Anthropic's multi‑pronged infra push

Anthropic is trying to solve the same problem from three directions at once: build more compute, buy more compute, and make compute easier for customers to use. In one week, it was reported to be exploring its own artificial intelligence chips, it signed a multi-year cloud deal with CoreWeave, and it launched Claude Managed Agents for enterprise deployments. (cnbc.com) (investors.coreweave.com) (datacenterknowledge.com) The immediate trigger is demand. Anthropic said on April 6 that its revenue run-rate had passed $30 billion, up from about $9 billion at the end of 2025, and that the number of business customers spending more than $1 million a year had doubled from 500 in February to more than 1,000. (anthropic.com) That kind of growth turns chips into a supply problem. Training and serving a model like Claude means renting huge numbers of processors for months at a time, and the companies that control those processors can dictate price, delivery dates, and how much capacity you get. (cnbc.com) (anthropic.com) So Anthropic is looking at custom silicon, which is the chip industry’s version of owning the factory instead of leasing shelf space. Reuters reported on April 10 that the company is exploring building its own chips, but has not committed to a design and has not assembled a dedicated team yet. (cnbc.com) It is not starting from zero on the hardware side. Anthropic said this week that it had expanded its compute partnership with Google and Broadcom, with Broadcom helping ship custom Google Tensor Processing Unit chips for Anthropic workloads. (anthropic.com) (bloomberg.com) At the same time, Anthropic is still buying outside capacity wherever it can get it. CoreWeave said on April 10 that it signed a multi-year agreement to support development and deployment of Claude models, with compute coming online later in 2026. (investors.coreweave.com) CoreWeave matters because it is basically a specialist landlord for artificial intelligence infrastructure. Instead of waiting for the biggest cloud platforms to free up graphics processing units, Anthropic can add another supplier that is built around renting dense clusters of those chips. (thenextweb.com) (cnbc.com) The third move was aimed lower in the stack, at the customers trying to run agents inside their own operations. Data Center Knowledge reported that Claude Managed Agents is designed to handle the infrastructure needed to run production-grade agents at scale, instead of forcing enterprises to stitch together that plumbing themselves. (datacenterknowledge.com) That changes what Anthropic is selling. It is no longer just selling a model that answers prompts; it is selling a managed system that handles orchestration, sandboxing, and deployment work that usually eats up engineering time in data centers and enterprise software teams. (datacenterknowledge.com) (winbuzzer.com) Put those three pieces together and the strategy looks less contradictory than it first sounds. Custom chips promise long-term cost control, CoreWeave adds near-term capacity, and managed agents help convert raw compute into higher-value enterprise revenue. (cnbc.com) (investors.coreweave.com) (datacenterknowledge.com) The result is a company trying to avoid dependence on any single choke point. If Anthropic can split its future across Google Tensor Processing Units, rented graphics processing units, possible in-house chips, and software that makes customers stickier, it gets more leverage in every negotiation with cloud providers and chip vendors. (anthropic.com) (investors.coreweave.com) (cnbc.com)