For run 2 of LHC they used 50000 CPU cores for their software triggers, after the hardware trigger has reduced the 40 MHz input rate down to 1 MHz. The final output of the software triggers is 12.5 kHz, which is persisted to disk. Keep in mind this is just for the LHCb detector.
For run 3, they're planning to remove the hardware trigger bit, running the software triggers directly from the 40 Mhz signal. This would allow them to reprogram the triggers during the run, in case some new interesting theory comes along which for some reason has a signal their current trigger won't identify.
The computation side of the LHC is really impressive. For a full software trigger, you have 25 nanoseconds in which to load all the raw collision data, reconstruct 100's of particle tracks, calculate their momentum, join them up to figure out their decay vertices etc etc, and then, decide whether to store the event.
I recall LHCb could afford higher trigger-rates than CMS/Atlas, an LHCb event is smaller (~100 kB vs 1 MB for CMS) because the detector only covers 300 milli-radians from the collision-axis, in one direction, whereas CMS/Atlas have full coverage.
I don't know what the author is proposing. We don't have enough storage to persist all collision events, that would require zettabytes of disk space. The detectors are bottlenecked to store a few hundred events a second. Therefore they need to filter out the majority of the 40 million collision events per second that occur in the LHC.
Even then, I recall around 1% or so of the stored events were saved via a 'minimal bias' trigger, one that doesn't apply any filter criteria. This was mainly for calibration purposes, and cross checking stimulation data. So we still have petabytes of collision events that didn't have any selection criteria applied.