TBPN Digest delivers summaries of the latest fundraises, interviews and tech news from TBPN, every weekday.