The 20logN number is real, but it is a floor, not the whole story. If you multiply a 10 GHz source by four to reach 40 GHz, the phase noise rises by 12 dB in theory; in practice I budget another 2 to 4 dB for the multiplier and its buffers, plus any spurious tones the multiplier generates. The multiplier also converts amplitude noise into phase noise, so a clean-looking source at the input can still come out worse than the textbook calculation.
That is why the choice is rarely just about the multiplication factor. A fundamental mmWave VCO avoids the 20logN penalty, but it usually has a lower resonator Q, a narrower tuning range for a given phase noise, and more frequency pushing from supply and load changes. A lower-frequency, high-Q VCO followed by a multiplier can win at close-in offsets, especially when it is locked to a good reference, because the PLL suppresses noise inside the loop bandwidth and the reference noise, scaled up, is often still better than a free-running mmWave oscillator.
Outside the loop bandwidth the multiplication penalty returns in full, so the answer depends on which offset matters for your system. For a communications link, the integrated jitter over the channel bandwidth and the phase noise at 100 kHz to 1 MHz offsets are usually more useful than a single spot number. For radar, close-in noise and spurious tones matter more.
When comparing numbers, always state the carrier, the offset, whether the source was locked or free-running, and the multiplier chain. Without those, a phase-noise figure at mmWave tells you very little, and the 20logN calculation alone will make a multiplied source look better than it measures on the bench.