The cost list for headless Chrome in serverless (400MB+ layers and 3 to 8 second cold starts) is a fair summary of why page.pdf() becomes painful at scale. Moving compilation into the browser also removes the step of sending internal documentation to a remote worker, which matters for API docs that are not public yet. The comparison between window.print as a native vector stream and direct vector assembly is the part I would want to see benchmarked, since running headers and footnote counters are where those two usually differ.