Rollup Stages: What Stage 1 and Stage 2 Actually Mean

TL;DR

A rollup's Stage label compresses five separate risk rows into one word, and the compression hides the thing that matters most to your users. On 14 September 2026, L2BEAT tracked 103 scaling projects: 4 at Stage 2, 6 at Stage 1, 73 still at Stage 0, and 20 marked not applicable. The four Stage 2 chains hold under $1M combined — Facet $664K, Cartesi PRT Honeypot v2 $1K, Aztec Network $1K, Ethscriptions at roughly zero. The six Stage 1 chains hold $29.32B. Stage 0 holds $8.48B. So the top label belongs only to chains nobody has funded, and almost all real money sits one rung down or lower. Worse, the Stage 1 badge tells you nothing about the upgrade path. Base is Stage 1 with fraud proofs described as single-round and ZK-proven, and its Exit Window row reads "None" — contracts are instantly upgradable behind two approving parties. Arbitrum One is Stage 1 and its Exit Window row reads "None (emergency upgrade path) / 10d (regular upgrade path)"; OP Mainnet's reads "None". Read the rows, not the badge.

What Stage 1 and Stage 2 Actually Mean on L2BEAT

The Stage framework exists because "is this rollup decentralised" is not a yes or no question. It is a set of separate questions, each with a different failure mode, and L2BEAT answers them individually in its risk row table↗ before rolling them up into a Stage.

The rough shape is this. Stage 0 means the training wheels are fully on: there may be a proof system, or there may not, and a privileged operator can override outcomes. Stage 1 means a working proof system governs state, but a council or multisig retains real power over the contracts. Stage 2 means the proof system is the authority and human intervention is narrowly bounded.

That is the honest summary, and it is where most articles stop. The problem is that a Stage is a summary of rows, and two chains can land on the same summary with materially different rows underneath. The rows are what you actually need. L2BEAT publishes five of them on its risk table:

• State Validation. How the chain proves its state transitions are correct. Fraud proofs, validity proofs, or nothing.
• Sequencer Failure. What you can do if the sequencer stops accepting your transactions.
• Proposer Failure. What you can do if nobody posts state roots.
• Data Availability. Where the transaction data lives. Onchain, or somewhere you have to trust.
• Exit Window. How long you have to withdraw if the contracts are about to change under you.

Upgradeability is not a separate row. Who can change the contracts, and how fast, is what the Exit Window row encodes — which is why that row carries so much weight.

I have deployed on several of these chains and the row I care about most is not State Validation. It is Exit Window, because that is the one that decides whether a user can get out when something goes wrong with governance rather than with the proof system. And that is exactly the row the Stage label buries.

Why the summary label leaks

A Stage is designed to be comparable across chains. Comparability is bought with information loss. A chain can satisfy the Stage 1 conditions on proofs and permissionless exit paths while still being instantly upgradable, and the badge will not tell you. If you are choosing a chain for a protocol that will hold other people's money, the badge is a filter, not an answer. Pull the rows.

The Stage Distribution Nobody Prints: 4, 6, 73

Here is the full distribution from L2BEAT's scaling summary API↗, pulled on 14 September 2026 across 103 tracked projects.

Seventy-one percent of tracked projects are Stage 0. Six projects in the entire ecosystem are Stage 1. Four are Stage 2 and they are empty.

One honesty note before you reuse these figures. The stage buckets do not add up to L2BEAT's headline total value secured of $42.05B on the same date. That is not an error — it reconciles once you read the type field rather than the category. The stage buckets cover all 103 tracked projects and total $48.89B; the $42.05B headline covers only the 89 layer 2s and drops all 14 layer 3s, Hyperliquid's $6.581B among them. Polygon PoS at $3.892B sits inside the headline despite not being a rollup at all, so category is not the dividing line — layer is. Check both fields before you fold a number into a comparison.

The headline TVS breaks down as $13.32B native, $14.67B canonical and $14.06B external, equal to 16,794,166 ETH at an ETH price of $2,504, L2BEAT's intraday reading at the hour of the pull. The daily close on 14 September 2026 was $2,477, which is the figure used for the ETH-denominated comparison later in this piece. Base Chain alone holds $14.679B — 35% of the $42.05B headline. Base plus Arbitrum One ($12.168B) is 64% of that headline and 55% of the $48.8B secured across all 103 tracked projects. Pick one denominator and name it.

Every Stage 2 Chain Is Empty

This is the finding that should reframe the whole conversation. The four Stage 2 chains on 14 September 2026:

Under $1M between them. One of them is literally a honeypot — a deliberate bounty contract built to test whether a proof system holds. Another holds nothing at all.

