AWS Anthropic Partnership on Amazon Bedrock Expands, Meta Joins with Graviton Deal
AWS closed out April with a packed week of announcements. The biggest headline involves the deepened AWS Anthropic partnership on Amazon Bedrock, a deal that now reaches all the way down to the silicon level.
According to AWS, Anthropic is now training its most advanced foundation models directly on AWS Trainium and Graviton chips. The co-engineering occurs in collaboration with Annapurna Labs, optimizing performance from hardware through the full stack. That means builders on Amazon Bedrock are getting Claude models tuned for the infrastructure underneath them, not just ported to it.
Alongside that, Claude Cowork is now live inside Amazon Bedrock. Rather than treating Claude as a standalone tool, Cowork positions it as a true team collaborator. Enterprise teams can now deploy it inside their existing Bedrock environments, keeping data within AWS while running collaborative AI workflows. AWS also teased Claude Platform on AWS, described as a unified developer experience for building, deploying, and scaling Claude-powered apps, though it’s still listed as coming soon.
Meta made a significant move this week, too. The company signed an agreement with AWS to deploy tens of millions of Graviton cores for CPU-intensive agentic AI work. The scope includes real-time reasoning, code generation, and multi-step task orchestration, exactly the kind of demanding workloads Graviton was built for.
Beyond those headline stories, AWS shipped several meaningful updates. AWS Lambda can now mount Amazon S3 buckets as file systems using a feature called S3 Files. Built on Amazon EFS, this lets functions run standard file operations without pre-downloading data. Multiple Lambda functions can share the same file system simultaneously, a welcome upgrade for AI pipelines where agents need to persist memory across steps.
Amazon Aurora Serverless also got a performance boost, up to 30% faster than its previous version. Its new scaling algorithm handles competitive workloads better, including busy APIs and agentic AI applications with burst traffic. It still scales to zero, and all improvements arrive in platform version 4 at no extra cost.
For developers building agents, Amazon Bedrock AgentCore introduced a managed harness in preview, a new CLI, and coding-assistant skills. The harness lets developers define an agent using a model, a system prompt, and tools, then run it immediately without writing orchestration code. When teams are ready for more control, they can export the setup as Strands-based code. The AgentCore CLI is already available across 14 AWS regions at no extra charge.
On the cost side, AWS published a guide to granular cost attribution for Bedrock, which lets organizations tag and track usage by team or project. AWS also detailed how its DevOps Agent integrates with a Salesforce MCP Server to automate incident investigation from detection all the way to customer notification. In a separate move, AWS Skill Builder micro credentials are now free in all supported countries, hands-on, scenario-based assessments with no cost barrier.
For developers who want to stay current, AWS is hosting What’s Next with AWS virtually on April 28. In-person AWS Summits are also arriving across Singapore, Tel Aviv, Warsaw, Stockholm, Sydney, and several other cities through May and June. Full schedules are available on the AWS events page.
The AWS Anthropic partnership on Amazon Bedrock is clearly not slowing down. Between the Trainium co-engineering, Cowork’s enterprise rollout, and the Meta Graviton agreement, AWS is building a dense ecosystem around Claude, and around agentic AI more broadly. Developers who build on Bedrock will want to watch this space closely.
