Today: several thousand tx/sec in "rollups," just by compressing transactions to minimal information and using zero-knowledge proofs that everything is valid. That's in production. The same transaction data is available on chain but e.g. instead of a sig on each transaction you have a compact proof that every tx had a valid sig.

Later: data sharding, where you're doing the same thing but each Ethereum node only has to store a portion of that data. That will multiply rollup capacity by 23X.

Details from Vitalik here: https://vitalik.ca/general/2021/01/05/rollup.html