The data problem Flare is meant to help with
An exabyte stored, 40 million crossings a second, and about ten times more collisions coming. The numbers, with their sources.
The Large Hadron Collider is a ring 26,659 metres round, about 100 metres under the French-Swiss border, colliding protons at 13.6 TeV (CERN). Every number around its data is large, and CERN gives the size of the problem as its main reason for wanting a foundation model.
The numbers
| What | How much | Source |
|---|---|---|
| Proton bunch crossings | 40 million a second, one every 25 nanoseconds | CMS |
| Collisions seen by ATLAS | Up to 1.7 billion a second | ATLAS |
| Kept by ATLAS's triggers | About 1,000 events a second | ATLAS |
| LHC data stored so far | About one exabyte, a million terabytes | CERN |
| Open to anyone | More than 5 petabytes on the Open Data portal | CERN Open Data |
| After the upgrade | About ten times the LHC's design collisions and data | HL-LHC media kit |
| Upgraded machine starts | Mid-2030 (June 2030 in the September 2024 schedule) | CERN |
Where the LHC is now
Run 3, the current run of physics, was planned to end on 29 June 2026 (CERN). The LHC is now in Long Shutdown 3, the years of works that turn it into the High-Luminosity LHC. CERN's other accelerators delivered their last beams on 31 August 2026 and restart in stages from 2028, with the upgraded LHC last (CERN). That gap gives time to build new software, Flare included, before the data arrives.
Ten times or five times?
CERN's release and its upgrade media kit say about ten times more collisions. A CERN article from February 2026 says roughly five times more (CERN). The two figures probably measure different things: ten times the original design's total over the machine's life, against how much faster collisions arrive at a time. CERN hasn't put both side by side, so we quote whichever source we're citing.
Why this calls for AI
More collisions per crossing means more overlapping tracks to untangle, and more data to reconstruct with the same computing budget. Learned models can run on GPUs, and one model can be retrained when the detector changes. CERN has used AI since the 1980s (CERN); the change is the size of the model it now wants.