What Are ZK-Rollups? zkSync, Starknet, Linea & Scroll

Introduction

One zero-knowledge (ZK) rollup closed this summer, another is retiring its virtual machine, and a third plans to turn private. The 11 ZK-rollups that L2BEAT tracks still secure $3.4B, about 9% of layer 2 (L2) value, yet only Starknet has reached Stage 1 (L2BEAT, as of 5 November 2026). L2BEAT renamed itself ETHBEAT on 4 November 2026, and this guide keeps the older name in citations. Behind those figures sit validity proofs, compatibility types and prover hardware that decide cost, speed and who can run the network. ZKsync, Starknet, Linea and Scroll took different routes to the same proof, and 2026 pushed all four off their earlier plans. This guide defines each in plain terms, places ZKsync, Starknet, Linea and Scroll, and uses the Polygon zkEVM shutdown to show what leaving a closed rollup requires of its users.

What Is a ZK-Rollup, and What Does a Validity Proof Actually Prove?

A ZK-rollup is a layer 2 (L2) chain that executes transactions off Ethereum and proves each batch correct before Ethereum accepts it. L2BEAT tracks 11 of them, holding $3.4B, about 9% of the $38.8B secured across the 83 layer 2s it lists ↗ (L2BEAT, about 10:45 UTC on 5 November 2026). The proof certifies the computation and nothing more.

From batch to proof to Ethereum

A sequencer, the operator that orders transactions, collects them into a batch and executes it off Ethereum. A prover then generates a validity proof, a short cryptographic argument that the batch moved the rollup from its old state to its new one under the rollup's rules. A verifier contract on Ethereum checks the proof and, only if it passes, accepts the new state root, one hash that commits to every balance and contract (ethereum.org, 2026). The batch data also goes to Ethereum, as calldata or blobs, so anyone can rebuild the state independently (ethereum.org, 2026). The wider split between layer 1 and layer 2 blockchains ↗ is set out in a separate guide.

What a validity proof does not hide

The name misleads. Zero-knowledge describes a property of the proof system, and rollups use it mainly for succinctness: a small proof that Ethereum verifies far more cheaply than it would re-execute the batch. Transactions on a ZK-rollup are public, as on Ethereum, and ethereum.org states that validity proofs do not by themselves make transactions confidential (ethereum.org, 2026). The proof also leaves other risks open. It shows the batch followed the rules coded in the verifier contract, not that those rules are free of bugs or that nobody can change them. Censorship resistance is a separate guarantee: published data cannot reveal a transaction the sequencer never included, so a forced-inclusion path through Ethereum matters (Alchemy, 2026).

Who runs the sequencer, prover and verifier

The sequencer orders transactions and gives the fast confirmations users see, the prover does the heavy computation, and the verifier is a smart contract, so Ethereum itself fills that role. L2BEAT lists whitelisted proposers for state roots on ZKsync Era and Linea, so withdrawals freeze if they stop, while Starknet relies on a Security Council minority to step in (L2BEAT, 2026). Scroll lets users submit proofs themselves through its source-available prover if its proposer fails (L2BEAT, 2026). Who may prove matters as much as how proofs work.

Optimistic rollups reach the same goal by assuming batches are correct and letting anyone dispute them. The comparison of optimistic rollups and ZK-rollups ↗ on exit time, cost and stage maturity is covered in a companion article.

The proof has to be built for something, and that something is Ethereum's virtual machine, which was not designed around ZK-friendliness (Buterin, 2022).

What Are zkEVM Types 1 to 4, and Why Does the Type Matter?

A zkEVM proves Ethereum's virtual machine (EVM), and its type states how closely it matches Ethereum. Vitalik Buterin's 2022 taxonomy runs from Type 1, identical to Ethereum and slowest to prove, to Type 4, fastest to prove and least compatible (Buterin, 2022). Every live ZK network sits somewhere on that line.

