The Empty Column
At 03:14 UTC on a Tuesday in early February, the batch submitter for one of the five largest Layer2 networks published its 1,847th consecutive block from the same address. No rotation. No external proposer. No challenge window. Just one ECDSA signature, every two seconds, for eleven straight days.
I found it while measuring something else. I was auditing escape-hatch latency — the wall-clock time a user needs to force a transaction through the L1 contract when a sequencer refuses to include it — and I had a dataframe open with 4.2 million block records. The rotation field was empty. Not null. Empty. No values existed, because the mechanism that would generate them has never run in production on any chain in my sample.
That empty column is the entire story of "decentralized sequencing" as it stands in February 2026. Not a scandal. Not a hack. Just a promise that has lived on a roadmap for twenty-six months while the block producer has not moved an inch from where it stood on day one.
The code whispered what the whitepaper hid.
Why the Sequencer Is the Whole Ballgame
Strip a rollup down and you find three things: a bridge contract on L1, a state transition function, and a sequencer. The first two are verifiable. The third is a node — usually a single one — that decides which transactions enter which block, in what order, and when they get posted to Ethereum as a compressed batch.
Everything a user cares about flows through that node. Fee levels. Inclusion. Ordering. MEV extraction. The difference between your liquidation executing at 41,200 or at 39,800. This matters more than bridge security in a bear market precisely because bridge exploits are loud and rare, while censorship is quiet and daily.
The decentralization pitch has been consistent since 2023. Replace the single sequencer with a permissionless proposer set. Share ordering across chains through a common marketplace — Espresso, Astria, Radius and a half-dozen others were going to be the plumbing. Let anyone propose blocks, let anyone force them, let the L1 contract be the final arbiter.
Ethereum's own researchers added a parallel track in 2024: based sequencing, where L1 proposers themselves order L2 batches, borrowing Ethereum's decentralization instead of building a new consensus layer. It was elegant. It was also slower and more expensive, which is a fatal combination for a product whose only marketing claim is cheap transactions.
Then, in mid-2024, EIP-4844 landed. Blobs cut L2 data costs by roughly an order of magnitude and turned rollups into genuinely profitable businesses overnight. That single upgrade did more to entrench the centralized sequencer than any governance vote could have, because it made the incumbent operator rich. And an operator who is rich has no reason to hand over the keys.
I have mapped this dynamic before. In 2020 I built a Python script that tracked fifteen thousand daily transactions across Uniswap, Compound and Aave to find recursive collateral cascades — the point where composability stops being a feature and becomes a transmission vector. The lesson transfers directly: in any system, the component that captures the revenue is the component that resists decentralization. Sequencers capture the revenue. The decentralization debate is therefore not a technical debate wearing technical clothing. It is a property-rights debate wearing technical clothing.
What I Actually Measured
Methodology first, because numbers without provenance are just adjectives.
I pulled 4,218,447 L2 blocks across ninety days, from the second week of November 2025 to the second week of February 2026, from eleven rollup networks. Five of those eleven account for 94% of the block sample. Data came from public RPC endpoints and block explorers, decoded with standard ABI definitions, normalized to UTC, and cross-checked against L1 batch transactions on Ethereum mainnet.
For each L2 block I extracted four fields: the producing address, the batch submitter address on L1, the force-inclusion contract interaction count, and the timing interval between L1 batch postings.
Four fields. That is all you need to test whether a sequencing layer is decentralized.
Finding One: The Rotation Column Is Empty
Across 4.2 million blocks, on all eleven networks, the number of distinct block-producing addresses was five. One per major chain. Zero on the smaller six, because those chains outsource production entirely to a shared infrastructure provider.
There is no proposer set. There is no auction. There is no rotation schedule. The "sequencer" is an address, and it is the same address it was on the day the chain launched.
I want to be precise here, because this is where most analysis goes soft. A single sequencer is not automatically a failure. It is fast, cheap, and produces clean UX. But the marketing around these chains has, for two years, described a transition that has not occurred. Somewhere between the whitepaper and the block explorer, "we will decentralize" became "we are decentralized," and nobody updated the field.
The most reliable indicator of a protocol's true architecture is not its documentation — it is the entropy of its block producer addresses. Mine came out at essentially zero.
Finding Two: The Escape Hatch Is Real, and Mostly Theatrical
Every major rollup ships a force-inclusion path: a contract on L1 that lets a user bypass the sequencer and get a transaction included directly, with the sequencer given a window to include it anyway before the forced transaction lands.
In theory this is the safety valve. In practice it is a queue.
Median measured latency from force-inclusion submission to actual inclusion across my sample: 18.4 hours. Ninety-fifth percentile: 41.2 hours. Worst case observed: six days and four hours, on a network whose documentation advertises a "24-hour maximum censorship resistance window."
Then there is the cost. The force-inclusion path requires an L1 transaction, not an L2 one. On a chain where the typical user pays fractions of a cent per swap, the escape hatch costs between 4 and 22 dollars depending on L1 gas, plus the L2 gas for eventual execution. During a congestion event, that number triples.
Which produces the following distribution of who can actually exercise censorship resistance: institutions, MEV searchers, and anyone with a reason to move more than ten thousand dollars. Not the retail user the mechanism was designed to protect.
Censorship resistance that prices out 98% of users is not censorship resistance. It is an insurance product with a deductible larger than the policy.
Finding Three: Who Are These Sequencers, Really
Here the data gets interesting, and this is the piece I have not seen published elsewhere.
I took the batch submitter addresses for each chain and traced their funding. Not the chain's treasury — the actual signer set behind the submitter contract. Three of the five major chains in my sample run their submitter through a multisig whose signer set overlaps with at least one other chain's signer set. Two share the same infrastructure vendor. One is a subsidiary of a US-listed exchange.
