Anthropic Signs Multi-Year GPU Deal With CoreWeave to Power Claude at Scale
Bloomberg reported today that Anthropic has agreed to rent data center capacity from CoreWeave under a multi-year agreement. The deal spans a range of Nvidia chip architectures at US-based CoreWeave facilities, giving Anthropic a dedicated GPU supply outside of its existing cloud commitments with Google and Amazon. CoreWeave confirmed that nine of the ten largest AI model providers now use its platform.
The timing is notable: Anthropic just crossed $30B in annualized revenue, launched Claude Managed Agents in public beta, and is running Claude Mythos Preview for Project Glasswing — all workloads that demand a significant and growing compute footprint. The CoreWeave deal buys headroom and negotiating leverage as Anthropic prepares to scale beyond what any single hyperscaler can absorb.
Anthropic Revenue Surpasses $30B Annualized Run Rate — Up 3x From Year-End 2025
Anthropic’s annualized revenue run rate surpassed $30 billion in early April 2026, up more than 3x from the $9 billion figure recorded at the end of 2025. Enterprise contracts account for roughly 80% of total revenue per recent disclosures, driven by expanded deals with Snowflake, Accenture, AWS, and Google — all announced in the past week. The number underscores how quickly the frontier model market has shifted from API experimentation to mission-critical production deployment.
Anthropic Exploring Custom AI Chip Design to Reduce Compute Dependency
Anthropic is considering designing its own AI chips to address ongoing GPU supply constraints, according to three sources familiar with the matter. No timeline has been confirmed and the effort appears to be in early exploration, but the move would follow a well-worn path: Google, Amazon, and Meta all developed internal silicon to reduce reliance on Nvidia and gain cost advantages at scale. For Anthropic, custom silicon would also allow tighter hardware-software co-design for Claude’s inference and training workloads.
Advisor Tool Launches in Public Beta — Smarter Agents Without the Cost of Full Opus
Anthropic’s new advisor tool is now in public beta. It pairs a fast executor model (typically Sonnet or Haiku) with a high-intelligence advisor model (Opus 4.6) that provides strategic mid-generation guidance for long-horizon agentic tasks — all within a single API call. The idea: get near-Opus quality on decisions that matter while keeping latency and cost down for routine steps. To use it, include the beta header advisor-tool-2026-03-01 and add an advisor tool with type advisor_20260301 and model: claude-opus-4-6. Advisor tokens are billed at Opus 4.6 rates; executor tokens at Sonnet or Haiku rates.
Message Batches API Now Supports 300k max_tokens for Opus 4.6 and Sonnet 4.6
Anthropic raised the max_tokens cap to 300,000 on the Message Batches API for Claude Opus 4.6 and Claude Sonnet 4.6. Add the beta header output-300k-2026-03-24 to unlock long-form outputs in batch jobs — useful for generating full documents, large codebases, structured data exports, and other tasks that hit the previous 8K-64K ceiling. Standard Batches API pricing applies, with the 50% cost reduction versus synchronous calls still in effect.
Claude Code: Vertex AI Setup Wizard, Linux Subprocess Sandboxing, and Monitor Tool
The latest Claude Code release adds three notable features. An interactive Google Vertex AI setup wizard is now accessible directly from the login screen, making it faster to point Claude Code at a GCP project. Subprocess sandboxing now uses PID namespace isolation on Linux, hardening the execution environment against privilege escalation. And a new Monitor tool lets scripts stream events from background processes in real time, which is useful for long-running build, test, or data pipeline workflows running alongside active Claude Code sessions.
Anthropic Tightens Claude Subscription Rules for Third-Party Agents
Anthropic updated its usage policies to restrict how third-party agents can access Claude through consumer subscription plans. Per PYMNTS reporting, agents built by external developers will need to use the API directly under their own credentials rather than routing through end-user Claude.ai subscriptions. The change closes a grey-area in how automation tools and agent frameworks have historically piggy-backed on personal plans. Developers running agents on behalf of users should review the updated usage policies and migrate to API keys if they haven’t already.
Claude Cowork Gets Enterprise Features Alongside Managed Agents Push
Alongside this week’s Managed Agents launch, Anthropic quietly shipped enterprise-grade updates to Claude Cowork, the desktop AI tool for non-developers. Per 9to5Mac, the updates include centralized admin controls, audit logging, and expanded file-type support — features that enterprise IT teams have requested since Cowork launched. The move signals Anthropic is positioning Cowork as a serious enterprise productivity tool, not just a consumer companion app, and aligns with the broader enterprise platform push that defined this week’s announcements.
Anthropic’s Compute Strategy Is Now As Important As Its Model Strategy
CoreWeave deal, custom chip exploration, Google partnership, AWS deep integration — Anthropic is quietly assembling a multi-cloud, multi-vendor compute stack. That’s the right play for a company at $30B ARR with ambitions to go higher. A single cloud dependency is both a margin problem and a reliability risk, as anyone who watched the April 6-8 outages can attest. Diversifying compute is how you build the kind of SLA guarantees that enterprise contracts actually require.
But there is a real tension here. Anthropic’s safety mission has always been about moving carefully. Infrastructure races don’t reward careful. Every data center deal, every chip partnership, every managed service layer is a commitment to scale fast and keep scaling. At some point the question isn’t whether Anthropic can build a reliable platform — it’s whether the pace of that build is compatible with the cautious deployment culture that made Claude trustworthy in the first place.