How L2s Actually Work
Layer 2 rollups reduce Ethereum transaction costs through two mechanisms: off-chain execution and compressed on-chain settlement.
The execution itself—token swaps, smart contract calls—happens on L2 infrastructure. This is cheap because it doesn't require global consensus across thousands of nodes. The L2 then posts a compressed summary of hundreds of transactions to Ethereum L1 as a single batch.
The fee structure reflects this split:
- L2 execution: Fractions of a cent, paid to L2 operators
- L1 settlement: The expensive part, but amortized across 500+ transactions in each batch
A basic Ethereum transfer costs 21,000 gas units. At 50 gwei, that's ~0.00105 ETH ($3.15 at $3,000 ETH). Layer 2s reduce this to near-zero by dividing L1 costs across batched transactions.
The Blob Storage Shift
EIP-4844, implemented in 2024, introduced "blob" storage—temporary data attachments to Ethereum blocks, deleted after 18 days. Before this, L2s stored transaction data in permanent calldata, which was expensive.
Blobs cut L2 fees by 90% or more because Ethereum charges significantly less for temporary storage. Transactions that previously cost $0.10 now run at $0.01.
Production Considerations
The L1 settlement cost still dominates L2 fees. When Ethereum gas prices spike, L2 fees rise proportionally—though the batching still provides 10-100x savings versus direct L1 transactions.
Most L2s currently use centralized sequencers to order transactions. If that server fails, the network pauses. This is the primary operational risk versus L1's distributed consensus.
For enterprise developers: fee estimation tools exist for Arbitrum Nitro, Optimism, and Base. The compression ratios vary by implementation—Arbitrum Nitro's calldata compression differs from Optimism's fraud proof structure.
What This Means
L2s treat Ethereum as a data availability layer, not an execution environment. The architecture works because they inherit L1 security while moving computation off-chain. The question for CTOs: whether centralized sequencer risk is acceptable for your use case.