What is a Layer 2? Optimistic vs zero-knowledge rollups, decoded
L2 TVL crossed $50bn in 2025, and EIP-4844 cut rollup costs by roughly 90% overnight. The optimistic-vs-ZK debate is no longer hypothetical — here is how to read it.
On 13 March 2024, at slot 8,626,176, Ethereum activated the Dencun upgrade and switched on EIP-4844 — proto-danksharding, the blob transaction type. Within a week, the average cost of an Arbitrum swap had fallen from roughly $0.40 to under $0.05, and Optimism reported a ~95% reduction in its L1 data costs. L2Beat‘s aggregate TVL chart, which had spent 2023 grinding sideways around $15bn, broke above $30bn within four months and crossed $50bn during 2025. The Layer 2 thesis — that Ethereum would scale by moving execution off-chain while keeping data and settlement on-chain — went from a roadmap slide to an operating reality in a single hard fork.
For anyone trying to understand what an L2 actually is in 2026, that timeline matters. The optimistic-vs-zero-knowledge rollup debate is no longer a thought experiment; both architectures are running production money, charging real fees, and competing on a roadmap that finally has dates attached. This piece walks through what a rollup actually does, where the two main families differ, where they are converging, and which numbers to look at when comparing them.
What a rollup is, defined by what it inherits
A rollup is a chain that executes transactions in its own environment but posts its data and a proof of correctness to Ethereum mainnet. The thing it inherits from Ethereum is settlement and data availability — meaning, if every node on the rollup disappeared tomorrow, the data on Ethereum would be enough to reconstruct the full state and let users withdraw their funds. That property is what separates a rollup from a sidechain. Polygon PoS, BNB Chain, and Avalanche C-Chain are sidechains: they have their own validator sets, their own security budgets, and no recourse to Ethereum if their validators collude. The Ethereum Foundation’s scaling page is the canonical reference on the distinction.
The two families of rollup differ on how they prove that the off-chain execution was honest. Optimistic rollups assume blocks are valid and rely on a challenge window during which anyone can submit a fraud proof. Zero-knowledge rollups attach a cryptographic validity proof to every batch, so there is nothing to challenge — if the proof verifies, the state transition is correct. Both write transaction data to Ethereum as blobs; the difference is in what kind of proof accompanies the data.
Optimistic rollups: cheap to run, slow to exit
Optimistic rollups are the production-dominant family. Arbitrum One and Optimism Mainnet alone hold north of $20bn TVL between them, and Base — built on the OP Stack — adds another $14bn. The shared design pattern: the sequencer (currently centralised on each of these chains) builds and orders blocks, posts the compressed transaction data plus state roots to Ethereum, and the network assumes those state roots are correct for a challenge window of approximately seven days. During that window, any honest party with the off-chain data can run an interactive fraud-proof game against the proposed root and slash the proposer if the root is wrong.
The seven-day window is the famous catch. A user who deposits ETH to Arbitrum sees it appear in seconds; withdrawing it through the canonical bridge takes a week. The bridge ecosystem — Across, Stargate, Hop, Synapse — exists to bridge this gap by extending short-term credit to withdrawers, charging a small fee (typically 5-30 basis points) for instant liquidity. The seven-day delay is not arbitrary; it is the upper bound on how long it could take for a fraud proof to be submitted and resolved under network congestion. There is active research, including the BoLD (Bounded Liquidity Delay) protocol on Arbitrum, to shorten it without weakening the security model.
Zero-knowledge rollups: heavy math, instant finality
ZK rollups replace the fraud-proof game with a cryptographic validity proof. After each batch, the prover generates a succinct proof — typically a SNARK or STARK — that the new state root is the correct result of applying the batched transactions to the previous state root. Ethereum verifies the proof on-chain in milliseconds; if it verifies, finality is immediate. There is no challenge window. The trade-off is that proof generation is computationally expensive, which historically made ZK rollups slower and more expensive to operate than their optimistic counterparts.
That gap has closed faster than most observers expected. zkSync Era, Scroll, Linea, and Polygon zkEVM are all production, all EVM-equivalent or EVM-compatible, and all settling validity proofs to mainnet on hourly-or-better cadences. Starknet, which uses Cairo rather than the EVM, has been live for years and has the deepest dedicated developer ecosystem of any ZK chain. The economics matter: a ZK rollup pays a proving cost per batch, often hundreds of dollars depending on circuit complexity, but it eliminates the need to run a separate challenge infrastructure and unlocks features like trust-minimised bridges that simply cannot exist on an optimistic chain.
