Shares of several cloud infrastructure companies fell sharply in the wake of new managed AI agent offerings, but analysts at Morgan Stanley contend that the market reaction misunderstands how those agents interact with the existing infrastructure stack.
Over the past week, the bank's infrastructure coverage saw heavy declines among names including Cloudflare, Akamai, Snowflake, Datadog, Dynatrade, and Elastic. Investors had treated the arrival of managed agent products from AI-native firms as a direct competitive threat to cloud infrastructure software.
In a different reading, Morgan Stanley's team, led by Sanjit Singh, argues that managed agent platforms primarily handle the orchestration layer - coordinating model calls, routing to tools, and maintaining agent workflows - while external execution, data, network services, and observability remain necessary to make agents functional.
"The point is that managed agents abstract away the complexity of coordinating agents, but still depend on external execution, data, network services, and observability to make agents useful," the analysts wrote.
The analysts see the more consequential effects emerging when agents are run at scale. They describe a potential future with "millions, billion or even trillions of AI agents," each capable of invoking multiple tools, accessing governed data stores, performing web searches, and downloading content. That scenario, the report contends, would materially reshape demand across the infrastructure stack.
For Content Delivery Networks (CDNs) and edge compute providers such as Akamai and Cloudflare, Morgan Stanley sees agent workloads acting as a multiplier of volume. More autonomous agents could generate additional web searches, application traffic, and demand for caching, routing, and security services.
Latency sensitivity in multi-step agentic workflows further raises the value of edge infrastructure. Morgan Stanley cited an Akamai blog post that notes a London-to-Virginia inference path adds roughly 28 milliseconds each way before a single token is generated - a delay that compounds with every sequential tool call.
"In other words, every extra model hop and tool call increases the value of proximity, caching, request routing, and traffic control at the edge," the analysts explained.
Data platforms also stand to gain, according to the Morgan Stanley note. Managed agent sessions are likely to require governed access to enterprise data, which could translate into increased compute and monetization opportunities for providers. The analysts point to Snowflake as an example where agent connections via MCP would invoke its SQL query engine and drive compute consumption while maintaining enterprise security controls. They also highlight Palantir's Ontology MCP as a mechanism through which agents could execute workflows and analyze real-time data that the company could monetize.
Observability and monitoring platforms may see demand rise as well. The proliferation of long-running, autonomous agents adds complexity and expands monitoring surface areas, the analysts argue. Tool failures, latency spikes, token overruns, and policy violations are cited as examples of issues that would sustain the need for platforms like Datadog and Dynatrace.
Morgan Stanley's view reframes managed agent offerings not as replacements for cloud infrastructure but as a change in where and how work is executed. If agent workloads scale as the analysts describe, the result could be greater traffic, more stringent latency requirements, increased data governance needs, and broader observability requirements across the infrastructure ecosystem.