Type 1 and Type 2, Ethereum-identical

A Type 1 zkEVM changes nothing in Ethereum, so existing execution clients and tooling work as they are, and block explorers and block-production tools carry over (Buterin, 2022). Buterin wrote that proofs of Ethereum blocks then took many hours, and that Type 1 is what Ethereum's own layer 1 would need (Buterin, 2022). A Type 2 zkEVM keeps the EVM identical but changes structures outside it, such as the state tree and hash function, which speeds proofs while nearly all applications keep working (Buterin, 2022). Type 2.5 adds higher gas costs for operations that are hard to prove. Linea's team has described its network as Type 2 with Type 1 as the target (Linea community forum, 2025).

Type 3 and Type 4, faster proofs and less compatibility

A Type 3 zkEVM drops features that are hard to prove, usually precompiles, so most applications run and some need rewriting. A Type 4 zkEVM compiles Solidity or Vyper source into a ZK-friendly virtual machine, which gives the fastest proofs; contract addresses can change and bytecode-level tools may not work (Buterin, 2022). In that taxonomy ZKsync was Type 4. Starknet runs the Cairo virtual machine rather than the EVM, so it sits outside the EVM types (L2BEAT, 2026). Buterin noted that no team wanted to stay at Type 3 and that most planned to add the missing precompiles, and he treated the types as points on a trade-off, not a ranking (Buterin, 2022).

Where the live networks sit

The positions are moving. Scroll's Euclid upgrade in April 2025 replaced its halo2 circuits with a RISC-V prover called OpenVM and its custom state tree with Ethereum's Merkle-Patricia trie (Scroll docs, 2025). That narrows the gap Buterin listed for Type 2. Linea announced on 29 March 2026 that it would rebuild its prover around RISC-V, promising Type 1 compatibility "from day one", a vendor claim (Linea, 2026). ZKsync OS runs EVM code on a RISC-V proving system called Airbender (ZKsync docs, 2026). RISC-V, a simple open instruction set, is becoming the common base. Buterin hoped everything would become Type 1 over time, through better provers and a more ZK-friendly Ethereum (Buterin, 2022).

Data current as of November 2026.

Compatibility sets what developers can reuse, and proof speed sets what the network pays to run.

Why Does Proving Cost So Much, and What Hardware Does a Prover Need?

Proving is the extra bill a ZK-rollup carries. Vendor figures now put it near $0.0001 for a simple transfer on one GPU, yet proving still needs specialised hardware, and L2BEAT's data shows that requirement keeping prover and proposer roles permissioned on Starknet and ZKsync Era.

What proving costs

ZKsync's Airbender prover offers the clearest public figures. ZKsync says it proves a simple token transfer for as little as $0.0001 and an average Ethereum block in about 17 seconds on one Nvidia H100 GPU (ZKsync, 2025). Those are vendor benchmarks. ZKsync OS also charges a separate native resource that reflects proving cost, next to gas and data costs, and its docs base the price on the prover machine's cost and speed (ZKsync docs, 2026). Linea reports making its prover twice as efficient to cut infrastructure costs (Linea docs, 2026). Proving sits beside Ethereum data costs ↗ and does not replace them.

What hardware a prover needs

Ethereum's own targets show the scale. The Ethereum Foundation's 2025 real-time proving goals ask for proofs of 99% of mainnet blocks within 10 seconds, on hardware costing no more than $100,000 and drawing no more than 10 kilowatts (Ethereum Foundation, 2025). ZKsync said on 13 October 2026 that its Airbender Prime prover proved 99% of 7,200 consecutive mainnet blocks in under 10 seconds each on one Nvidia RTX PRO 6000 GPU, a result from its own test (ZKsync, 2026). ZKsync's open-source prover lists a GPU with at least 24 GB of memory for its fast proofs and about 140 GB of RAM plus a GPU for the final SNARK proof (Matter Labs, 2025). Buterin reported hours per Ethereum block in 2022 (Buterin, 2022). GPUs and newer proof systems, such as StarkWare's S-two, live on Starknet since November 2025, closed much of that gap (StarkWare, 2025).

