‘Could change AI’: Google Gemini is performing strong even without Nvidia GPUs
Google’s Gemini models increasingly run on Google’s own TPU chips instead of Nvidia GPUs, raising questions about whether the shift could reshape the AI hardware landscape. Here’s what’s actually true about Gemini’s performance and Google's reduced reliance on Nvidia.
1. Gemini Runs Primarily on Google’s TPU Chips
Google has publicly confirmed that Gemini models were trained and deployed largely on its in-house Tensor Processing Units (TPUs), particularly TPU v4, v5e, and newer Ironwood variants. These chips are optimised for large-scale AI workloads and give Google full control over its infrastructure. This means Gemini is not dependent on Nvidia GPUs for its core training and cloud inference operations.
2. Google Claims Its New TPUs Outperform Some Nvidia Systems
According to Google and third-party reporting by Tom’s Hardware, the company’s seventh-generation Ironwood TPUs and Axion CPUs outperform some Nvidia hardware setups in specific training and inference workloads. Google presented benchmark data showing competitive or superior performance for certain tasks, suggesting that its hardware can now rival established GPU-based systems.
3. Analysts Highlight Gemini’s Ability to Run Without Nvidia GPUs
Several industry analysts, including independent researchers, note that Gemini 2.5 Pro and later variants run end-to-end on Google’s TPU infrastructure rather than Nvidia’s GPU ecosystem. This reflects Google’s broader strategy of controlling both the hardware and software stack, reducing supplier dependency and allowing customized AI optimization.
4. Gemini’s Architecture Is Designed for Efficiency Across Devices
Google says Gemini models were engineered to run efficiently across data centres, on-device environments, and distributed computing systems. This scalability is partly enabled by TPUs, which allow tight integration between hardware and model architecture. While “efficient” does not always mean “higher quality,” it does mean faster, cheaper, and more energy-efficient performance in many use cases.
5. On-Prem Deployments Still Use Nvidia Blackwell GPUs
Despite Google’s strong reliance on TPUs in its cloud, the company has confirmed that Gemini 2.5 offered through Google Distributed Cloud for on-premises deployments will run on Nvidia Blackwell GPUs. This indicates that while Google can run Gemini without Nvidia, some enterprise customers still rely on Nvidia hardware for compatibility and local infrastructure needs.
6. Google’s TPU Strategy Reduces Reliance on External Suppliers
Using TPUs allows Google to control production, scale infrastructure more predictably, and reduce exposure to global GPU shortages. Moving away from Nvidia dependence strengthens Google’s AI independence and lowers operational costs, which could translate into faster model rollouts and more reliable cloud performance.
7. Early Testing Shows Strong Performance, Not Guaranteed Superiority
Tech reports and independent tests show Gemini performing competitively against top models like OpenAI’s GPT series, but the advantage depends heavily on the task. While TPU optimization delivers efficiency and speed, reviewers note that Gemini does not universally outperform all Nvidia-powered models in every benchmark. The hardware shift helps, but it is not a magic breakthrough.
8. The Real Impact: More Competition in AI Hardware
If Gemini continues to demonstrate high performance on TPUs, it could reduce the industry’s dependence on Nvidia GPUs for large-scale AI systems. This diversification matters because Nvidia dominates the global AI chip market. Google’s successful TPU-first strategy could push other companies toward custom AI hardware, increasing competition and lowering long-term costs.
)
)
)
)
)