Angular's built-in i18n tags templates (i18n attributes, $localize in code), extracts XLIFF/JSON, and can compile a locale into the bundle. Each language is a build (or a set of files with localize). Switching language at runtime then means reload or a second deployment artifact. Runtime libraries (@ngx-translate/core and similar) keep one bundle and swap a JSON map. That is slower to fail at compile time and easier to hot-swap for user preference.

Build-time: $localize and extract-i18n

html
<h1 i18n="@@home.hello">Hello</h1>

IDs (@@) should be stable. Extract in CI, send XLIFF to translators, import, ng build --localize. ICU plurals and selects are in the syntax. You get unused-string checks and no missing key at runtime if the pipeline is strict. Cost: N locales ≈ N builds (mitigated by Angular's localize in one process for many locales, still more output). Locale-specific lazy chunks need care.

RTL: dir and locale data (registerLocaleData) for pipes (DatePipe) are separate from message catalogs. A translated string in an LTR layout is still a bug.

Runtime maps

A TranslateService and en.json is how many teams start. Missing keys show the key in production unless you lint. Changing copy without a redeploy is a plus for marketing sites and a versioning mess for apps that need legal-reviewed strings. SSR must load the right map per request (Accept-Language or cookie) or you hydrate-mismatch.

Hybrid: $localize for app chrome, CMS for editorial pages.

Choosing

Regulated product, few locales, strong CI: build-time Angular i18n. Many locales, translators in a CMS, language switch without reload: runtime, with a fallback and a test that every key exists. Don't mix two frameworks on the same view.

Read Angular's i18n guide for ICU and extraction, then time a locale switch. If product requires instant switch on a huge app, runtime wins. If product requires "the German legal string is in the signed artifact," compile-time wins. That is the trade, not "which blog post used ngx-translate in 2018."