Look at where the directive sits. In the metadata example `use cache` is written above an async function that selects title, price and description; in the review component it is the first statement inside the async function body [18]. The scope you attach it to is the cache unit, and the route around that scope keeps its own rendering mode [6].
The call semantics matter more than the syntax. Without the directive the function executes on every request; with it, Next.js serves the stored result until the revalidation window closes [12]. The directive takes an optional key and revalidation strategy [11]. Skip the key and Next.js derives one from the arguments, so a metadata function called with two product IDs produces two entries [14].
Multiply that out before you adopt it. Entry count is cached functions times distinct argument tuples [20]. A function keyed by user ID is one entry per user; five such functions behind one page are five per user. The invalidation surface you were trying to shrink at the route level comes back as cardinality at the function level.
The strongest paragraph in the piece reverses an assumption I have heard in review more than once. A dynamic route already pays server execution on every request, which the article argues is exactly why it benefits most from caching the expensive functions inside it [16]. Route-level static generation needs the whole page deterministic at build time [13], and that requirement is what made one request-time value expensive for everything around it.
The server-cost claim [9] transfers only if the expensive work is already separable into named functions, the arguments are low cardinality, and the cached output contains nothing request-specific. That last condition is why the product-page example streams personalized recommendations while roughly 90% of the markup stays static [5]. On the source's account, thrash is what happens when a dynamic component forces synchronous rendering, and the streaming boundary is what prevents it [8].
What the article does not supply is a measurement. The only quantity in it is that illustrative 90% [21], and it names no Next.js release or configuration flag for enabling Dynamic IO [22], which is worth knowing because it carries a note saying it was written with AI assistance under human review [19]. The review sample also stops mid-query at `await db.revi` [23], which is one way to demonstrate streaming.
