Bots and AI agents now represent 31% of global HTTP traffic, according to Cloudflare. By H1 2027, non-human traffic is projected to exceed human traffic. AWS, Google, Microsoft and Databricks are redesigning their infrastructure to absorb the change.
Key Takeaways
- 31% of HTTP traffic is generated by bots; AI crawlers and agents represent 25% of all bot requests
- Non-human traffic is projected to surpass human traffic in H1 2027
- AWS, Google, Microsoft, Databricks and Snowflake are rebuilding cloud architectures to handle unpredictable agent traffic spikes
Have an AI Sum Up This Article
ChatGPT31% of the Web Is No Longer Addressed to Humans
The data comes from Cloudflare and covers the past six months. Of all HTTP traffic they observe, 31% is generated by machines: bots, crawlers, automation agents. Within that total, AI crawlers, search engines and AI assistants alone represent 25% of all bot requests. This is not a fringe phenomenon. It is the new normal operating state of the web.
The H1 2027 projection is more significant. At this pace, non-human traffic will overtake human traffic. In fewer than twelve months, the majority of requests on the web will come from machines, not from people behind a browser. The internet designed for humans is becoming shared infrastructure, and that cohabitation is already destabilizing architectures built around predictable human behavior.
The practical problem: AI agents do not behave like human users. A human loads a page, reads it, then loads another. An agent can fire dozens of simultaneous API requests, trigger sudden and brutal traffic spikes, then sit completely idle for hours. Traditional infrastructure, sized for steady human traffic, is not designed to absorb these usage patterns.
AWS has already responded with OpenSearch Serverless: a system that decouples compute from storage, scales up instantly during agent activity spikes, and scales down to zero during idle periods. Customers pay nothing when agents are inactive. That is a direct break from the previous model, which required maintaining at least one running instance at all times regardless of actual load.
An Industry-Wide Infrastructure Rethink
AWS is not alone in revisiting fundamentals. Google, Microsoft via Azure, Databricks and Snowflake have all announced or deployed architectural updates to better handle agent traffic. This is not a product improvement: it is a service model redesign, moving from “serve humans continuously” to “absorb unpredictable agent bursts.”
Vercel and Kiro are among the integration players also adapting. The full stack, from front-end to data warehouse, is being rethought for a world where the primary client is not a browser but an agent session. The security implications are significant. The spread of Shadow AI and the new attack surfaces it opens take on additional weight when agents are directly connected to data systems. For context, see our earlier piece on Horizon: Opus 4.8: Anthropic Ships Fast and Rewrites the Rules.
The cost impact is real and still poorly understood. Agent traffic does not resemble human traffic in terms of consumption patterns. Monthly billing models or fixed usage plans no longer match reality. Companies that deployed AI agents at scale are discovering cloud bills very different from their projections. The AI budget is no longer just the cost of tokens: it also includes the infrastructure cost of running them.
Microsoft experienced this in a visible way. Claude Code was suspended after consuming the entire allocated AI budget, illustrating what happens when autonomous agents operate without a predefined cost ceiling. This was not an isolated incident. It was the first warning of a structural pattern.
Also on Horizon:
- Vibe: Mistral Turns Le Chat Into a Full Work Agent
- Anthropic Series H: $65 Billion Round Approaches a Trillion
- Opus 4.8: Anthropic Ships Fast and Rewrites the Rules
Short Term, Medium Term: What This Changes for Teams
In the short term, infrastructure teams must revisit their capacity models. Load benchmarks based on human traffic are no longer valid in environments where agents run in production. Peaks can be ten times more intense than human peaks, and idle periods much longer. Autoscaling thresholds and alerts need recalibration.
Over the medium term, governing agent traffic becomes a central concern. Who can run how many agents, on which systems, with what consumption limits? These are architectural decisions that can no longer be left to product teams alone. Agent usage policy becomes as important as API governance and access security policy.
For content publishers and site owners, the signal is different but equally structural. Designing a site “for humans” is no longer sufficient. Accessibility for AI agents (LLM crawlers, AI search assistants) is becoming as important as traditional SEO accessibility. Tomorrow’s web will be consumed primarily by machines, and the players who anticipate this will maintain a lasting advantage.
Follow the story on Horizon.
