Google–Thinking Machines Deal

Google Cloud has signed a multibillion-dollar infrastructure deal with Thinking Machines Lab, tying Mira Murati’s startup to Google’s latest Nvidia-powered systems. (techcrunch.com) Google announced the agreement on April 22 at Cloud Next ’26 in Las Vegas. The company said Thinking Machines will expand its footprint on Google Cloud for research, platform development, and frontier model training. (googlecloudpresscorner.com) The hardware at the center of the deal is Nvidia’s GB300 NVL72, a new Blackwell Ultra system built for running and training large artificial intelligence models. Google said Thinking Machines will use A4X Max virtual machines and will be one of the first Google Cloud customers on that setup. (googlecloudpresscorner.com) Cloud infrastructure is the rented computing backbone behind modern AI systems: the chips, networking, and storage needed to train models and answer prompts at scale. Google is pitching that backbone as a reason companies should build “agentic” software on its cloud, meaning programs that can take multi-step actions instead of only generating text. (cloud.google.com) At Cloud Next, Google said nearly 75% of Google Cloud customers now use its artificial intelligence products, and 330 customers processed more than 1 trillion tokens over the past 12 months. Sundar Pichai also said more than half of Google’s machine-learning compute investment in 2026 is expected to go to the cloud business. (cloud.google.com) (blog.google) Thinking Machines is one of the highest-profile new AI labs because Murati previously served as chief technology officer at OpenAI. The startup has kept a low public profile, so a named cloud supplier and a named chip platform give one of the clearest views yet into how it plans to build. (techcrunch.com) Google said Thinking Machines saw training and serving speeds double in early tests on A4X Max virtual machines compared with the prior generation of graphics processors. That kind of gain matters because model builders buy time on clusters by the week or month, and faster runs can mean lower cost or more experiments with the same budget. (googlecloudpresscorner.com) The deal also shows how the AI compute race is spreading beyond Microsoft, Amazon, and OpenAI’s long-running alliances. Google is using outside customers like Thinking Machines to show that its cloud can sell not just models and software, but scarce high-end capacity tied to Nvidia’s newest chips. (techcrunch.com) (cloud.google.com) For now, the clearest takeaway is simple: before Thinking Machines has shown a finished product, it has locked in one of the most expensive ingredients in AI — compute. (techcrunch.com)