They held back. Had they used 32Gbps GDDR7, they would have reached 2.0TB/sec memory bandwidth. 36Gbps GDDR7 would have let them reach 2.25TB/sec. The GB202 also reportedly has significantly more compute cores, TMUs, ROPs, tensor cores and RT cores than the 5090 uses:

https://www.techpowerup.com/gpu-specs/nvidia-gb202.g1072

Maybe there is a RTX 5090 Ti being held in reserve. They could potentially increase the compute on it by 13% and the memory bandwidth on it by 25% versus the 5090.

I wonder if anyone will try to solder 36Gbps GDDR7 chips onto a 5090 and then increase the memory clock manually.