Not all data is detector data. This is usually what we'd call Level 0 or Level 1 data. The lower levels aren't usually that interesting for analysis, because th

Not all data is detector data. This is usually what we'd call Level 0 or Level 1 data. The lower levels aren't usually that interesting for analysis, because they might not be properly calibrated or even reconstructed, depending on the detector. Reconstruction takes an immense amount of CPU time so it's not practical to do on the fly, so you have to store reconstructed data, and that takes up space. Then you might store (re)calibrated data and have several generations of that. Then you reprocess all of your data as a new pass, then you have several passes of data, constructed with different calibrations and algorithms. Not mentioned is simulation (monte carlo) datasets. Their size is usually within an order of magnitude of your actual data size. For some experiments, the amount of simulation data may eclipse the amount of actual data they have. I can't comment on the size and composition of ATLAS/CMS data, but my guess is the rate at which data is generated for the experiments, using a broader definition, is easily within an order of magnitude to that of Youtube.