It's not about storing off-chain, you compute off-chain and post to the network a cryptographic proof of the computation.
Rollups provide something akin to lossless compression proofs of computation. You execute the computation off-chain with high throughput and post a proof to the Ethereum network, this proof contains everything necessary to verify the correct execution of the computation. Sounds mind-blowing but we can thank smart mathematicians for figuring out the cryptography that makes it possible.
This means that we can scale off-chain while enjoying the security guarantees of the blockchain.
These techniques allow scaling the network throughput towards the 1000-4000 transactions per second range. Which together with data-availability sharding will push the throughput towards the 100K transactions per second range.
To be noted that this is not anymore a theoretical construction, these things are live running today on Ethereum. You can check the different solutions as well as detailed explanations of their trade-offs, limitations, etc... Here: https://l2beat.com