I am not saying these are bad projects. Cartesi's honeypot is exactly the right way to put a proof system under adversarial pressure, and Aztec is solving a harder problem than throughput. I am saying that the sentence "Stage 2 is next" implies a path that has been walked by chains carrying real load, and it has not. No Stage 2 chain has been stress-tested by billions of dollars trying to leave it at once.

So when a rollup roadmap says Stage 2 in 2026 or 2027, treat it as an engineering intention rather than a demonstrated outcome. The category has no proof-of-load yet.

Base Is Stage 1 and Has No Exit Window

Base is the single largest L2 by value secured, at $14.679B. It reached Stage 1. Here are its risk rows verbatim from L2BEAT on 14 September 2026.

The State Validation row is genuinely strong. "1R, ZK" means single-round, ZK-proven — a dispute resolves in one transaction rather than through a multi-round interactive game. That is a real improvement over the interactive model, though it is not what cleared Stage 1: Arbitrum One and OP Mainnet hold the same badge with interactive proofs.

Now the Exit Window row, quoted exactly:

There is no window for users to exit in case of an unwanted upgrade since contracts are instantly upgradable. Upgrades need to be approved by 2 parties: the Base Coordinator Multisig and the Base Security Council.

Read that twice. The proof system is strong. The upgrade path does not give you time. If the two approving parties agree to change the contracts, the change can take effect immediately, and there is no period during which a user can observe the pending change and withdraw before it lands.

That is not a scandal and it is not a bug report. Two-party approval with a Security Council is a deliberate design that trades user exit time for the ability to patch a live chain quickly. There are real arguments for it, particularly on a chain onboarding users who have never held a private key. But it is a different risk to your users than a chain where an upgrade is visible and delayed, and the Stage 1 badge says nothing about the difference.

If you are building on Base — and I do, it is a good chain to ship consumer products on, which I go through in my Base developer guide — you should know this row and be able to say it out loud to whoever is funding the deployment.

Exit Windows Compared: Base, Arbitrum One, OP Mainnet

Three chains, three Stage 1 badges. The upgrade timelines underneath are not what the badge implies.

The 6 day 8 hour figure quoted for Arbitrum One is its fraud-proof challenge period, not an exit window. Its Exit Window row is two numbers: none at all on the emergency upgrade path, and 10 days on the regular path, where an upgrade clears an 8 day L2 timelock and a 3 day L1 timelock, less the 1 day force-transaction delay. OP Mainnet's Exit Window is None outright, because its Security Council can upgrade instantly and without notice. Base is None as well. So the honest comparison is not zero against ten days. It is three Stage 1 chains that all retain a no-notice upgrade path, one of which also offers a 10 day route when governance is not in a hurry.

That is the comparison the consensus write-up never puts in one table, and it is the one that changes decisions. A user with funds on Arbitrum One has 10 days of visible delay when an upgrade goes through the regular DAO path, and none at all when the emergency path is used. A user with funds on Base or OP Mainnet has none on any path. Same label, and the gap is narrower than the badge debate suggests: all three of the largest Stage 1 chains keep a route to upgrade without giving users time to leave.

Why the exit window is the row that decides user risk

Proof systems fail rarely and loudly. Governance changes happen quietly. Rollup contracts get upgraded routinely, and no production rollup's fraud proof system has been broken in the wild to date. I have not put numbers on either rate and I have no source that has, so read that as a direction rather than a probability. For an ordinary user's expected loss, the upgrade path is the live risk and the proof system is the tail risk.

The exit window is the only row that gives a user agency over the live risk. Everything else is a promise about what the operator will do. An exit window is a promise about what the user can do.

When I evaluate a chain for a client protocol, the order I read the rows in is: Exit Window, Data Availability, Sequencer Failure, State Validation. That is close to the reverse of the order they are usually discussed in. The full chain-by-chain breakdown of the four largest rollups is in my Layer 2 comparison, and the Arbitrum-specific mechanics are in my Arbitrum deep dive.

Stage 0 Is Where a Fifth of the Rollup Money Sits

Seventy-three of 103 tracked projects are Stage 0, holding $8.48B on 14 September 2026 — a fifth of the $42.05B layer 2 headline, and more than five times the value on OP Mainnet.

The Stage 0 list includes chains with serious capital. Robinhood Chain holds $2.740B and is Stage 0. Mantle holds $1.451B. Lighter holds $1.313B. These are not experiments — they are chains with real balances that have not cleared the Stage 1 bar.

The pattern I read from this: reaching Stage 1 is not a natural consequence of getting users. It is a separate engineering programme that a team has to choose to fund, and most teams have not. Six chains have. If your mental model is that rollups mature toward Stage 1 as they grow, the distribution says otherwise. A chain can hold billions and stay at Stage 0 indefinitely, because nothing in its business forces the upgrade.

