Let's simplify the world down to a single pixel sensor. It has a very simple API (this being modestly simplified, but not by much):

interface Pixel {
void SetZero();
uint16 GetValue();
}

If you want to take a picture, you tell the pixel to zero itself, then you open the shutter, wait a while, close the shutter, and finally read out the value of the pixel. You'll note that the pixel has no clue about ISO speeds and whatnot. It's just a little capacitor that builds up charge from photons that arrive.

When you read it out, you get an integer. If too many photons arrived, you'll get MAXINT. If none arrived, you get zero. To keep things simple, let's assume that the sensor is perfectly linear in its response to photons. Double the number of photons, double the integer you get back. (In real life, there's an analog-to-digital converter in there that tries to convert the charge in the capacitor into something resembling linear response. Let's ignore that for now.)

Instead, let's talk about mapping this linear uint16 to color values in a JPEG (which go from 0-255). Every sensor has a "base ISO" of some sort. This indicates that an exposure for that base ISO should get close to MAXINT and a standard 50% grey card should come out to MAXINT>>1 (i.e., half of maximum). The simple answer is to just take the high eight bits and call it a day.

When you're going to a higher ISO value, all you're doing is shifting the 8-bit window of that uint16. The high bits are zero anyway, so ignore them. Every bit you shift is exactly equivalent to doubling the ISO value.

So.... if your pixel gives 16 perfectly absolutely awesome noise-free bits, then you can shift 8 times. If your base ISO was 100, then your final ISO is 100 * 2^8 = ISO 25600. In reality, the low-order bits are noisy, and the lower you go, the less useful information you get. Furthermore, say you only had 14 real bits of information (i.e., the bottom two bits of that uint16 were always zero). You could still shift by 8, and you're still technically shooting at ISO 25600, but you've only got 6 bits of non-zero signal, and the low-order bits of that are going to be noisy crap.

To complicate things modestly, when you have a larger pixel, your base ISO might be higher. The Nikon D700's base ISO is 200, versus 100 on the D800. That means the D700 kinda gets the first shifted bit for free. (Downside: if you're shooting something that's stupidly really bright, you might find yourself wanting a neutral density filter on your lens. For most people, this is a non-issue.)

Whew!

What this boils down to is that many cameras will claim to support ISO 25600 or higher, but they're not exactly doing it very well. That's why DPReview (and others) conduct various tests to measure noise, resolution, and so forth. Many camera vendors will play post-processing games to try to clean up the mess they get from their sensor. You can run all manner of noise reduction algorithms, but to really clean up a mess, you're also going to get rid of very fine details. Or, you can just downsample the image altogether, pooling several adjacent source pixels into one final pixel. You give up image resolution, but you gain back the ability to have a non-crappy high ISO exposure.

If you're hoping to print something 6 feet tall, sharp as a tack, you'll be unsatisfied with this tradeoff. If, however, you're shooting for a newspaper or web site, and you don't need anything more than 1000 pixels across for the final image, you're going to be perfectly happy shrinking things down from the source resolution. (Which is exactly what newspapers and so forth do.)

After all that, I can finally address your question. If you're comparing a hypothetical 22 megapixel vs. 36 megapixel camera, otherwise made with the same generation of sensor technology, the 22 megapixel camera may have less noise in the low bits. The 36 megapixel camera has more pixels. When you reduce the image from both cameras to your target output resolution, they'll probably perform very similarly.

When you want to talk about the specifics of the Canon this vs. the Nikon that, you instead need to look at the test measurements. Those data seem to suggest that the Nikon D800 kicks everybody's ass. Today. We'll see what comes out tomorrow.