Google Eyes Custom Chips
- Google is reportedly in talks with Marvell to build custom AI inference chips aimed at challenging Nvidia's lead. - Reports say the discussions focus on bespoke silicon for inference workloads across cloud and edge deployments. - Developing custom inference chips could shift cloud vendor competition and reduce dependence on off-the-shelf GPU providers. (x.com)
Alphabet’s Google is in talks with Marvell Technology to develop two new chips for running artificial intelligence models more efficiently, according to a Reuters report citing The Information. (reuters.com) The report said one chip is a memory processing unit designed to work alongside Google’s Tensor Processing Units, and the other is a new Tensor Processing Unit built specifically for inference, the step where a trained model generates answers, images or predictions. (reuters.com) Inference has become the expensive part of the artificial intelligence business because companies must keep serving models after training is finished. Google said in April 2025 that Ironwood, its seventh-generation Tensor Processing Unit, was its first chip designed specifically for inference. (blog.google) Google’s current Ironwood family is already on Google Cloud, where the company says TPU7x is the latest Tensor Processing Unit available and the first release in its seventh-generation Ironwood line. Google also said Ironwood delivers more than four times better performance per chip for training and inference than TPU v6e. (docs.cloud.google.com) (cloud.google.com) The Marvell talks point to a supply-chain shift as much as a product shift. CNBC reported Marvell shares rose on April 20 after the report, while Broadcom shares fell, even though Broadcom remains a key Google chip partner. (cnbc.com) Google has been trying to turn its in-house chips into a cloud selling point as it competes with Amazon Web Services and Microsoft for artificial intelligence workloads. Alphabet told investors in February that Google Cloud ended 2025 at an annual run rate of more than $70 billion, driven by demand for artificial intelligence products. (abc.xyz) (blog.google) Nvidia still dominates the market for the general-purpose graphics processing units that power much of today’s artificial intelligence boom. Google’s pitch with Tensor Processing Units is narrower: custom chips tuned for its own software stack and, increasingly, for cloud customers that want an alternative to Nvidia systems. (reuters.com) (docs.cloud.google.com) If the talks lead to a deal, Google would be adding another specialist chip designer as it tries to lower the cost of serving artificial intelligence models at scale. The immediate question is not whether Google will stop buying outside processors, but how much more of its artificial intelligence infrastructure it can pull in-house. (reuters.com)