Why prover hardware shapes who can prove

Hardware cost decides who can take the prover role. L2BEAT notes that running a Starknet prover requires significant infrastructure, so it is unclear the Security Council can cover the role quickly if the operator stopped (L2BEAT forum, 2025). ZKsync Era and Linea keep whitelisted proposers, and only Scroll lists self-proposing through a source-available prover (L2BEAT, 2026). ethereum.org warns that specialised hardware can concentrate proving in a few hands (ethereum.org, 2026). Linea's roadmap says a single-constraint prover lowers memory needs and widens hardware options (Linea, 2025). Cheaper proofs make open proving feasible, but no network has finished that work.

Those mechanics play out differently on each network, and 2026 changed course on all four.

Where Do ZKsync, Starknet, Linea and Scroll Stand in November 2026?

Only Starknet is Stage 1 on L2BEAT; ZKsync Era, Linea and Scroll are Stage 0 with no exit window (L2BEAT, as of 5 November 2026). Each also changed course in 2026: ZKsync is retiring its original virtual machine, Linea is rebuilding its prover, Scroll plans to go private and Starknet is studying a move to its own layer 1.

ZKsync Era

ZKsync Era secures $235M, of which about $112M is the ZK token itself, and ZK sits 99.1% below its all-time high (L2BEAT and CoinPaprika API, 5 November 2026). On 4 September 2026 Matter Labs announced it will retire EraVM, with a transition for EraVM chains starting within six months. Wallet-held funds need no action; guidance for funds in smart contracts is pending (Matter Labs, 2026). Era's execution delay rose from three to 24 hours on 7 October. ZKsync Lite, the first-generation network, stopped producing blocks on 4 May 2026 (ZKsync, 2026). The company now centres on Prividium, a private chain product that Germany's Bundesbank is testing (The Block, 2026).

Starknet

Starknet is the only ZK-rollup at Stage 1, with $363M secured, about $195M of it STRK (L2BEAT, as of 5 November 2026). It uses STARK proofs and the Cairo virtual machine, and L2BEAT lists no trusted setup. A shared prover service called SHARP serves it alongside other StarkWare chains (L2BEAT, 2026). StarkWare's S-two prover replaced Stone on mainnet in November 2025, and the Grinta upgrade put three sequencers into consensus, still operated by StarkWare (Starknet, 2025). Stage 1 rests on a Security Council minority registered as a second permissioned prover, a design L2BEAT questioned in December 2025 (L2BEAT forum, 2025). Upgrades can still be instant through a 9-of-12 council (L2BEAT, 2026). On 8 October 2026 Starknet said it was considering becoming an independent layer 1 to reach quantum resistance in 2027; a non-binding STRK poll that closed on 22 October backed researching the idea with 93.5% of the voting power cast, and any move would need a further community vote (Starknet Foundation, 2026).

Linea

Linea secures $342M, mostly ether, at Stage 0, and L2BEAT lists no mechanism to force transactions through Ethereum if its sequencer fails (L2BEAT, as of 5 November 2026). Linea, which scales Ethereum as a zkEVM rollup ↗, posts full transaction data to Ethereum, where most ZK-rollups post state diffs (L2BEAT, 2026). Its prover used direct arithmetization of the EVM, and on 29 March 2026 Linea said it would move to a RISC-V design (Linea, 2026). Linea's fee model burns 20% of net gas income as ETH and 80% as LINEA (Linea, 2025). Its release notes list finality under 30 minutes for a Q4 2026 beta, a roadmap item (Linea docs, 2026).

Scroll

Scroll secures $39M at Stage 0, the smallest of the four, though users can self-sequence through Ethereum with up to a seven-day delay (L2BEAT, as of 5 November 2026). A prover upgrade to OpenVM v2.0.0 and the SWIRL proof system finished on 22 September 2026 (Scroll governance forum, 2026). On 8 September the team said it plans to turn Scroll from a public chain into a private, application-specific network around its AI products over about nine months (Scroll forum, 2026). A formal proposal posted on 3 November asks the DAO to approve the switch to a permissioned network and a move of SCR to Ethereum mainnet as a single decision, with the upgrade planned for year-end; no vote had been held by 5 November 2026 (Scroll governance forum, 2026).

