Layer 2s 101

A blockchain that thousands of independent machines all verify is, by construction, slow and expensive. That redundancy is the product — it is what nobody being in charge costs. But it puts a hard ceiling on how many transactions can fit, and when demand hits that ceiling, fees rise until people stop.

A layer 2 is the response: process transactions somewhere else, in bulk, then post a compressed record back to the base chain along with something that lets anyone check the result. You pay the base chain's security costs once for a whole batch instead of once per transaction.

The trade the whole design rests on

The base chain — the layer 1 — settles its own transactions and provides its own security. Everything on top borrows that security rather than creating its own. The genuinely important question about any layer 2 is therefore always the same: how much of the base chain's security does it actually inherit, and where does it substitute a promise instead?

A system that posts enough data for anyone to reconstruct and challenge its state inherits a great deal. One that posts only a summary and asks you to trust its operators inherits much less, whatever it calls itself.

Two ways to prove a batch

Rollups are the dominant design, and they split on how they convince the base chain that the batch is honest.

The optimistic challenge window is not a bug being optimised away — it is the mechanism. The system is safe because a dishonest batch can be disputed for long enough that someone will. Shortening the window shortens the safety.

Zero-knowledge here means the proof reveals that a computation was done correctly without redoing it, not that anything is private. It is a confusing name for a validity proof, and most zk-rollups are no more private than anything else.

The sequencer, which is where the honesty lives

Somebody has to decide the order of transactions on a layer 2. That role is the sequencer, and on most networks today it is run by a single operator — usually the team that built the network.

This is the part that gets glossed over, so state it plainly: a single sequencer can delay your transaction, reorder transactions around it, or stop accepting them. It generally cannot steal your funds, because the proof or challenge system constrains what it can post. Censorship and theft are different powers, and layer 2s typically remove the second while leaving the first.

Well-designed networks provide an escape hatch: a way to submit directly to the layer 1 and force inclusion if the sequencer ignores you. Whether that hatch exists, and whether an ordinary user could actually operate it, is a fair question to ask of any of them.

Bridges, and why they keep being the thing that breaks

Getting assets onto a layer 2 usually means locking them in a contract on the base chain and issuing a representation on the other side. That contract accumulates the deposits of everyone who ever crossed.

Which makes it a concentrated target holding a great deal of value, and bridges have accordingly been among the most profitable things to attack in crypto's history. A rollup's own bridge, secured by the same proofs as the rollup, is a different and safer animal from a third-party bridge connecting two unrelated chains — but “it is on a layer 2” is a statement about cost, not about safety.

What this means in practice

• Fees are a fraction of the base chain's, which is the entire reason most activity has moved. That part delivers.
• Withdrawing back can be slow. On an optimistic rollup, plan around roughly a week unless you use a third party who fronts you the funds for a fee — and that party is now a counterparty.
• Assets are not automatically portable. A token on one layer 2 is not the same token on another, and sending to an address on the wrong network is a common, unrecoverable way to lose funds.
• Two networks can share a name and share almost nothing else. The interesting facts are who runs the sequencer, whether data is posted to the base chain, and whether you can exit without permission.

The honest summary is that layer 2s are a real answer to a real constraint, and that the range between the best and worst of them is much wider than the shared label suggests.