So which one?

Architecture is all about tradeoffs. Neither is better; the right choice depends on what your screen needs from consistency, availability and latency. Measure the read and write paths here, then read the table with your own numbers.

Ready
GraphQL FederationEvent-driven projections
Query latencyHigher: the router waits for several services, in sequence where they depend on each otherLower: one local database
Data freshnessReal-time: every read asks the ownerEventually consistent: as fresh as the last event applied
Service couplingTight: every owner must be up for a full answerLoose: owners and readers meet only through the bus
Failure modePartial or failed answers when an owner is downStale answers when anything upstream is down
Read after writeImmediateAfter propagation; needs a plan (wait, optimistic UI, or read the owner)
Write complexitySimple: one mutation to the ownerMore: outbox, relay, idempotent consumer, replays
Moving partsRouter and subgraphsPlus topics, relays, consumers and projection stores
Rules of thumb. Federation suits screens that must show the truth this second and can live with an owner's outage taking part of the page with it (account balances, permissions). Projections suit high-traffic reads that combine many owners and can be a few hundred milliseconds behind (search, dashboards, feeds). Many systems use both: federate the few fields that must be live, project the rest.