That last one deserves its own sentence. When the sequencer and the custodian and the exchange order book share a corporate parent, the distinction between "on-chain" and "off-chain" stops being a technical boundary and becomes an accounting convention. The compliance department does not need to read the chain. It already has the keys.
This connects to something I have argued for years and can now demonstrate structurally: KYC on a rollup is theater in both directions. The chain cannot enforce it, because a user can bridge to a permissionless venue in a single transaction. The regulator cannot rely on it, because a well-funded wallet simply never touches the gated front end. What KYC actually does is route honest, small users through an identity funnel while leaving flow-of-funds analysis exactly where it was — dependent on whoever happens to run the block producer. And as the signer-set overlap shows, that is frequently the same institution asking for your passport.
In 2017 I spent four months reverse-engineering roughly fifty thousand lines of contract code from a failed ICO, tracing where 40% of raised funds had locked themselves in unoptimized multisig wallets. The lesson then was that compliance paperwork and custody reality have almost nothing to do with each other. Nine years later, the paperwork has gotten better. The relationship has not changed.
Finding Four: Blob Economics Moved the Profit, Not the Risk
EIP-4844 is the best thing that ever happened to rollup margins and the least discussed disappointment for rollup users.
In my sample, L1 data costs per posted batch fell 92.6% year over year. Sequencer operating margins — estimated as fee revenue minus L1 data and settlement cost — expanded dramatically. User fees, meanwhile, fell only about 31% on the median chain.
Read that gap again. Costs down ninety percent. Prices down thirty. The difference became margin and, in some cases, token buybacks.
I am not making an accusation of bad faith. This is simply what happens when a single operator controls both the supply of blockspace and the schedule on which fees adjust. A competitive proposer market would have compressed that spread, because competition does that. A monopoly keeps it.
And this is the structural point the decentralization debate keeps missing. Sequencers are not failing to decentralize because the technology is hard. They are failing to decentralize because the business model is excellent. You do not auction off a printing press while it is still printing.
Finding Five: The Bear Market Test
Now the part every reader in February 2026 actually wants.
Fee revenue across my sample fell 71% quarter over quarter. Some of that is price. Some is genuine activity decline. But composition matters more than the headline: on the five major chains, revenue fell roughly in line with activity. On the six smaller chains, revenue collapsed faster than activity, because their cost base is fixed in L1 blob space and L2 infrastructure while their fee take is variable.
Translation: the tail of the rollup market is now running at negative gross margin and being subsidized by treasuries. Treasuries denominated in tokens that are down 60% to 80% from their highs.
Four years of ledgers never lie, only distort. The distortion here is that TVL counts bridged assets rather than economic activity, so a chain can look alive while bleeding out. The signal I trust instead is batch posting frequency adjusted for blob price — if a chain slows its L1 postings while claiming unchanged activity, it is optimizing for survival, not for users.
The Institutional Story Is Not Where You Think
I built a real-time tracker last year for spot Bitcoin ETF inflows, and it produced a finding mainstream coverage still has not absorbed: roughly 70% of institutional volume occurs during low-volatility periods. Not during crashes. Not during rallies. During the quiet weeks when nobody is writing headlines.
The same pattern holds across my L2 dataset. Wallets that behave like professionals do not rotate between chains on yield news. They sit in one place, accumulate during calm, and pay almost nothing in fees because they route through private mempools and batch their own transactions. Whale tails flicker in the NFT gallery shadows — but the institutional whale does not swim in a gallery at all. It swims in the quiet.
Which means the decentralization argument that matters most is not the one being made to regulators. It is the one being made to the shrinking set of sophisticated participants who understand exactly where the single point of failure sits and have already priced it into their allocation.
The Contrarian Read: Correlation Is Not Causation, and Decentralization Is Not a Free Option
Here is where I have to argue against my own framing, because the data will not support the easy conclusion.
It is tempting to read the empty rotation column as evidence of capture. It is not. It is evidence that the cost of decentralization is currently higher than the market is willing to pay. Sequence a chain with a permissionless proposer set and you inherit latency, coordination overhead, and — most importantly — a fee market you no longer control. On a chain with a two-second block time and a sub-cent fee, that overhead is not a rounding error. It is the product.
The escape-hatch latency is bad, but it is measured against a hypothetical alternative that has never shipped at scale. Shared sequencer networks have been in testnet for over two years. Based sequencing has no meaningful mainnet deployment. It is easy to indict the incumbent when the challenger is a slide deck.
And the signer-set overlap I flagged is a structural observation, not a smoking gun. A shared signer set between an exchange subsidiary and a compliance-adjacent infrastructure vendor can mean capture. It can also mean that the same three competent engineers are the only ones who can operate the thing — a talent concentration problem, not a governance one. The two look identical on-chain. Only the off-chain context separates them, and off-chain context is precisely what block explorers cannot show you.
What the data actually supports is narrower and more uncomfortable: the industry has spent two years marketing an architectural transition it has not begun, and the market has not punished it because the centralized version works better for everyone who currently matters.
What to Watch Next Week
Stop watching governance forums. Start watching three numbers.
First, batch posting intervals on L1. If a chain stretches them beyond its historical band without a matching activity drop, it is running a treasury, not a network. Second, force-inclusion call counts, which are public on-chain and which almost nobody tracks. A sustained rise there is the earliest available signal of user migration, and it will show up weeks before TVL moves. Third, signer-set overlap between L1 submitters, because that is where the real architecture lives — and it changes without an announcement.
The rotation column will stay empty for a while. The question is who notices the day it does not.
