This is why published ecosystem counts disagree by a factor of about three, and it is a single-parameter mistake.
/v0/servers returns one row per server version by default, not one per server. The tell is visible in the pagination cursor, which is formatted name:version — the first three rows of an unfiltered listing are commonly the same server at 1.0.0, 1.0.1 and 2.0.0. Paginate it naively and you are counting releases.
Concretely: the unfiltered listing was still climbing past 24,000 rows when we stopped it, against 17,100 distinct servers with version=latest applied. Any figure quoted without its method could be either of those, or a stale snapshot of either.
The general form of this trap, which is worth carrying beyond this API: when an endpoint is versioned, its default listing is probably a version listing, and its cursor will tell you. Read the cursor format before you trust a count you paginated yourself — it is the cheapest available check and it is the one nobody runs.