Google's AI capex push
- Google told investors it will spend $175–185 billion this year, with just over half earmarked for machine-learning compute. - At Cloud Next it unveiled the Ironwood TPU and previewed an eighth-generation split between training and inference chips at TSMC 2nm. - The move aims to make Google Cloud the default substrate for enterprise AI, and Google even signed a multibillion-dollar infrastructure deal with Thinking Machines Lab ( ).
Google is telling investors and customers the same thing: it plans to spend up to $185 billion in 2026 to build more artificial-intelligence computing capacity. (abc.xyz) Alphabet said on its February 4 earnings call that 2026 capital spending will land between $175 billion and $185 billion. Reuters reported on April 22 that “just over half” of that budget is earmarked for machine-learning compute. (abc.xyz; reuters.com) At Google Cloud Next in Las Vegas on April 22, Google paired that spending plan with new products for companies building AI agents, the software assistants meant to carry out tasks across business apps. Reuters said Google presented agents as a core part of how it plans to turn AI into enterprise revenue. (reuters.com; blog.google) The hardware pitch starts with a simple split: training chips teach a model by processing huge amounts of data, while inference chips run the trained model when customers ask it questions. Google said its eighth-generation Tensor Processing Units will separate those jobs into TPU 8t for training and TPU 8i for inference. (blog.google; thenextweb.com) Google also introduced Ironwood, a seventh-generation Tensor Processing Unit built for inference, and said it delivers 4,614 teraflops of peak compute per chip. The company said Ironwood can be linked into pods with up to 9,216 chips. (blog.google) The next step is smaller manufacturing and more specialization. The Next Web reported that Google plans to build the training and inference versions of its eighth-generation chips on Taiwan Semiconductor Manufacturing Co.’s 2-nanometer process, with Broadcom on training and MediaTek on inference. (thenextweb.com) Google is not betting only on its own chips. TechCrunch reported on April 22 that Mira Murati’s Thinking Machines Lab signed a multibillion-dollar deal to expand its use of Google Cloud infrastructure, including systems powered by Nvidia’s GB300 graphics processors. (techcrunch.com) That mix of custom Tensor Processing Units and rented Nvidia systems shows how Google Cloud is trying to serve both sides of the market: customers that want Google-designed hardware and customers that want the Nvidia stack they already know. Google said nearly 75% of Google Cloud customers are using its AI products, and 330 customers processed more than 1 trillion tokens each over the past 12 months. (blog.google; techcrunch.com) The backdrop is a cloud market where Amazon, Microsoft and Google are all racing to add data centers, power and chips fast enough to meet AI demand. Google’s message this week was that it wants enterprises to buy the agents, the models and the compute from the same provider. (reuters.com; blog.google)
