Broadcom: Software, agent identity controls key to production AI

Chris Wolf, global head of AI and advanced services in Broadcom's VMware Cloud Foundation Division, said companies should not buy hardware before making software decisions. "That's a horrible idea, because you have to make sure that your software choices are compatible with the hardware you bought," Wolf said. He said software provides flexibility to switch between cloud and local models and to choose accelerators, while "sovereignty considerations" and "tokenomics considerations" should push enterprises to reserve frontier models for deep reasoning and use specialized local models where appropriate. Wolf said IT operations teams are now caught in the middle, hosting frontier models, small local models and fleets of agents on the same hardware pool. He added that many customers have "buyer's remorse" after purchasing turnkey systems, and that Broadcom's AI factory approach provisions infrastructure from bare metal to model runtimes and exports YAML files to clone clusters.

Clayton Donley, vice president and general manager of Broadcom's Identity Management Security Division, said governance is becoming the foundation of AI infrastructure as agents gain access to corporate data and application programming interfaces without an audit trail. "It's not happening in an environment where we have 50 years of figuring out how to deal with employees and giving them their rights. It's happening in a brand new world," Donley said. He said enterprises should treat agents as identities and follow three principles: identity, intervention and inspection. A practical first step, he said, is to passively watch agent traffic, identify agents and then replace employee-maintained Claude or OpenAI keys with centrally managed keys. "Now you can make sure they can't circumvent you by using their keys through some other app," he said.

Purnima Padmanabhan, general manager of Broadcom's Tanzu Division, said agentic applications are essentially microservices applications, making the platform a natural place to enforce control. She said Tanzu Platform Agent Foundations, now part of VMware Private AI Cloud, runs agents in a sandboxed environment where every model, tool and data set must be explicitly bound. "The right way to secure an agent is to put it in a black box and give it nothing, but then you won't get any intelligence," Padmanabhan said. She said curated data products handle chunking, vectorization and access control so agents do not need direct access to original sources.

Padmanabhan also stressed the importance of code trust, noting that Broadcom scans its commercial and open-source repositories with Mythos and is extending Spring and Java security work into Python and Node.js. She said June's patch release was the largest in 23 years of Spring stewardship, and that discovered vulnerabilities are not just low-level issues but can "completely bring down an enterprise."

Padmanabhan told CIOs that private AI is now a revenue opportunity and competitive advantage. "If you don't do it, it's a competitive disadvantage, because everybody else is going to be doing it," she said.