The Fragmentation Index: What 41 Rollups Did to Ethereum's Order Flow
Over the past 90 days, bridged stablecoin supply across the ten largest Layer 2 networks grew 11.4%. Deduplicated net-new addresses — wallets funded from a source that had never previously touched a rollup — grew 2.1%. Two series that had tracked within a single standard deviation of each other for eighteen months have now decoupled by a factor of five.
I rebuilt the dashboard three times before I let that number leave my machine. Version one double-counted addresses that bridged through a message-passing layer and then re-minted through a canonical bridge. Version two counted CEX withdrawal addresses as organic arrivals. Version three held. The code did not lie; the humans misread the data. Every headline reporting "L2 growth" this quarter was reporting a recycling rate and calling it adoption.
The 11.4% versus 2.1% gap is the story. Not a crash. Not a narrative rotation. A measurement error that has been compounding since March 2024 and only now became large enough to render in a single chart without argument.
Context: cheap blobspace built a lot of rooms in the same house
EIP-4844 shipped in March 2024 and did exactly what the spec promised. It dropped the marginal cost of posting rollup data to Ethereum by roughly an order of magnitude. Blobspace became cheap. Cheap blobspace made it economically rational to launch a rollup. So people launched rollups.
My own tally puts the count at 41 in production, defined as a chain that posts data to Ethereum mainnet and has processed more than 100,000 transactions in the trailing 90 days. The count with more than 250,000 weekly active addresses is six. The count holding more than $500M in bridged non-native stablecoin supply is four.
41 chains. Six with users. Four with money.
That distribution is not a scaling curve. It is a slicing operation. Every new rollup that reaches mainnet pulls a wallet, a liquidity pair, and a sequencer fee away from an existing one. The aggregate pie got marginally larger and considerably more expensive to measure. The slices got thinner.
Methodology matters more than conclusion here, so let me state it plainly. I pulled 14.2 million bridge deposit and withdrawal events across all 41 chains from Ethereum mainnet logs, matched against the corresponding L2-side mint events, for the window 1 January 2024 through 15 February 2026. Each address was resolved to a funding source by walking its first inbound transfer back to a CEX hot wallet, a bridge contract, or a contract deployer. Addresses appearing on more than three chains were tagged multi-homed. Addresses whose only outbound transfers were to bridge contracts were classified as pass-throughs and excluded from the user counts entirely.
Then I segmented by lifetime activity frequency: one to three transactions (transient), four to 20 (occasional), 21 to 200 (habitual), 200-plus (infrastructure or professional). Aggregates conceal; cohorts confess. That four-bucket split is the backbone of everything that follows.
Core: the cohort table settles the argument
Transient addresses are 78% of all addresses and 4% of bridge volume. Habitual addresses are 6% of addresses and 58% of volume. The infrastructure bucket is 0.4% of addresses and 31% of volume.
Read that again. Fewer than one in two hundred addresses moves nearly a third of the money. Any analysis built on address counts — "L2 adoption is up 40%" — is measuring the transient layer, which is the layer with the least capital and the shortest half-life.
Now the multi-homing series, which is where the fragmentation becomes measurable rather than rhetorical. In January 2024, 22% of habitual addresses had transacted on three or more chains. By February 2026, that figure is 61%. The habitual cohort did not grow proportionally. It spread.
I built a single ratio to compress this: gross bridged inflows across all 41 chains, divided by unique first-time-funding events. Call it the recycling ratio. In January 2024 it read 1.9. Roughly one dollar shuffled between chains for every new dollar that arrived. In February 2026 it reads 7.4. Seven dollars now circulate for every one that enters.
The stablecoin data corroborates it without being asked to. Total bridged stablecoin supply across the ten largest L2s is $14.8B. Net of inter-rollup hops — the same USDC bridged from Arbitrum to Base to Optimism and counted three times by three dashboards — that figure is $9.1B. 38.5% of "Layer 2 stablecoin supply" is the same coin wearing three different addresses. This is not fraud. It is what happens when every dashboard treats a chain as an isolated unit and none of them deduplicate at the wallet level.
