As the article says, the first example in the post does include the time that it takes to go out and fetch the static HTML that swaps in, and it added about a second to the server response of the experiment run that doesn't show up in the control run. For a big distributed site, a second may be more time than it would really take to put together a dynamic response.
Even with that 1-second additional delay included though, the improvement in that first test is still large (over 8 seconds faster in those tests). If the experiment took a few seconds longer on the server, it still would be 5 or 6 seconds faster to render content than the control.
I agree that server rendered would be faster, just the article had presented the absolute best-case scenario for the server. That said there could be other tradeoffs at play. Maybe loading the JavaScript and requesting a small amount of JSON each page is faster loading the initial page and then scrolling 10 pages down is faster than the server rendering out each page and appending it to the end.
Even with that 1-second additional delay included though, the improvement in that first test is still large (over 8 seconds faster in those tests). If the experiment took a few seconds longer on the server, it still would be 5 or 6 seconds faster to render content than the control.