OK, so I cocked that one up. Apologies to just about everyone - researchers, PRs, other hacks, and especially you, distant reader - for missing the point.
However, there are benefits. I now have the research paper in question and, even better, some very nice emails from the researchers involved.
The key aspect of IBM's development is that yes, it really is much smaller than the Intel (and other) modulators. I didn't understand the key aspect of the physics involved - so, if you're sitting comfortably, it's time for Fun With Photons.
Let's start with basics. What IBM, Intel, NEC, HItachi and tons of others want to do is chop up beams of light really, really fast so they can carry lots of data - and they want to do this on a chip, rather than in a big external light chopper.
One of the best ways to do this is with the awesomely named Mach-Zehnder modulator (which I've now learned to spell). This works by splitting light along two paths, electrically altering the refractive index of one to delay its light by just the right amount, then recombining the two paths to one again. If you get the delay just right, the light from the first path is exactly out of phase with the light from the second, and they cancel out. Result, darkness. Get them back in phase, and you get your original beam - light is restored. Do it fast enough, and you've got your data rates.
Now, one of the fortuitous facts of physics is that silicon is transparent to infra-red light - and, if you alter the number of electrons in a piece of silicon, its refractive index at infra-red changes. So if you make the two legs of your Mach-Zehnder modulator out of silicon, and make one of those legs part of a semiconductor diode, you can put a charge on that diode, change the electron density in that bit of the silicon, and modulate your light with something that looks a lot like any other bit of semiconductor technology.
The limiting factors of such a design is that you need to have a certain number of electrons (or their absence) to affect enough light to make it work, so M-Z modulators tend to be quite long, at least as far as semiconductors go. And when you send light down lengths of transparent material - called waveguides - it leaks out of the sides unless you keep them substantially straight, and that means you can't have sharp bends. This means your splitters and combiners have to be quite long, with gentle curves.
What IBM has done that's clever is make the cross-sectional area of their silicon waveguides exactly the same as the wavelength of the infra-red light. I'd missed that. I'd got the wavelength of IR in my head as 1500 nanometres, which it is - in air. In silicon, which has a refractive index of 3.5, it's much closer to 500nm.
Lots of interesting things happen when you make a waveguide that light just fits into. The two important points, from IBM's point of view, is that the waveguide is really small, so you can fill or empty it electronically really fast with a very high density of photonic interaction, and that you can have really tight bends - it doesn't work in the same way as larger waveguides.
That means you can build your M-Z modulator far smaller than before; you can fold the non-modulated leg into a hairpin, have tiny and widely divergent spltters and combiners, and achieve a deep phase shift over a very short leg.
All this is so small, in fact, you can pack hundreds or thousands on a chip - it is much better suited for on-chip use than the Intel device, at least at these stages of their various developments - while using very little power.
So, thanks to Yurii Vlasov and Will Green of IBM for putting me right,with a great deal of grace,
I just wish the PRs would put more physics in the press releases to start with...