Nvidia says its GPUs are ahead of Google’s AI chips

Nvidia said its GPUs remain a generation ahead of competitors, following a 3% drop in its shares after reports that Meta may use Google’s tensor processing units (TPUs) for data centers.

Nvidia holds over 90% of the AI chip market with its graphics processors, according to analysts, but Google’s TPUs have gained attention as an alternative.

Nvidia’s latest Blackwell chips are designed to run all AI models, while Google’s TPUs are used internally and available for rent via Google Cloud.

Google recently released Gemini 3, an AI model trained on TPUs, and said demand for both its TPUs and Nvidia GPUs is up.

Nvidia CEO Jensen Huang addressed increased TPU competition in a recent earnings call, noting that Google remains a customer and that Gemini can also run on Nvidia chips.

🔗 Source: CNBC

🧠 Food for thought

Implications, context, and why it matters.

Meta AI spend and chip mix decide if TPU use tests Nvidia

Meta plans $40–50 billion in AI accelerator spend for 2026 1, with talks on on‑premises tensor processing unit (TPU) installs from 2027 and Cloud TPU rentals as soon as next year 2. A 10–20% shift would put $4–10 billion of Nvidia revenue at risk each year.
Google Cloud (Google’s cloud computing division) says TPU growth could take up to 10% of Nvidia’s annual sales if many hyperscalers (the largest cloud providers that run massive global data centers) adopt it 2.
Meta will weigh compute density, power use, and total cost of ownership in its data centers 1. If TPUs hit those marks, the shift turns strategic and lifts Google’s hardware push.

Tool makers can ride portability as firms hedge between Nvidia and Google silicon

OpenXLA (an open-source compiler and runtime project) runs PyTorch and TensorFlow across TPUs, GPUs, and CPUs 3. Alibaba reports 72% speedups for GPT-2 plus 88% for Swin Transformer on Nvidia GPUs using OpenXLA 4.
Startups or integrators can build serving layers plus cost tools that price‑shop between TPU and GPU backends. PyTorch/XLA 2.6 (the PyTorch bridge to XLA (Accelerated Linear Algebra)) runs PyTorch models on Cloud TPUs with minimal code changes 5.
Enterprise AI teams will need unified observability, scheduling, and resource management across mixed fleets. Vendors shipping Kubernetes operators (extensions that automate operations on Kubernetes, the open-source container orchestration system) or MLOps platforms or inference engines that hide hardware quirks can win budget.