That is also the honest answer to "should I launch my own rollup". Most rollups that launch stay Stage 0. I go into the economics of that decision — and why it is not a cost argument in 2026 — in the piece on Fusaka, blobs and rollup economics.

Read the Rows, Not the Badge: A Five-Row Checklist

Here is the procedure I run before deploying anything that holds user funds on an L2. It takes about twenty minutes per chain.

1. Pull the risk rows. Open the chain's page on L2BEAT's risk view and copy all five rows and their tooltip descriptions into your own notes. Do not summarise them. Copy the wording.
2. Read Exit Window first. If it says None, write down who the approving parties are and how many of them there are. For Base that is two: the Base Coordinator Multisig and the Base Security Council.
3. Read Sequencer Failure. Base's row says self sequence with up to a 12h delay. That is your users' fallback if the sequencer censors them. Know the number.
4. Read Data Availability. Onchain is the strong answer. Anything else means you are trusting a committee to hand back the data your users need to construct a withdrawal.
5. Read State Validation. Note whether proofs are live, and whether disputes resolve in one round or many. Multi-round interactive games take longer to settle and have a larger surface.
6. Check the archive. Confirm the chain is not on L2BEAT's archived list before you write it into a comparison table.

Then write one sentence for your README: "If governance goes wrong on this chain, our users have X to get out." If you cannot fill in X, that is your answer.

The reason I insist on copying the wording rather than summarising it is that the rows are written precisely and summaries lose the precision. Base, OP Mainnet and Arbitrum One all show the identical Sequencer Failure value, "Self sequence" — and the delay that actually differs, 12h on the first two and 1d on Arbitrum, lives only in the tooltip description. Copy the description too, not just the value. "Fraud proofs (1R, ZK)" and "fraud proofs" are different systems. Every time I have seen a team get a chain's risk profile wrong in an architecture document, the error entered at the summarising step, not at the reading step. Copy the row, paste the row, cite the date you pulled it.

# L2BEAT publishes the same data its site renders.
curl -s https://l2beat.com/api/scaling/summary > summary.json

# Inspect the shape before you write a filter against it.
# Field names here have changed more than once, so do not
# hardcode a path from a blog post — including this one.
jq 'keys' summary.json

That last comment is not false modesty. Any code that parses a third-party dashboard's JSON is a maintenance liability, and I would rather you check the shape than copy a selector that broke last quarter. The same principle applies to measuring anything on-chain yourself, which I work through in detail in the piece on account abstraction measured on-chain.

TVS Is Not TVL, and Mixing Them Breaks Your Comparison

This trips up more comparison articles than any other single error. L2BEAT reports total value secured. DefiLlama reports total value locked. They measure different things and the gap between them is not a fixed ratio. On 14 September 2026 the gap ran 2.6x on Base, 3.8x on OP Mainnet and 8.8x on Arbitrum One — $12.168B TVS against $1.39B DeFi TVL.

TVS counts everything the chain's bridge and contracts secure: bridged assets, natively issued assets, external assets. TVL counts deposits sitting inside DeFi protocols on that chain.

On 14 September 2026, Base was $14.68B TVS and $5.63B DeFi TVL. Same chain, same day. If you put one chain's TVS next to another chain's TVL in a table, the table is wrong.

The DeFi TVL series tells its own story, separate from stages. Pulled from DefiLlama's historical chain TVL↗ on 13 September 2026:

Arbitrum peaked at $4.20B on 7 October 2025. Base peaked at $5.71B on 4 September 2026. Two Stage 1 chains of similar size moving in opposite directions in the same price environment.

One caveat you must apply to every dollar-denominated decline in 2026: ETH is down 46% year on year, from $4,607 on 15 September 2025 to $2,477 on 14 September 2026, and down 14.9% across the 1 January to 13 September window this table covers. Total DeFi TVL fell 22.6% over that same window, so in ETH terms the decline is about 9%: roughly two fifths of the dollar fall is capital leaving and three fifths is price. Denominate in ETH before you call anything a collapse. Arbitrum's -52% against Base's +26% in the same price environment is real divergence, not a price artefact — that is precisely why the comparison is worth making.

The 65 Archived Chains Comparison Sites Still Quote

L2BEAT's archived page↗ carries 65 entries, most of them with residual value still on them. A sample from 14 September 2026:

Loopring has two thousand dollars on it. l2fees.info still lists it in its live fee comparison table, alongside zkSync Lite, DeGate and Boba.

That is the state of L2 comparison content in 2026: a widely linked tracker presenting a chain with $2,000 of value as a live option, ranked second-cheapest, alongside Arbitrum One and Optimism — while Base, the largest L2 by value secured, does not appear in the table at all. If your chain selection process starts with a comparison table you found through search, check every row against L2BEAT's archived list before you trust any of it.

