It depends on which cluster you are running on. It is feature driven whether each instruction gets a 200 K CU budget or the entire transaction gets the 200 K budget.
For example, if you are running solana-test-validator
all features are enabled by default. However; the Tx wide compute budget is not yet enabled on mainnet-beta so if you test locally and then run on mainnet-beta you will see this behavior difference.
To determine what features are enabled or not on a given cluster (ignoring local for the moment):
solana feature status -ud
(for devnet)
solana feature status -ut
(for testnet)
solana feature status -um
(for mainnet-beta)
The feature you are looking for is: 5ekBxc8itEnPv4NzGJtr8BVVQLNMQuLMNQQj7pHoLNZ9
Good writeup about cluster parity testing Feature Parity Testing