Cloudflare used where it improves the product.

When visitors are geographically spread, or you serve substantial static assets and media. A CDN moves cacheable content closer to people; it does nothing for a slow origin, expensive database queries or personalised pages that cannot be cached. Measure where the time actually goes before buying one.

Vercel removes operational work — builds, image optimisation, caching and rollbacks are handled — at a cost that scales with usage. Self-hosting is cheaper at steady high volume and gives full control, but you take on the Node runtime, process management, cache coordination, image handling and zero-downtime releases yourself.

Edge functions suit small, latency-sensitive work with no heavy dependencies — redirects, geolocation routing, header rewriting, A/B assignment. They are a poor fit for anything needing a database connection pool, large libraries or long execution. Most application logic belongs in a regional runtime near the data.

Deliberately, or it becomes the main source of complaints. Decide what is cached, for how long, and what triggers a purge — ideally tag-based, so publishing one page does not invalidate everything. A caching strategy without an invalidation path produces stale content and people who stop trusting the site.

It changes what your origin sees. Visitor IP arrives in a header rather than the connection, SSL mode has to match your origin configuration or you get redirect loops, and aggressive caching rules can serve authenticated pages to the wrong person. Each is straightforward once known and confusing when it is not.

Your site is down, and that is the trade for not operating infrastructure yourself. Mitigation is mostly about detection and communication: independent uptime monitoring rather than relying on the platform's own status page, and a documented plan for what you tell customers. True multi-provider redundancy is expensive and rarely justified.

Usually on bandwidth, function invocations, image optimisations and build minutes rather than a flat plan. Costs are modest until traffic or server rendering grows, then rise faster than most teams expect. Model your actual page views and rendering strategy against the pricing before committing, and set spend alerts.

Easier than most vendor lock-in, harder than the marketing suggests. A standard Next.js application ports reasonably well; platform-specific edge functions, image handling, caching primitives and analytics integrations do not. Keeping those boundaries thin is what preserves the option.