Nvidia‑Google chip competition note
- Recent reporting highlights Google launching inference chips that challenge Nvidia's inference lead. - Market commentary links Nvidia share moves to Google's new AI chip efforts testing inference leadership. - The competitive push could open alternative inference hardware options beyond Nvidia in the medium term (ts2.tech)
Google has moved deeper into Nvidia’s strongest artificial-intelligence business with custom chips built to run models after training, a step aimed squarely at inference. (blog.google) Inference is the stage when a model answers a prompt, generates an image, or ranks search results, and Google said on April 9, 2025 that its seventh-generation Ironwood Tensor Processing Unit was its first chip designed specifically for that job. Google said Ironwood would be offered to cloud customers later in 2025 in 256-chip and 9,216-chip configurations. (blog.google) Google said a 9,216-chip Ironwood pod delivers 42.5 exaflops of compute, and the company pitched the chip as more power-efficient than its prior Trillium generation. Reuters reported the launch from San Francisco on April 9, 2025 as Alphabet’s latest attempt to speed AI applications with in-house silicon. (blog.google) (finance.yahoo.com) Nvidia still leads the broader market for AI accelerators, but inference has become the part of the business cloud providers most want to cheapen and scale. Nvidia markets its Blackwell systems on lower cost per token for inference, showing how central that workload has become to the company’s pitch. (nvidia.com) Google’s push is not limited to one product cycle. At Google Cloud Next on April 22, 2026, the company introduced separate eighth-generation Tensor Processing Units for training and inference, a split that MarketWatch described as two distinct chips and CNBC described as a direct shot at Nvidia. (marketwatch.com) (cnbc.com) Investors have treated that effort as more than a lab project. Reuters reported on November 25, 2025 that Meta Platforms was in talks to spend billions of dollars on Google chips for data centers starting in 2027, a report that would cast Google as a more serious rival to Nvidia. (finance.yahoo.com) That report hit Nvidia shares immediately. CNBC said Nvidia stock fell 4% on November 25, 2025 after The Information reported Meta was considering Google’s Tensor Processing Units for future data-center use. (cnbc.com) Google is still competing with a software and ecosystem gap. Reuters noted that more than 4 million developers rely on Nvidia’s CUDA platform, and Google’s own cloud documents say Ironwood is exposed through software layers including JAX, PyTorch and the XLA compiler rather than CUDA. (finance.yahoo.com) (cloud.google.com) The contest now is less about who can train the biggest model once and more about who can serve billions of responses every day at lower cost and power. Google has put custom inference chips at the center of that fight, while Nvidia is trying to hold it with Blackwell and its software stack. (nvidia.com) (blog.google)
