Google Pushes Vertical Stack

Google is heading into Cloud Next this week with a tighter pitch: buy the whole stack from one vendor, from chips and servers up through models and apps. (googlecloudevents.com) Cloud Next 2026 runs April 22-24 in Las Vegas, and Google has spent the past year turning its Tensor Processing Units, or TPUs, into a broader cloud product line rather than an internal Google-only advantage. Google’s event pages and blog posts put generative AI, infrastructure and security at the center of the conference. (googlecloudevents.com) That stack starts with custom silicon. Google said its Ironwood TPU, introduced at Cloud Next 2025, is its seventh-generation accelerator and delivers 10 times the peak performance of TPU v5p and more than four times better performance per chip than TPU v6e for training and inference workloads. (cloud.google.com) Google has also been filling in the software layers that make those chips usable to outside buyers. In an April 7 post, Google said TorchTPU lets developers run PyTorch natively on TPU infrastructure, a step aimed at reducing the switching cost for teams that already build on Nvidia-heavy tooling. (developers.googleblog.com) The business argument is that a cloud provider can tune the chip, the networking, the compiler and the model service together instead of letting customers assemble those pieces themselves. Google’s Ironwood engineering post says the TPU is “codesigned” with software including the XLA compiler plus JAX and PyTorch ecosystems. (cloud.google.com) That puts Google in a different lane from companies selling merchant accelerators, the off-the-shelf chips cloud customers can buy across multiple platforms. SiliconANGLE reported April 18 that Google is framing the market less as a race for raw chips and more as a contest between integrated platforms and modular accelerator strategies. (siliconangle.com) Reports published Monday suggest Google is still pushing deeper into chip design. Reuters said Marvell shares rose 7% in premarket trading on April 20 after The Information reported Google was in talks with Marvell to develop two new AI chips aimed at running models more efficiently. (usnews.com) Bloomberg reported the same day that Google is pursuing new chips to speed AI results and build on recent cloud deals with Meta and Anthropic. That matters because Google is no longer selling only access to computing capacity; it is selling a package in which the hardware roadmap supports the cloud sales pitch. (bloomberg.com) Google has been building toward this for years. Its latest TPU documentation says TPU7x, the first Ironwood release, is now the newest TPU available on Google Cloud and is designed for large-scale AI training and inference, which gives enterprise buyers a current product rather than a future promise. (docs.cloud.google.com) The immediate test comes this week in Las Vegas. If Google uses Cloud Next to tie TPUs, infrastructure, Gemini services and developer tools into one buying story, customers will be weighing a simpler question than “which chip is fastest” — whether to rent a platform or assemble one. (googlecloudevents.com)