Data current as of November 2026.

One network on that list was shut down this summer, and its closure is the clearest test of what a user owes themselves before bridging.

What Did the Polygon zkEVM Shutdown Teach About Leaving a Closed Rollup?

Polygon zkEVM Mainnet Beta stopped producing blocks on 3 July 2026, so it is a case study and not a live option. Wallet-held funds are recoverable through a claims interface open until 31 December 2027; funds locked in smart contracts are not (Polygon, 2026). The shutdown changed nothing about how validity proofs work. It changed what users had to do.

What happened and when

Polygon Labs announced the sunset in June 2025 with twelve months' notice and repeated the warning on 22 April 2026 (Polygon forum, 2026). The sequencer was sunset on 3 July 2026. The network no longer produces blocks, and the Agglayer Bridge no longer processes withdrawals (Polygon, 2026). Funds locked in DeFi protocols were not migrated automatically, and every protocol on the chain stopped working once transactions stopped (Polygon forum, 2026). Alchemy ended support for the chain on 1 July and offers no migration path (Alchemy docs, 2026). L2BEAT still lists Polygon zkEVM with about $9.4M, under category Other, stage "Not applicable" and state validation "None" (L2BEAT, as of 5 November 2026).

What users had to do

Before the sunset, wallet holders had to bridge to Ethereum through Agglayer, and DeFi users had to withdraw from protocol contracts first, because those funds were not migrated automatically (Polygon forum, 2026). Afterwards, the Claims interface covers only assets held in a self-custodied wallet at the sunset. A user connects the wallet, the interface scans for claimable assets, and the user signs a claim transaction on Ethereum; Polygon Labs never takes custody of the funds. Assets in DeFi protocols, smart-contract wallets and third-party bridges are excluded (Polygon Support, 2026). Polygon's terms say recovery may not be practicable after the interface retires.

What it means for the four networks

Shutdowns and retirements now touch three of the four networks. ZKsync Lite stopped on 4 May 2026 with about $33.9M still canonically bridged when the date was set (The Block, 2026). EraVM is being retired, and Scroll plans a private chain. The pattern repeats: months of notice, no action for wallet-held funds, and extra steps for funds inside smart contracts (Matter Labs, 2026). Protocol teams carry the burden too: Polygon told teams with locked value to unwind positions or provide a withdrawal path before the sunset (Polygon forum, 2026). Rollup users should hold assets in a plain wallet when a network announces a change.

Data current as of November 2026.

The networks that remain have tokens, and the tokens tell a plain story.

How Far Have the ZKsync, Starknet, Linea and Scroll Tokens Fallen From Their Highs?

ZK is 99.1% below its all-time high, and SCR, STRK and LINEA are 98.7%, 98.8% and 94.8% below theirs (CoinPaprika API, as of 5 November 2026). Ether sits 51.1% below its peak of $4,946 on 24 August 2025. The four tokens have lost between 95% and 99% of their peak value.

Drawdowns from all-time highs

ZK has a market capitalisation of about $36.8M; its peak came on 13 June 2024, at $1.12 a token. STRK is valued near $335M, against a peak of $3.59 on 20 February 2024. LINEA is valued near $36.7M, and its peak came on 10 September 2025. SCR is valued near $3.8M, against a peak of $1.49 on 21 October 2024 (CoinPaprika API, as of 5 November 2026). No network's technical changes have reversed these moves.

