It really wasn't. Itanium was a gamble that hard-coding static parallelism into the ISA would beat dynamic OoO. And just like architectural delay slots, it inevitably failed - using a few more transistors isn't that expensive since the alternative is "waste your time doing nothing at all".
"Compilers just need to keep up" was Intel's marketing apologia, not reality.
Compilers simply could not keep up - that was reality, not just marketing. Ideally, Intel should have coded the compilers too and removed the rough edges. But they were a bit too slow and AMD ate Itanium's target market.
You have to admit though that the EPIC (Explicitly Parallel Instruction Computing) model was quite innovative. The philosophy influenced the LLVM project and some of the principles are used in GPU's and AI accelerator chips, even if hardware-based dynamic scheduling won the game.
"Compilers just need to keep up" was Intel's marketing apologia, not reality.