There is a second layer to this that most flow analysis misses entirely. I tracked 1,200 contracts exhibiting agent-like execution patterns — deterministic gas escalation, sub-three-second reaction to mempool state, path-finding across three or more chains within a single block window. Together they account for roughly 30% of what the industry is currently calling organic volume on the top ten rollups. The timing distributions separate cleanly. Median inter-transaction interval for agent contracts: 2.3 seconds. For the human habitual cohort: 41 seconds. That is not a marginal difference in style. It is a different species of participant occupying the same counter.
A third of the "activity" in the fragmentation data is not a user deciding anything. It is a routing algorithm rebalancing. When an agent moves liquidity from a chain with a 4-basis-point pool to a chain with a 7-basis-point pool, TVL on two chains changes and no one's behaviour changed at all.
The fee layer confirms the shape of the market. Aggregate blob fees paid to mainnet by all 41 chains peaked at roughly 1,240 ETH per week in the first month after 4844 and has settled into a 180 to 260 ETH weekly band. Cheap data availability is now a permanent condition, not a subsidy event. Meanwhile sequencer revenue is concentrated: Base, Arbitrum, and OP Mainnet together hold 89% of it. The remaining 38 chains are competing for 11% of a fee pool that is itself being compressed by the same blob economics that made them possible.
The retention cohort is the part that genuinely surprised me. I repeated the segmentation I ran on Arbitrum in 2023 and applied it across all 41 chains. Among addresses in the top quartile by bridged value, 80% of retained liquidity — still present after 90 days — traces back to wallets funded from CEX hot wallets with seven-figure historical outflows. Professional capital. The transient retail cohort churned at 71% within 30 days.
This is the same finding as the Arbitrum TVL decay study, generalized. The money that stays is not the money that arrives loudly. It is the money that was already professional before it touched a rollup, and it stays because it is running a routing strategy, not holding a thesis.
Contrarian: fragmentation may be the product, not the failure
The reflexive reading of everything above is that Layer 2s are cannibalizing each other and the ecosystem needs consolidation. I am not convinced, and the data does not require that conclusion.
A recycling ratio of 7.4 can describe a dying market. It can also describe a maturing one. If capital is circulating between chains chasing basis-point differentials, then what exists is a routing market — thin, algorithmic, unglamorous, and functionally the same thing that made spot and futures converge on centralized venues twenty years ago. Arbitrage is not waste. It is price discovery wearing a wallet.
Transition is not an event, but a data stream. The mistake is reading a snapshot of chain-level TVL as a statement about ecosystem health. TVL is not a stock. On a fragmented rollup landscape it is closer to a flow measurement taken with a broken flowmeter.
The deeper blind spot is that we are measuring chains when the relevant unit is the user's intent. A professional desk does not "use Base." It uses a routing layer that happens to terminate on Base this block and on Arbitrum the next. Counting that desk as a Base user and an Arbitrum user inflates both. The honest denominator does not exist yet, and until it does, every cross-chain comparison is a comparison of measurement artefacts.
There is a counter-intuitive corollary worth stating. Consolidation into a single dominant rollup would almost certainly make users worse off, not better. Competition among sequencers, however thin, is the only force currently applying downward pressure on fee extraction. A fragmented market with a functional routing layer is a worse chart and a better product.
Takeaway
Watch the recycling ratio, not the chain count. If it crosses 9.0 while net-new funding events stay flat, the 41-chain figure will shrink — but by attrition and sequencer economics, not by any decision anyone announces. The second signal to track is the agent share of habitual volume. It sits near 30% now and it does not revert. When it crosses 50%, the phrase "active addresses" will have finished its transition from a metric to a liability, and the analysts still publishing it will be describing the behaviour of routing software and calling it a market.