The side-by-side
| Property | Optimistic rollup | ZK rollup |
|---|---|---|
| Validity assumption | Honest-minority fraud prover exists | Cryptographic — no assumption |
| Canonical withdrawal time | ~7 days | Minutes to hours |
| Per-tx cost (post-EIP-4844) | $0.01-0.05 | $0.02-0.20 (depends on prover share) |
| EVM equivalence | Type 1-2 (full) | Type 2-4 (varies) |
| L1 verification cost | State root commitment only | Proof verification (~200-500k gas) |
| Production examples (TVL 2026) | Arbitrum ($14bn), Base ($9bn), Optimism ($6bn) | zkSync, Scroll, Linea, Polygon zkEVM, Starknet |
EIP-4844 and the blob economy
The single most consequential thing that happened to L2 economics in the last three years was EIP-4844. Before Dencun, rollups posted their compressed transaction data as calldata, which competed with every other use of Ethereum block space. After Dencun, rollups post data as blobs — a separate fee market with its own base fee and its own gas-limit equivalent. Blobs are pruned by the consensus layer after roughly 18 days, which is fine: they only need to be available long enough for any party to download the data and reconstruct the rollup state.
The result was a step-function drop in costs. A typical Optimism transaction that cost 40-80 cents in early 2024 cost 2-5 cents by late 2024. Arbitrum saw similar reductions. The blob fee market has its own dynamics — when many rollups post at once, the blob base fee can spike sharply, which is exactly what happened during the Pectra activation week and during high-volume memecoin events on Base. Track current blob fees on our gas dashboard, and watch the events calendar for the next protocol upgrade — full Danksharding will expand blob capacity from the current 3 target / 6 max per block to something closer to 64.
The decentralisation roadmap, by stage
L2Beat’s stage classification — Stage 0, Stage 1, Stage 2 — is the most useful single rubric for comparing L2s on a security basis. Stage 0 means training wheels are on: the team can upgrade the contracts with little or no delay, the sequencer is fully centralised, and exits go through the team’s bridge. Stage 1 means the contracts have a meaningful timelock (typically 7-30 days), there is a permissionless escape hatch, and the proof system is operational. Stage 2 means the only allowable upgrades are bug fixes through a multi-month process, and the system is functionally final.
- Stage 0 (most L2s): Operator-controlled, fast iteration, real but bounded user risk during early deployment.
- Stage 1 (Arbitrum, Optimism, Base, dYdX v4): Timelocks in place, fraud or validity proofs operational, exit via L1 contracts possible if sequencer goes dark.
- Stage 2 (none yet at scale): Fully trust-minimised. The aspirational endpoint.
This classification is not theoretical — it has been used in real-world legal and risk disclosures, and it is the right anchor for any due-diligence conversation. L2Beat’s risk page shows the current stage for every tracked L2.
Sequencers, MEV, and the centralisation footnote
Every major L2 in production runs a single sequencer today. That sequencer orders transactions, builds blocks, posts data to L1, and — crucially — collects all priority fees and MEV. Sequencer revenue is the operating margin of an L2: Base reportedly cleared more than $100m in sequencer profit in 2024, all of which flowed to Coinbase. Arbitrum and Optimism publish theirs through governance reports. The single-sequencer model is the L2 ecosystem’s biggest pending governance question, and the shared-sequencer projects — Espresso, Astria, Radius — are the leading attempts to solve it without making each rollup re-implement consensus.
The MEV story on L2s is also different from mainnet. Many L2s offer encrypted mempools, FCFS (first-come-first-served) ordering, or pre-confirmations, all of which mechanically eliminate some categories of MEV — particularly sandwich attacks. Arbitrum’s classic FCFS rule made sandwiches essentially impossible until the chain moved to Timeboost in 2024, which introduced a fast-lane auction that brought sandwich-style behaviour back in a controlled form. Our deeper market dashboard tracks the MEV revenue split per L2.
Where the two families are converging
The most interesting trend of the last 18 months is that optimistic and ZK rollups are converging on a shared end state. Several optimistic stacks now have ZK fraud-proof modes in production or testnet — Optimism’s Cannon and Kona, Arbitrum’s BoLD, and the Risc0 zkVM integration with the OP Stack. The idea is that you keep the optimistic operational model for everyday flow but use a ZK validity proof to resolve any fraud challenge instantly, collapsing the 7-day window without changing the upgrade contract. On the ZK side, every major chain has moved or is moving toward Type 1 EVM equivalence — meaning bytecode-level compatibility with mainnet Ethereum — which is the last technical gap between the two families.
In two years it is plausible that the “optimistic vs ZK” framing will look as quaint as “POW vs POS chains” looked in 2022. What will remain is the more honest distinction: who runs the sequencer, who controls the upgrade key, what stage is the chain at on L2Beat, and how much do users actually pay per transaction. Those are the numbers to anchor on.
How to read an L2 in 2026
The short version: ignore branding, look at L2Beat stage, look at TVL on canonical bridges (not on third-party bridges), look at sequencer revenue and where it goes, look at the upgrade keys and their timelocks, look at the proof system and whether it has ever been challenged. Compare the per-transaction cost on our tools page and the upgrade calendar on the events page. For most users, the right L2 to be on is the one where the apps they use are deployed, the bridge they trust has a corridor, and the stage is at least 1. The architectural debate matters for the people building these chains; for everyone else, the operational details matter more.
The L2 thesis has gone from “promised future” to “the place where most Ethereum activity already happens.” The next protocol upgrades, the move to full Danksharding, and the eventual decentralisation of sequencers are the three storylines worth watching. Everything else is implementation detail.