Nvidia had their moat. A GPU castle surrounded by the trenches of the CUDA software ecosystem. "Nobody can come close!" everyone said. Everyone except Google.
It's a common enough adage in investing communities: “During the gold rush, sell shovels.” And that's what Nvidia did best. They sold the shovels for the likes of OpenAI, Anthropic, Meta, Google, etc. to dig for gold. And with so much software written specifically for CUDA, developers had few options other than Team Green's shovels.
Google's Gemini 3 being trained and run on TPUs marks an important turning point. They cut out Nvidia and beat them at their own game. Not only did they develop and train the AI using all in-house software, but their newest suite of LLMs is comparable to other frontier labs. Simultaneously, they're building the surrounding software ecosystem- XLA, JAX, PAX...these are the roads necessary for an “If you build it, they will come” approach to work. Their aim is to create a world where you either pay them for LLM inference, or pay them for TPU usage.
They have shown there is blood in the water at nearly every layer of the stack. From chipset to driver to compiler to model, Google wants it all. (Well… nobody can come close to decoupling from TSMC, but that’s another topic.)
Google is aiming for the crown, with a laser dot and trained hitmen behind the scope.