Cloudflare’s AI Engineering Stack: What 8,700 Weekly Merge Requests Actually Measure

Cloudflare published detailed figures from its internal AI engineering rollout — and the most-cited number, a near-doubling of weekly merge requests, is also the one that requires the most careful reading.

What follows is analysis of figures a company published about itself. It is not investment advice, and it predicts nothing about adoption, engineering output or the tools market.

What Happened

On April 20, 2026, Cloudflare published a detailed post on its engineering blog describing the rollout of its internal AI coding stack. The figures — confirmed against the blog itself — cover eleven months of deployment across a company of roughly 6,100 employees. Tools named include OpenCode and Windsurf, frontier models from OpenAI, Anthropic, and Google, and Workers AI for open-source inference including Kimi K2.5 on security workloads. This is a company reporting publicly on its own internal operations; the numbers below are Cloudflare’s own.

The adoption curve is the cleanest part of the data. Cloudflare reports 3,683 active users against a stated ~6,100 employee base — which works out to about 60.4%, matching the 60% company-wide figure the post states. The parts agree with the whole, which is rarer than it sounds: most adoption percentages published without their numerator and denominator are unauditable, and most never supply the inputs needed to check them. This one does, and the check clears.

The headline figure is the merge-request count. The four-week rolling average of merge requests climbed from roughly 5,600 per week to over 8,700, with one week reaching 10,952 — described by Cloudflare as a level the company has “never seen” in a quarter-to-quarter increase. Cloudflare reports this as a count. It does not claim a productivity gain, and nothing below suggests otherwise.

The Structural Read

The Harness Theory lens is the right one here. Cloudflare is not building foundational AI models — it is harnessing them across its existing engineering organization, deploying frontier models from OpenAI, Anthropic, and Google alongside its own Workers AI infrastructure for open-source inference. The question Harness Theory asks is not “did you adopt AI?” but “what does the adoption actually change in the system?” And here, the data is carefully bounded in ways worth unpacking.

A merge request is an event in a version-control system of record. It increments when a change is necessary, when it is redundant, and when it is reverted the following week — and nothing in the count distinguishes those cases. That is precisely why the count is easy to produce: it was already instrumented. The measurable thing and the meaningful thing came apart, and the measurable one won because it required no additional inference. Throughput and value are different quantities, and only one of them is cheap to count.

The direction of the gap is not knowable from the figure either. A near-doubling of merge requests is consistent with substantially more work being completed. It is equally consistent with the same work arriving in smaller, more granular pieces — a well-documented effect of AI-assisted authoring, where the natural unit of a change shrinks. Both produce the same chart. Neither reading is the correct one without additional data that the post does not contain.

The adoption split — 93% across R&D against 60% company-wide — is informative in its own right. It locates the penetration in the function where the tooling fits rather than collapsing both populations into a single blended number. That decision reflects either disciplined reporting or organizational reality, probably both. Either way, the distinction matters: a 60% company-wide figure that included a 93% R&D spike and a much lower non-R&D base would mean something quite different from a uniform 60%. Cloudflare shows you the seam.

Three Implications