Because decapitation strikes are too easy, or at least too dangerous. A large country can mitigate this threat through dispersal of strike/retaliation capability, and by building early warning systems. However, early warning only really works reliably when a strike can be detected a substantial amount of time (“a few minutes”) before it takes effect.
This is the logic behind bans on orbital weapons (de-orbiting weapons given less warning than ballistic weapons) and on the mutual withdrawal of nuclear weapons from Cuba and Turkey at the resolution of the Cuban Missile Crisis. This is also the game theory logic behind countries like South Korea and Taiwan effectively outsourcing their nuclear deterrent to the United States. In The Dead Hand, David Hoffman quotes Zbigniew Brzezinski as saying that after detection, initial cross-checking, and message delay, the President has about seven minutes from launch to impact of a Russian (Soviet) ICBM to decide whether or not to order retaliation. That window might have gotten slightly larger since the 1980s, but not by much. The United States’ (and other existing powers’) nuclear strike capability is sufficiently dispersed not to be neutralized by local subterfuge, and highly likely to survive a a first strike. This further reduced the likelihood of a launch on false-positive early warning.