The internet was built around humans. A person visits a website, clicks through a few pages, and leaves. Traffic is steady and predictable enough that cloud companies could design their systems around it. AI agents break that model completely.
An AI agent handling a complex task might sit completely idle for hours, then suddenly fire off hundreds of database queries and web searches in ten seconds, then go idle again. The old cloud infrastructure had no good answer for this. Companies running AI systems had to pre-pay for computing capacity sized for those peak moments, even though most of the time that capacity sat unused. Think of it as renting a full warehouse because you occasionally need to store a lot of boxes, even though it is mostly empty.
AWS just launched a rebuilt version of its search and data retrieval system, called OpenSearch Serverless NextGen, designed specifically for this problem. The core change is simple: computing power and data storage are now billed separately. When an agent fires up a task, computing resources appear within seconds. When the agent goes idle, those resources drop to zero and the bill drops with them. AWS says this can cut infrastructure costs by up to 60 percent compared to the old approach of always keeping resources running.
The previous version had the same scaling problem it was trying to fix. As AWS's own team noted, even the prior serverless product required at least one instance running at all times, because compute and storage were coupled together. The new architecture breaks that dependency entirely, with compute scaling up to 20 times faster than before.
This matters beyond the technical details. If your company is buying software that uses AI agents, or if your IT team is building internal AI tools, the cost of running those tools is directly tied to this kind of infrastructure. Cheaper and faster infrastructure means vendors can price their products more aggressively, and it means internally built tools become viable at smaller scales.
AWS is not alone in doing this. Microsoft has updated Azure to handle the burst-and-idle traffic patterns that agents create. Databricks and Snowflake, both of which store and manage large amounts of corporate data, are repositioning themselves specifically as memory and retrieval systems for AI agents. Cloudflare, which handles a significant share of global internet traffic, launched similar agent-ready infrastructure last month.
The scale of machine traffic on the internet is already significant and growing fast. Bots currently account for about 31 percent of all internet traffic, according to Cloudflare's network data. AI-specific crawlers, agents, and search tools make up a growing share of that. Cloudflare's senior product leadership has stated that non-human traffic will exceed human traffic sometime in the first half of 2027.
For business operators, this plays out in two ways. First, the AI tools you are already using or evaluating, from customer service automation to procurement assistants to document analysis tools, are getting cheaper to run as their underlying infrastructure improves. Second, if your company runs any web-based services, the volume of automated traffic hitting your systems is rising steadily and will continue to do so. That has real implications for system costs, security, and analytics accuracy, since bot traffic inflates visitor numbers and distorts usage data.
The companies spending heavily to rebuild this infrastructure are doing so because they see agents moving from experimentation into real production use inside enterprises. The plumbing being built now is the foundation those deployments will run on.