Just four months after ChatGPT’s launch, Satya Nadella explained his ambitions for Microsoft in the rapidly accelerating AI race: “I want people to know that we made them dance.” Google was clearly caught off guard by the rise of LLMs. The company’s rushed launch of Bard was criticized as emblematic of a broader “bureaucratic lag” at Google. But now Google has most certainly come out to dance. The recent launch of Gemini 3 Pro reportedly “tops the leaderboard on reasoning, coding, multimodal reasoning, and tool-use benchmarks,” outperforming most rival models. Now, it’s Google’s turn to make Nvidia dance.
News broke this week that Meta is considering using Google TPUs in its own data centers beginning in 2027, and may rent TPUs from Google starting in 2026. Google has not previously sold or leased TPUs to other companies. In response to this news, Alphabet stock rallied over 6% on Monday, and Nvidia shares fell as much as 7% before partially retracing. Nvidia responded in an X post saying, “We’re delighted by Google’s success — they’ve made great advances in AI and we continue to supply to Google… NVIDIA is a generation ahead of the industry — it’s the only platform that runs every AI model and does it everywhere computing is done.”
Google’s TPU offering positions it as the first serious merchant-silicon competitor to Nvidia in frontier-scale AI compute. One report from SemiAnalysis explains how, in particular, Google’s system-level design (scale-up networking, liquid cooling, optical circuit switching, compiler efficiency, and internal software tooling) allows TPUv7 to achieve lower cost per effective FLOP than Nvidia Blackwell.
Beyond the broader system-level improvements, Google has made important positions around its software ecosystem. Google has begun shifting its TPU software stack from an internally focused system built around JAX and XLA to a more open, externally usable model that adds native PyTorch support and integrations. Google is attempting to reduce the friction of adopting TPUs for large-scale training and inference. In contrast, Nvidia’s CUDA ecosystem remains the most mature and widely adopted accelerator software stack, offering first-class PyTorch integration, highly optimized libraries, and broad community support built over more than a decade. While CUDA still provides a more complete and battle-tested environment, Google’s new software investments are designed to narrow that usability gap and make TPUs a feasible alternative for labs with the engineering capability to extract high utilization.
However, SemiAnalysis argues that, while TPUv7 represents Google’s first true Nvidia-competitive platform, with superior TCO, massive scale, and enough ecosystem support to win deals from Anthropic, Meta, and others, TPUv8 is only incrementally better, while Nvidia Rubin is aggressively better, meaning TPU’s window of advantage may narrow beyond 2027. The key questions going forward will be around the TPU ecosystem’s success software openness, external adoption, and Google’s ability to accelerate silicon cycles as fast as Nvidia.
