A blockchain does four things: execute transactions, reach consensus on order, settle disputes about validity, and make the data available so anyone can reconstruct the state. A monolith does all four on every node. A modular stack lets each be somebody's business — and the reason modularity became the dominant design is one boring line item: data availability was eating every rollup's budget.
Why data availability sets the cost
A rollup must publish enough data for anyone to rebuild its state. Without that, a sequencer could lie about balances and nobody could prove otherwise. The requirement is what makes a rollup trust-minimised, and it is also what it pays for on every transaction.
For years that meant posting to Ethereum calldata, which is expensive and permanent — you pay for the bytes forever. That single constraint set the cost floor for the entire rollup category and made the "cheap L2" pitch difficult to deliver outside of quiet periods.
Two changes broke it. Blobs — EIP-4844, live since March 2024 — store rollup data as a separate resource with a bounded retention window, cutting per-byte cost by roughly an order of magnitude in the process. And separate DA layers move the data off Ethereum entirely, at a lower price and a weaker guarantee.
The DA layers
Celestia was built as a dedicated DA network: no execution, just ordering and availability, with sampling-based verification so light clients can check that data was published without downloading all of it. EigenDA uses restaked ETH to buy economic security for the same job, inheriting the slashing conditions of the restaking layer. Avail and nearDA compete on throughput and cost, and Ethereum's own blob roadmap sits in the same design space with a much stronger security inheritance.
The trade everyone is making is legibility for price. Ethereum's DA inherits the base chain's security, which means your worst case is the base chain's worst case. A separate DA layer inherits its own validator set and therefore its own trust assumptions — a smaller set of operators, with their own incentives and their own failure modes.
The question that actually matters
For a rollup choosing a DA layer, the question is not cost per byte. It is: if this DA layer's validators stop publishing, can the chain still be reconstructed, and does the bridge degrade safely?
Three failure behaviours are possible and only one is acceptable. The rollup can halt — annoying, recoverable, honest. It can continue producing blocks whose data nobody can retrieve, which means users cannot exit and the state is unverifiable: this is the catastrophic mode. Or it can fall back to posting to Ethereum, which preserves the guarantee at a higher cost and is the right design if it has been implemented.
Every DA claim should be read as two numbers that are always quoted together and never separately: throughput, and the window during which data must remain retrievable. More throughput with a shorter window trades one failure mode for another rather than removing it.
What modularity buys, and what it costs
Splitting the stack lets a chain choose its own VM, its own fee market, and its own governance without asking anyone — which is why app chains exist and why a team with a specialised workload often prefers one. It also lets specialised layers compete on the thing they do best, which is the real source of the cost improvements.
The costs are as real. Composability fragments: an asset on one rollup is not natively usable on another, so bridges and intents multiply, and each is a trust assumption. Interoperability becomes a product problem rather than a protocol property. And the number of independent components that can fail goes up with the number of layers, which is exactly the trade a monolith avoids by being slow and simple.
The honest summary is that modularity is a bet on specialisation winning over simplicity, and the bet is currently paying off on cost. Whether it pays off on safety depends on how the fallback paths are built — which is a design decision each rollup makes quietly, and the one worth reading the documentation for.
References
Next, read L2 Scaling Solutions 101. More in Infrastructure & Scaling.