A Layer 2 is a blockchain that settles to another blockchain. The L2 runs its own execution environment, processes user transactions, batches them up, and then periodically posts a summary (plus a proof of some kind) back to an L1 like Ethereum. The L1 does not need to know about every individual L2 transaction; it only needs to verify the summary. This lets the L2 handle far more throughput at a fraction of the cost, while ultimately relying on the L1 for security: if the L2 tries to cheat, the L1 can reject its summary, and users have a path to withdraw their funds to L1 even if the L2 operator goes rogue or disappears.
The main Ethereum L2s as of 2026 are Arbitrum, Optimism (and the Superchain ecosystem around it including Base and several smaller chains), zkSync Era, Starknet, Linea, Scroll, and Polygon zkEVM. Their combined TVL is now larger than Ethereum mainnet’s, and most everyday Ethereum activity — swaps, lending, NFT mints — happens on one of them rather than on L1.
Optimistic vs ZK
There are two main architectures, which differ in how they prove to the L1 that the L2’s state transitions are valid. Optimistic rollups (Arbitrum, Optimism, Base) post transaction data and a new state root to L1 and assume the state root is correct unless someone challenges it within a week-long window. If a challenger posts a fraud proof showing a specific step was executed wrong, the L1 recomputes that step and reverts the chain. Zero-knowledge rollups (zkSync, Starknet, Linea, Scroll, Polygon zkEVM) post a cryptographic validity proof alongside the state root, which the L1 verifies immediately. There is no challenge window, so withdrawals can be instant, but producing the proof is computationally expensive and the proving systems are newer and less battle-tested.
Optimistic rollups are currently further along in terms of production usage, EVM compatibility, and tooling. ZK rollups are catching up, with zero-knowledge proof generation getting cheaper every year, and most observers expect ZK to be the dominant architecture long-term because of the faster withdrawals and the cleaner security model.
The Cost Story
L2 transactions are dramatically cheaper than L1 transactions. A Uniswap swap on mainnet Ethereum might cost $5-$15 depending on gas prices. The same swap on Arbitrum or Base typically costs between two cents and fifty cents. The compression factor is roughly 10-100x, which is the difference between Ethereum being unusable for small transactions and Ethereum being usable for them. This cost difference is why retail activity has almost entirely migrated off mainnet since 2022.
The March 2024 Dencun upgrade introduced EIP-4844 (“blob transactions”), which gave rollups a dedicated, cheaper way to post data to Ethereum. After Dencun, L2 transaction costs dropped by roughly another 10x, reaching single cents for most routine actions. This was a watershed moment for the L2-centric scaling roadmap — it made L2s cheap enough that the original argument against them (“but Solana is already this cheap”) largely stopped applying.
Security Assumptions
L2 security is more nuanced than L1 security. Every L2 has a sequencer — an operator that orders transactions before they are batched to L1 — and during normal operation you are trusting the sequencer to not censor or front-run you. If the sequencer goes down or acts maliciously, most L2s have an escape hatch that lets users force transactions directly through the L1 or withdraw their funds, but this is an emergency mechanism, not a daily-use one.
There is also a spectrum of decentralisation across L2s. Some, like early-stage ZK rollups, have admin keys that can upgrade contracts or change parameters unilaterally — effectively, the team can pause or modify the chain. Others, like Arbitrum, have multi-sig control with time-locked upgrades and are on a public roadmap to further decentralisation. The site L2BEAT publishes a widely-used rating of each L2’s current security guarantees, and it is worth checking before you assume that “inherits Ethereum security” means “as secure as Ethereum in every respect”. It does not, at least not yet.