OpenAI is in discussions to buy AI computing chips directly from Google, according to a person familiar with the matter, a step beyond the TPU rental arrangement it has already had running through Google Cloud since June 2025. Final terms of any purchase aren't settled yet, and the person requested anonymity because the talks are private, but the direction is clear: OpenAI wants to own tensor processing unit capacity, not just borrow it on Google's infrastructure.
The hires that telegraphed this move
The buying discussions aren't appearing out of nowhere. OpenAI has reportedly been hiring TPU compiler engineers, the specialized talent needed to get its own models running efficiently on Google's chip architecture rather than Nvidia's, which is a meaningfully different skill set from optimizing for Nvidia's GPUs and CUDA software stack. Building that expertise in-house before a chip purchase is the kind of groundwork that suggests this isn't a speculative trial balloon, it's infrastructure planning that's already underway.
Why 'too expensive' is doing a lot of work here
Coverage of the talks out of Asia has framed the move bluntly around cost, reporting that Nvidia chips are too expensive as the operative reason OpenAI is turning to Google's TPUs. That's a notable thing to be circulating about the company that is, by a wide margin, Nvidia's largest and most closely watched customer. Nvidia's own growth narrative to investors leans heavily on enormous, concentrated AI lab spending, and OpenAI diversifying its chip bill is a direct test of how much of that spending was ever exclusive to begin with.
This isn't OpenAI's only hedge against Nvidia
The Google talks sit alongside another, separate bet: OpenAI is developing its own custom AI processor in partnership with Broadcom, targeting deployment by the end of 2026. Taken together, a Google TPU purchase and a Broadcom-built custom chip describe a company trying to reduce its dependence on any single supplier, Nvidia included, rather than one opportunistic deal. For a company whose compute bill is one of its largest and fastest-growing costs, that diversification is as much a negotiating position against Nvidia as it is a genuine technical hedge.
What it means for the Google-versus-Nvidia chip rivalry
For Google, landing a direct chip sale to OpenAI, rather than just a cloud rental relationship, would be a meaningful validation of its TPU hardware business, which has mostly powered Google's own models and a smaller circle of cloud customers compared to Nvidia's sprawling customer base. Selling silicon to the company that makes ChatGPT, one of Nvidia's most referenced customers in its own earnings commentary, turns TPUs from an internal Google tool into a credible second option other AI labs can point to when they negotiate their own Nvidia contracts.
The bigger picture for anyone buying AI infrastructure
The practical lesson here extends past the three companies involved. If the industry's most prominent AI lab is actively diversifying away from a single chip supplier, after years of a near-default assumption that serious AI infrastructure meant Nvidia GPUs, that's a signal worth watching for any business planning its own AI compute spending, media and marketing infrastructure included. Supplier concentration that looked like a reasonable bet two years ago is increasingly something even the largest buyers are moving to unwind.
