[zorm]

Very few companies run the vertically-integrated stack because it is prohibitively expensive to do so with current NWP versus what you can sell it for with only marginal forecast improvements. I know several companies have tried this with integrating their own observation sources and ended up with worse performing forecasts. Oops.

I'm very interested to see how the ML modeling revolution changes this. The ability to perform global forecasts on a single GPU should make it cost competitive for more companies. I know several companies are already deriving their own weights for the forecasting component so that they can sell them. Google appears to be working on the next piece of the puzzle too with using ML for the data assimilation step, or skipping that altogether and using observations to go directly to forecasts.

[counters]

There are a few groups working on leveraging observations more directly in the ML forecast models and skipping over the assimilation/analysis step. However, unlike the original ML forecasting problem (which, let's be honest - was grossly over-simplified by the existence of ERA-5, which has been treated as "ground truth" for the atmosphere and used to teach models how to simply go from state at t=1 to state at t=1+\delta t), there's reason to believe that incorporating the observations will be substantially more difficult, given the complexity and bounty of the observations themselves and the challenge of framing a tractable, useful ML problem on top of them.