NetworkNews

Cloudflare: machine traffic does not equal origin load

On this page
  1. Correct the speaker and the measurement boundary
  2. A simple counterexample to automatic server growth
  3. Measure the work your application actually performs

Cloudflare’s statement concerns traffic on its network. It is not a census of every Internet connection, and the number of edge requests alone does not size an application server.

Fictional cacheable HTTP examples: one million requests at 95% hits leaves 50,000 origin requests; two million at 99% leaves 20,000, or at 80% leaves 400,000. Simplified model excludes revalidation and background fetching; not Cloudflare measurements.
Fictional cacheable HTTP examples: one million requests at 95% hits leaves 50,000 origin requests; two million at 99% leaves 20,000, or at 80% leaves 400,000. Simplified model excludes revalidation and background fetching; not Cloudflare measurements. Chart : PeopleAreGeek. Data source.
View full-size image

Correct the speaker and the measurement boundary

In the official August 6 earnings transcript, pages 6 and 8, CEO Matthew Prince says non-human traffic exceeded 50% on Cloudflare’s network during Q2. The earlier article wrongly attributed the forecast to CFO Thomas Seifert. Prince describes earlier forecasts being overtaken and offers a conditional five-year extrapolation, not a measured future outcome.

Non-human traffic is a broader category than malicious traffic or a single type of AI crawler. Nor does a majority of requests establish a majority of bytes, unique visitors or customer purchases. Those denominators answer different questions.

A simple counterexample to automatic server growth

Our cover uses fictional, cacheable HTTP requests. At one million requests and a 95% hit rate, 50,000 reach the origin. At two million requests and a 99% hit rate, only 20,000 do. Edge traffic doubled while origin requests fell by 60%.

With the same two million edge requests but an 80% hit rate, origin requests become 400,000: eight times the first scenario. Both outcomes are compatible with growing traffic. This simplified calculation excludes revalidation, background fetches and other cache behavior; it demonstrates why the hit rate and request class matter.

Measure the work your application actually performs

For an operational investigation, split the logs by route and response behavior. Repeated retrieval of the same public document and a personalized search that runs a database query can have very different costs. Two origin requests can also differ substantially in CPU time and response size.

Track origin request rate, response time and resource use alongside cache status. Compare the same observation window and state whether the bot classification is supplied by a vendor or inferred locally. That connects the headline to a capacity decision without turning a network-wide observation into a universal rule for every website.

September 8: correct speaker to CEO Matthew Prince and scope to Cloudflare network; separate observed Q2 majority from extrapolation and origin-load arithmetic.