RE: forecastting on supercomputers vs. desktops

A local area model running on a desktop is also a not doing anything clever - a local area model that's of any use has a resolution that's so much higher than a global model that it also needs a supercomputer. In fact it generally takes longer to run a good mesoscale model than it does to run a global one.

Then there's the computer resources needed to process GB upon GB of observation data from all the different satellites, radar, weather stations and balloons etc. to produce a good starting point for the model. That should be done by a good mesoscale model too and is computationally expensive (and for a global model it's very very expensive indeed).

Then to give a much better forecast, to get around the butterfly effect, many different forecasts can be run each with very slightly different starting conditions (within the errors of the observations). This ensemble of model forecasts can be used to quantify the probability of events occurring... It's particularly useful for severe weather events. Of course it also multiplies the computer cost by however many model runs you can do but the civil contingency people and utilities companies will pay good money for this. Sadly this sort of stuff never makes it to the BBC weather forecasts or the weather websites. There's a perception that the public are just too dumb to understand it (and they may be right: http://www.theregister.co.uk/2007/11/08/scratchcard_anarchy/ ).

Of course those smaller companies you mention might not actually have a better forecast but I'd bet good money that they're able to disseminate the information they get from their forecasts in a more useful (and profitable) way than many of the bigger organisations.