Cloud players shifting compute mix
Big cloud vendors are diversifying the chips that run AI, not just buying more GPUs, and that shifts who benefits from AI spending. (Intel is expanding AI/cloud ties with Google beyond Xeon into IPU territory, and AWS is training Anthropic’s Mythos on its Trainium chips.) (x.com) That means investors and infra planners should watch non‑GPU accelerators and the software stacks that make them useful, since model operators are already experimenting with alternatives.
Amazon Web Services and Google Cloud are both still spending heavily on artificial intelligence, but the new twist is that they are spreading that spend across more kinds of chips than the usual graphics processing units. Intel said on April 9 that Google will keep deploying Xeon central processing units and expand joint work on custom infrastructure processing units, while Anthropic said Amazon remains its primary training partner and works with it on Trainium chips. (intel.com) (anthropic.com) A cloud data center does not run on one chip type any more. Central processing units handle general computing, graphics processing units do the dense math used in many artificial intelligence jobs, and infrastructure processing units move networking and storage chores off the main server like a warehouse loading dock that keeps trucks from blocking the factory floor. (intel.com) (cloud.google.com) Google has been using that split for years in public cloud machines. Its C3 instances were introduced with fourth-generation Intel Xeon processors plus a custom Intel infrastructure processing unit, which Google said improved how storage and networking work around the main compute chip. (cloud.google.com 1) (cloud.google.com 2) The April 9 Intel announcement pushes that model further into the artificial intelligence buildout. Intel said Google Cloud will use Intel Xeon 6 processors in C4 and N4 instances and will expand co-development of custom application-specific integrated circuit infrastructure processing units for large-scale systems. (intel.com) (intc.com) Amazon is making a different bet with Trainium, which is its in-house chip for training artificial intelligence models. Anthropic said in late 2024 that it was working with Amazon Web Services and Annapurna Labs on future Trainium generations, and Amazon said Anthropic had become its primary training partner for those chips. (anthropic.com) (aws.amazon.com) That partnership is no longer just a promise on a roadmap. Amazon said in November 2025 that Anthropic was already training and running inference for Claude on Project Rainier, and Amazon expected that footprint to scale to more than 1 million Trainium2 chips across direct usage and Amazon Bedrock by the end of 2025. (aws.amazon.com) This week added a second signal. Amazon Bedrock launched gated access to Claude Mythos Preview in the US East, Northern Virginia region on April 7, and Anthropic described Mythos Preview as a model built for advanced cybersecurity work, showing that one of the most demanding model developers is shipping new systems through Amazon’s stack while still using multiple cloud partners. (aws.amazon.com) (anthropic.com) Anthropic is also widening its own compute mix instead of picking one winner. On April 6, Google said Anthropic was expanding use of Google Cloud and Google tensor processing units, while Anthropic said Amazon remains its primary cloud provider and training partner, which means the same model company is actively spreading workloads across different clouds and different accelerators. (googlecloudpresscorner.com) (anthropic.com) That is why the money trail is getting harder to read if you only watch graphics processing units. Some of the value is moving to the chips that train models, some is moving to the chips that free central processors from networking overhead, and some is moving to the software work that makes a model run on Trainium, tensor processing units, Xeon, or a custom infrastructure processor without breaking. (intel.com) (anthropic.com) (cloud.google.com) The old picture was simple: artificial intelligence demand went up, so graphics processing unit demand went up with it. The new picture is that hyperscale clouds are building mixed fleets, and the winners increasingly include whoever supplies the extra chip, the compiler, and the cloud service that makes those extra chips usable at production scale. (intel.com) (aws.amazon.com)
