OpenAI plans to buy Google's in-house TPU chips for its expanding compute build-out, purchasing them alongside Nvidia GPUs, The Information reported on October 28. Alphabet shares rose on the news.
The weight of the deal is not its size — neither scale nor delivery timeline has been disclosed — but its structure. OpenAI has long been one of Nvidia's most important customers, with training and inference heavily dependent on its GPUs; The Information framed the plan as a direct hit to Nvidia's bond with its biggest customer. Other US outlets were blunter: Google's AI chip business is getting harder for Nvidia to ignore.
TPUs are Google's custom tensor processors, previously serving Google's own services and cloud customers. An OpenAI deal would bring the platform its first external customer at this scale — a new phase for Google's silicon exports, and another crack in the default equation of AI compute with Nvidia GPUs.
The logic of diversification is straightforward. Top labs' compute demand is growing faster than any single supplier's capacity elasticity, worsening queue times and bargaining positions; moving part of the load to custom ASICs or third-party silicon is both leverage and a hedge. OpenAI already works with Broadcom and AMD; adding Google further dilutes Nvidia's exclusive color in its chip mix.
For Nvidia, customer diversification is the real test of its moat. For Google, whether TPU can carry OpenAI-scale training loads will soon have an answer. Structural shifts in compute supply tend to hide in a big customer's second order form.