During the opening keynote at the company’s Frontier Tour event in London, Microsoft CEO Satya Nadella discussed how the company would be addressing the growing concern of tokenomics.
Analyst Gartner has predicted that by 2030, a failure to adopt adaptive artificial intelligence (AI) governance runtime controls will triple the cost of AI deployments, eroding value. It warned that standard cloud billing alerts are a lagging indicator of overspending.
In its Cost governance is a non-negotiable addition to AI governance report, Gartner said: “Relying on these escalations hours or days after overspending has been identified delays mitigation and fails to prevent immediate curtailing of runaway inference costs at runtime.”
Discussing how Microsoft would be tackling potential runaway costs associated with overuse of AI inference, Nadella spoke about a Microsoft project that provides AI model orchestration, enabling organisations to select a model, either run locally or in the public cloud.
“We now have HydraFusion, which is a learned model where you are now able to take advantage of powerful local models or an RTX Spark,” he said. “You now have very capable coding models. You can start using HyperFusion whenever a task or one of the sub-agents of a task can be done and can be satisfied by using just a local model.”
According to Nadella, this means IT departments can essentially provide “an infinite token factory”.
The company made Project HydraFusion available on GitHub in September as a research preview. The project’s goal is to deliver what Microsoft describes as “frontier intelligence through runtime orchestration”. In the GitHub post, Microsoft said HydraFusion creates a full execution plan, choosing from models across multiple providers to draft, critique and revise, or cascade to more powerful models to complete a task.
Along with demos of the tools Microsoft offers for developing agentic AI, Nadella used the opening keynote to discuss how Microsoft was addressing AI trust. “It starts with having a set of principles of your trust system,” he said. “I describe them as principles of observability and governance, but you need to identify every agent and you need to be able to log every activity of these agents.”
Nadella said that when an AI agent uses an AI model, it is necessary to log its chain of thought in a human-readable form. This log needs to show that the agent has complied with the organisation’s governance and compliance rules. Given that multiple AI agents may be required to complete a task, each one needs to be uniquely identifiable.
Nadella also discussed how Microsoft Agent 365 provides the agentic AI control plane for IT operations teams. He said Agent 365 “holds every agent identified, and manages and governs their policies”. According to Nadella, it also supports security operations and the budgeting requirements for FinOps.
He described how AI agents could work internally in an enterprise, where some work as penetration testers to attack enterprise IT systems, while others are tasked with fixing the security holes identified.
“You now have red agents that are continuously simulating the attacks, and you have blue agents that are continuously defending,” said Nadella. “Then you have green agents that are continuously remediating the vulnerabilities they find.”
