IBM and CoreWeave co-design workload controls for AI agent infrastructure

IBM and CoreWeave Inc. have partnered to develop controls for AI agent workloads, focusing on infrastructure challenges such as workload isolation, identity management, and security. The collaboration aims to support IBM Research’s evolving needs, including reinforcement learning and the execution of agent workloads on CoreWeave’s compute infrastructure.

Partnership targets AI workload infrastructure challenges

According to SiliconANGLE, IBM and CoreWeave Inc. have partnered to co-design controls for agent workloads, addressing infrastructure challenges like isolation, identity management, and security. These issues have become critical as research teams shift from training models to running and testing AI agents at scale.

The collaboration aims to support IBM Research’s evolving needs, particularly as reinforcement learning introduces a task-execution stage that requires systems to accommodate interactive and dynamic workloads. This marks a shift from traditional model training to more complex, agent-driven workflows.

IBM’s infrastructure demands drive collaboration

IBM’s work on its Granite family of models required significant computing power, leading the company to build a large H100 cluster internally. However, SiliconANGLE reports that the cooling and power demands of newer hardware generations influenced IBM’s decision to partner with CoreWeave for additional infrastructure support.

The partnership has expanded into joint engineering efforts around identity management and workload controls. IBM extended its internal identity systems into CoreWeave’s platform, refining the implementation through iterative testing to ensure seamless integration and security.

CoreWeave Sandboxes enable secure agent execution

A key outcome of the partnership is the integration of CoreWeave Sandboxes, which provide isolated execution environments for agent workloads. These sandboxes allow researchers to run agent code on dedicated infrastructure or through a managed serverless runtime, with control over resource access and execution parameters.

IBM also measures the performance impact of security controls against benchmark results to balance tradeoffs between security and efficiency. This approach ensures that workload isolation and enterprise identity integration align with IBM’s broader security architecture.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

This partnership underscores a critical evolution in AI infrastructure: as teams move from training models to deploying and testing agents, the demands on compute, security, and workload management become far more complex. IBM’s collaboration with CoreWeave reflects a strategic approach to scaling AI research—combining internal expertise with external infrastructure to address challenges like workload isolation and identity management.

For founders and operators, this signals that AI workloads are no longer just about raw compute power. The ability to isolate workloads, integrate enterprise-grade identity systems, and optimize security-performance tradeoffs will determine which teams can scale efficiently. Startups focused on AI agents should prioritize these infrastructure details early, as retrofitting later can be costly and disruptive to operations.

Key takeaways

• IBM and CoreWeave are co-designing controls for AI agent workloads to address infrastructure challenges like isolation and identity management.
• Reinforcement learning introduces new demands, requiring systems to support interactive and dynamic workloads.
• IBM’s H100 cluster and partnership with CoreWeave were driven by cooling, power, and scalability requirements.
• CoreWeave Sandboxes enable isolated execution of agent workloads on dedicated or serverless infrastructure.
• IBM balances security and performance by measuring tradeoffs against benchmark results.

FAQ

What is the goal of the IBM-CoreWeave partnership?

The partnership aims to co-design controls for AI agent workloads, focusing on infrastructure challenges like workload isolation, identity management, and security to support IBM Research’s evolving needs.

Why did IBM partner with CoreWeave?

IBM partnered with CoreWeave due to the cooling and power demands of newer hardware generations, which required additional infrastructure support. The collaboration also enables joint engineering efforts around identity management and workload controls.

What are CoreWeave Sandboxes?

CoreWeave Sandboxes are isolated execution environments for AI agent workloads, allowing researchers to run agent code on dedicated or serverless infrastructure with controlled resource access.

How does IBM balance security and performance?

IBM measures the performance impact of security controls against benchmark results to optimize tradeoffs between security and efficiency in its infrastructure.

Related on Lazyfounder

• CoreWeave Inc.
• Cognition AI and CoreWeave unveil advances in AI agent infrastructure
• More technology news