What the tokens do and do not capture

L2BEAT's value secured counts each network's own token: ZK is about $112M of ZKsync Era's $235M, STRK $195M of Starknet's $363M, LINEA $58M of Linea's $342M and SCR $4M of Scroll's $39M (L2BEAT, 5 November 2026). Value secured is not TVL and not ether locked. ZKsync Era's value secured is about six times ZK's market capitalisation, and the two measures differ, so the gap is context rather than a valuation. Linea ties its token to fees through the 80% LINEA burn (Linea, 2025). Starknet accepts ETH and STRK for fees (L2BEAT, 2026), and Linea uses ETH as its gas token (Linea, 2025).

FAQ

?Is a validium the same as a ZK-rollup?

No. A validium also verifies validity proofs on Ethereum, but it keeps transaction data off Ethereum, so users depend on another data layer to rebuild the state ↗. A ZK-rollup publishes that data to Ethereum (ethereum.org, 2026). L2BEAT lists Sophon, a ZK Stack chain that uses Avail for data, as a validium with about $5.6M secured (L2BEAT, 5 November 2026).

?Does a ZK-rollup need a trusted setup?

It depends on the proof system. L2BEAT notes that SNARKs, short for succinct non-interactive arguments of knowledge, require a trusted setup, while STARKs, scalable transparent arguments of knowledge, do not. Starknet uses STARKs only, Linea uses SNARKs, and ZKsync Era and Scroll wrap STARK proofs in SNARKs (L2BEAT, 2026). L2BEAT's Stage 1 rules now bar trusted setups it rates red (L2BEAT forum, 2026).

?Are ZKsync Lite, ZKsync Era and a ZK Stack chain the same network?

No. ZKsync Lite, launched as ZKsync 1.0 in 2020, supported transfers but not smart contracts, and it stopped producing blocks on 4 May 2026 (The Block, 2026). ZKsync Era is the zkEVM chain that L2BEAT tracks. ZK Stack is the toolkit behind other chains such as Abstract and Sophon, and Prividium is the private-chain product built on it (L2BEAT, 2026).

?Will Ethereum itself verify blocks with ZK proofs?

That is the stated direction, not a finished upgrade. Ethereum Foundation engineers outlined optional layer 1 zkEVM clients ↗ in July 2025, with validators checking proofs instead of re-running transactions (The Block, 2025). ethereum.org describes it as research and says current implementations can take minutes to hours per block, while ZKsync said in October 2026 that its Airbender Prime prover proved 99% of 7,200 consecutive blocks in under 10 seconds each on one GPU, a result from its own test (ethereum.org, 2026; ZKsync, 2026).

?Can Solidity contracts run on Starknet?

Not natively. L2BEAT lists Starknet as running the Cairo virtual machine and the Cairo programming language, and it is not an Ethereum Virtual Machine (EVM) chain (L2BEAT, 2026). Developers either write in Cairo or use a compiler. A Solidity-to-Cairo compiler called Warp was under development in 2022 (Buterin, 2022), so check current tooling before porting.

?Can I force a transaction through Ethereum if a ZK network's sequencer censors me?

It differs by network. Scroll lets users self-sequence through Ethereum with up to a seven-day delay. Starknet and ZKsync Era let users submit to an Ethereum queue but cannot force inclusion, and Starknet's Security Council minority can step in. L2BEAT lists no mechanism at all for Linea (L2BEAT, as of 5 November 2026).

?Does ZKsync Era's longer execution delay slow my transfers?

The execution delay is the minimum time before a batch becomes final on Ethereum. Token holders raised it for ZKsync Era from three to 24 hours on 7 October 2026, with the longer window meant to give more time to detect and stop an attack (Matter Labs, 2026). Matter Labs does not say ordinary transfers on Era slow down, but anything waiting for Ethereum finality, such as a withdrawal, waits at least that long.

?Is Polygon zkEVM the same as Polygon PoS?

No. Polygon zkEVM was a separate chain, and its Mainnet Beta sequencer was sunset on 3 July 2026 (Polygon, 2026). Polygon PoS is a proof-of-stake (PoS) chain that L2BEAT classes as Other, with state validation "None" and about $3.9B secured, so it is not a ZK-rollup (L2BEAT, as of 5 November 2026).

References / Sources