The zillions of sources of EM/thermal/etc noise. It's so prevalent, most of the time the effort is spent in trying to get rid of it.

Very wideband Zener noise sources are a common piece of measurement equipment; and many folks built their own noise sources (also for RNGs) using Zener junction breakdown. While building my own I needed quite a bit of shielding to avoid interference issues by RF sources (mostly HF fluoerescent lights and LEDs), so these sources might be susceptible to bias by EMI...?

Other junction breakdowns have noise as well, resistive noise is very low power. Atmospheric noise and background noise is a thing too, but I'd guess they are too susceptible to interference by malicious actors...

Such sources are always post-processed to deal with bias and correlation. No sane implementation just takes the straight measured output, whether it's keyboard interrupt timing or radioactive decay.

the random noise from the camera sensor is plenty good randomness. It would work just as well without the lava lamps.

If you want the same kind of visual display but with much lower power consumption, a much higher bit rate, and better reliability, I'd suggest a few plasma spheres, e.g.: https://www.scientificsonline.com/product/nebula-plasma-ball

There are an infinite sources of random audio out there that aren't susceptible to human interference.

busy streets webcams?

Put a fountain in the lobby and have a microphone listening to the splashes.

That would be susceptible to traffic patterns.

Cars don't look exactly the same and are not in the exact same places. Lighting changes. People look very different and walk different paths. It's going through PRNG so even very small pixel differences (which you can find even on an empty street thanks to changing weather, litter, wind and so on) are good enough if they are unpredictable. In other words, when you compare pictures of the same lava lamp with pictures of some random cars, I think you will be able to find more patterns in the lava lamp movement. You would also likely be able to predict it better (well that's the same thing I guess).

My guess is there’s still enough noise it doesn’t matter.

Not meaningfully. The input doesn't need to be uniform, just unpredictable. Statistical patterns are fine, as long as no one can predict exactly what's going to happen (to the pixel level). Even if all the drivers on the road were collaborating, they couldn't coordinate well enough to feed the camera adversarially predictable inputs. At least, not human drivers...

I was thinking more along the lines of busy/non-busy times may affect how much random data you get at different times of the day/week.

Not specifically some sort of giant coordinated attack to modify the randomness generated to specific values.