Microsoft opens its Ignite conference in San Francisco on Tuesday with a quiet pitch aimed at the fastest-growing line item in enterprise AI: the cost of keeping agents alive. According to a Bloomberg report citing three people familiar with the plan, the Azure team has built a new hosting system that lets an agent save a snapshot of its work, shut down completely, and later reactivate from that snapshot when a new request arrives. While the agent is idle, the customer pays essentially nothing.
The economics explain why the feature exists. Agents that monitor inboxes, watch databases or wait for triggers burn compute around the clock even when nothing is happening. Meta chief executive Mark Zuckerberg said on his company's last earnings call that Meta had not launched its Muse assistant globally because of how many server resources running it continuously would consume. Microsoft's answer is to make waiting free: the agent's state lives in a snapshot rather than in allocated GPU memory.
The move also closes a competitive gap. Google Cloud launched a similar service, the Agent Compute Engine, on October 22, letting agents suspend and resume without paying for idle infrastructure. Bloomberg's sources said Microsoft wants Azure to undercut rival clouds and standalone agent startups on the total cost of running agents, and the snapshot system is the centerpiece of that argument at Ignite this week.
Trust is the second pillar of the pitch. The agents and their snapshots would live inside a sandboxed Azure environment, and Microsoft plans to demonstrate that software built with its vibe-coding tools — which move to broad availability this week after a September preview — stays inside sandboxes that are disconnected from the internet and isolated from other customers' data. The company intends to contrast that posture directly with the data-handling policies of Anthropic and OpenAI, according to the report.
Internal numbers cited by Bloomberg suggest the defensive strategy is working so far. Claude for Microsoft 365, the feature launched in May that lets Office users work with Anthropic's assistant inside the Microsoft suite, has weekly active users below one tenth of those of the paid Copilot, according to a Microsoft employee with access to the data. Those figures are Microsoft's own and have not been independently verified.
The competitive backdrop is tense. Last week Anthropic added features that let Claude produce spreadsheets, whiteboards and dashboards without leaving its interface, and Microsoft shares dropped about 3 percent on the news as investors revisited the "SaaSpocalypse" debate over whether AI agents erode traditional office software. Microsoft says the number of customers paying for usage-based Copilot features grew 75 percent in the September quarter, and at least two large private-equity firms now spend more on pay-as-you-go Copilot tools than they used to spend on their entire Office 365 subscriptions, Bloomberg reported.
The rivalry is also a dependency. Microsoft plans to invest 1 billion dollars in Anthropic's IPO, while Anthropic has committed 49 billion dollars over seven years to rent Nvidia servers through Azure — a contract that helps Microsoft offset its reliance on OpenAI as a leasing customer. At the same time Microsoft pays Anthropic billions per year to power Copilot features such as the Cowork agent, and is beginning to swap some of those models for cheaper OpenAI alternatives under its commercial agreement.
If the snapshot system lands as described, the practical takeaway for enterprises is that always-on agents stop being a flat infrastructure tax and become a usage bill. That favors exactly the kind of high-volume, mostly-idle workloads — monitoring, triage, scheduled research — that early adopters keep saying they want. Whether it shifts real workloads from Google Cloud and the agent startups will depend on the pricing detail Microsoft has yet to publish.
Comments (0)
Log in to join the discussion
Log InNo comments yet