Unfortunately he used a relatively weak test signal (-20 dBm) and 10 dB of RF attenuation. So even though he overdrove the reference level by 10 dB, he got worse results than he should have beyond 30 kHz.
Yes, the DANL for the TR4172 is typically -146dBm/Hz and up at 1GHz it's more like -145dBm/Hz.
With the 10dB attenuator selected this will degrade to -135dBm/Hz. So for a -20dBm test tone this becomes -115dBc/Hz so it's no surprise to see the noise plateau at this level in that plot.
However, that analyser looks to be a couple of dB worse than this and I wonder if it has a problem? A while back my TR4172 developed an age related fault in the IF1 section where the adhesive they used for some RF absorber had gradually eaten away at a printed resistor trace (used as the bias resistor for the IF1 amplifier) and this made my analyser slowly go deaf and it also lost a lot of SFDR.
I removed the adhesive and used RTV instead and I fixed the part of the PCB with the damaged resistor. The noise floor improved back to where is normally was and the system IP3 went back up to normal. I think the production people at Advantest glued the absorber in place and immediately assembled and sealed the IF1 module and the adhesive then caused corrosion as it cured and the chemicals stayed entombed inside the module. It also rotted the nearby PCB traces making them go soft like rotten paint and it also caused all the plated screws to corrode. The sections without the adhesive still looked shiny and new inside.
The same thing happened inside the attenuator module. The adhesive was used on some RF absorber ant it rotted through an adjacent PCB trace and the analyser then lost an attenuator setting (now repaired). So I suspect that every TR4172 will have this issue at some point. I have a spare TR4172 here and the same thing was happening in both modules in that one too (but not as bad, as the PCB traces were soft but still intact)