Fees Are Not the Differentiator Any More

There is a second reason the stage rows matter more in 2026 than they did two years ago: the fee argument has collapsed into noise.

I read eth_gasPrice from each chain's public RPC endpoint on 14 September 2026 at 14:58 UTC and multiplied by 21,000 at an ETH price of $2,504. These are single point samples taken from one machine and are not reproducible from any public archive, so treat the ordering as the signal and the digits as indicative:

Two caveats, both important. First, these are execution gas only. OP-stack chains add an L1 data fee per transaction which is not in this table; I was rate-limited before I could get a clean all-in median for Base and Arbitrum, and the one OP Mainnet reading I did get came to $0.001115 all in. Second, these are point samples, not daily averages.

For context on the same date, growthepie's transaction cost leaderboard↗ had Arbitrum Nova at $0.06 as its top chain — a different methodology producing a much higher number, which is exactly why you should measure your own path rather than quote a leaderboard.

When the spread between the cheapest and the most expensive major rollup is roughly two cents on a simple transfer, fees stop being a decision input for most applications. What remains as a real differentiator is exactly the thing in the risk rows: who can change the contracts, how fast, and what your users can do about it.

What I Actually Do With This When Choosing a Chain

Concretely, for a protocol holding user funds, this is the decision I make and the tradeoff I state to the client.

If the application is consumer-facing, high-frequency, and the balances per user are small, Base's risk profile is acceptable. The Exit Window of None is a real cost, but the amount at stake per user is small and the chain's distribution advantage is large — $14.679B secured and 35% of all L2 value is not an accident. I say the tradeoff out loud in the architecture document rather than pretending it is not there.

If the application holds large balances per address, or institutional money, or anything where a governance failure produces a headline, Arbitrum One's 10 day regular-path exit window is worth the extra $0.0007 per transfer. That is not a close call on fees. It is a closer call on governance than it looks, because the same row records no exit window at all on the emergency path — the 10 days is what you get when the DAO is not calling an emergency, not a floor you can rely on.

If someone proposes launching a new rollup, the first question is not throughput. It is whether the team will fund the multi-year programme that gets a chain from Stage 0 to Stage 1, because 73 of 103 tracked chains did not. If the answer is no, you are building a Stage 0 chain, and you should say so in the documentation on day one rather than in an incident post-mortem.

The sequencer is the other half of this picture, and centralisation there is moving the same direction as block building on L1 — I cover the measured state of that in the piece on MEV and restaking in 2026.

Key Takeaways

• The Stage 2 category is empty. On 14 September 2026 all four Stage 2 chains held under $1M combined — Facet $664K, Cartesi PRT Honeypot v2 $1K, Aztec Network $1K, Ethscriptions roughly zero. No Stage 2 chain has been tested by real load.
• Only six chains are Stage 1 at all, holding $29.32B: Base $14.679B, Arbitrum One $12.168B, OP Mainnet $1.662B, Starknet $0.388B, Ink $0.336B, Unichain $0.085B.
• Seventy-three of 103 tracked projects are still Stage 0, holding $8.48B — including Robinhood Chain at $2.740B. Growth does not produce Stage 1; a funded engineering programme does.
• Base's Exit Window row reads "None." Contracts are instantly upgradable with approval from two parties, the Base Coordinator Multisig and the Base Security Council. Arbitrum One's Exit Window row reads "None (emergency upgrade path) / 10d (regular upgrade path)" and OP Mainnet's reads "None". The 6d8h quoted for Arbitrum One and the 3d12h quoted for OP Mainnet are fraud-proof challenge periods, not exit windows.
• TVS and TVL differ by anywhere from 2.6x to 8.8x on the same chain. Base was $14.68B TVS against $5.63B DeFi TVL on 14 September 2026; Arbitrum One was $12.168B against $1.39B. Never mix them in one table, and never assume a fixed ratio between them.
• Fees no longer separate the majors. Execution cost for a 21,000-gas transfer ran from $0.00001 on Scroll to $0.02104 on Linea on 14 September 2026, execution gas only, excluding L1 data fees.
• Check the archive before you quote a chain. L2BEAT lists 65 archived projects, and l2fees.info still shows Loopring — $2K of residual value — in its live comparison table.

About the Author

I'm Uvin Vindula — a Web3 and AI engineer based between Sri Lanka and the UK. I deploy production contracts on Base and Arbitrum, and the risk rows in this article are the ones I read out loud to clients before a mainnet deployment rather than after one. You can see my work at iamuvin.com or reach out about a project at hello@iamuvin.com↗.

If you're choosing an L2 for a protocol that will hold other people's money and want the risk rows read properly before you commit, let's talk about your project.