Over the past 90 days, the median rollup on my tracking sheet consumed 0.6% of the data availability bandwidth it pays for. Not sixty percent. Zero-point-six. I ran the query because a counterparty wanted me to structure a hedge against rising DA costs, and the resulting series was unusable โ you cannot hedge a line item that barely registers against the P&L it supposedly threatens. Ethereum's blob fee market cleared at its floor price of 1 wei in 71% of blocks during that window. When a fee market sits at its floor that often, it is not a market. It is a parking lot with a toll booth nobody drives through.
That is the anomaly I want to unpack, because it sits underneath a trade that a lot of desks are still carrying.
The context matters here, and it is mostly mechanical. EIP-4844 introduced a separate fee market for blobs in March 2024. Blobs are cheap by design โ 128 kilobytes each, with a target of three per block and a maximum of six. Crucially, the blob base fee adjusts independently of execution gas, using its own exponential controller. When demand drops below target, the fee decays fast. It does not decay to a comfortable equilibrium. It decays to the minimum, 1 wei, and it stays there until demand returns. That design choice solved the rollup cost problem so thoroughly that it destroyed the business model of anyone selling the same solution as a product.
Meanwhile, the DA supply side has done the opposite of consolidating. Celestia shipped its own blobspace with data availability sampling. EigenDA, Avail, and a handful of alt-DA layers launched into the same window, each with a token, each with an incentive program paying out emissions to attract rollup teams that were already posting to Ethereum for effectively free. The result is a supply glut in a category whose demand curve is flat and whose marginal buyer is price-insensitive because the price is already zero.
I have seen this pattern before. In 2017, I manually audited three ERC-20 contracts in Remix and found integer overflow vulnerabilities in two of them before public launch. The teams had raised money on the narrative that their token was the product. The code said otherwise โ the token was a receipt, and the receipt was for a claim nobody had priced. A token that prices a cost line trending toward zero is not a growth asset. It is a short with a delay fuse.
Now to the numbers, because the ledger remembers what the ego forgets.
Decompose a rollup's marginal cost per transaction and the DA share collapses. On a typical OP Stack chain, the cost stack breaks into four components: L1 data posting, proof generation and verification, state growth and archive maintenance, and sequencer overhead. In 2023, data posting was the dominant term โ 70% or more of marginal cost for most chains. Today, on my own measurements across four chains, data posting runs between 8% and 22% of marginal cost depending on batch cadence and compression ratio. Compression is the quiet variable here. Teams that invested in better state-diff compression โ zlib to brotli to custom dictionary encoding โ cut their data footprint by 40% to 60% without touching which DA layer they used. Those teams are now indifferent to DA pricing entirely.
The cost moved. It moved into proving and state. ZK rollups pay for proof generation, and that cost is roughly linear in circuit complexity, not in bytes posted. Optimistic rollups pay for the seven-day challenge window in capital lockup, not in DA fees. Both of them pay for state growth, which is the cost that never appears in a DA pitch deck because it cannot be externalized to a third party. Every state leaf you write is a permanent obligation on the full node set.
Here is where the DA thesis and the actual cost structure diverge: the pitch is that rollups scale and therefore need more DA. The reality is that rollups that scale do so by compressing harder and proving cheaper, both of which reduce DA dependence. Scaling and DA consumption are negatively correlated once you get past the low-hanging fruit.
I watched this exact dynamic in 2021 during the Azuki launch. I spent roughly $2,000 in gas to save $15,000 in slippage, and I mapped the fee spikes block by block. The lesson was not that gas is expensive. The lesson was that the fee market is a queue, and queues clear. Every congestion event I have traded since โ NFT mints, airdrop claims, liquidations โ resolved within a predictable window. Alpha hides in the friction of chaos, but the friction is transient, and the alpha decays with it. DA congestion is the same shape. There is no persistent rent to extract from a queue that clears at 1 wei for three quarters of the year.
So where does the value actually accrue? Not in storage. Storage is a commodity with near-infinite supply and a buyer base of maybe forty teams. The value accrues to the parties who control sequencing and the parties who control verification. Shared sequencer auctions extract real fees because they price ordering, and ordering is scarce. Provers extract real fees because verification is a fixed obligation that scales with chain count. DA extracts nothing because the marginal cost of an additional blob is approximately the cost of a few dozen bytes in a block that was going to be produced anyway.
I ran a similar decomposition in 2022 when I backtested TerraUSD's peg mechanism against historical volatility. The flaw was visible three days before the crash โ not in the code's logic, but in the liquidity pool imbalance that the logic assumed away. The lesson generalizes. Code does not lie, but it does obfuscate, and the obfuscation usually lives in an assumption about demand that nobody stress-tested. The DA assumption is that rollup data demand grows superlinearly with rollup adoption. It does not. It grows with rollup transaction count times bytes-per-transaction, and bytes-per-transaction is falling every quarter.
Now the contrarian read, which is where most people are going to disagree with me.
The retail position in DA tokens is a bet on a narrative โ modularity, the separation of concerns, the idea that data availability is the layer where value concentrates. The smart money position is different and it is not a bet on DA at all. It is a bet on the sequencer margin. Look at where the desks with actual size are positioned: they are long the chains that own their own sequencing, because those chains capture ordering revenue, MEV, and the spread between what users pay and what the chain pays for settlement. That spread is the business. DA is a line item on the cost side, and it is shrinking.
Silence in the order book is louder than noise. I pulled the order book depth on two DA tokens and one sequencing-adjacent asset over the same 30-day window. The DA tokens show thin depth, wide spreads, and a persistent sell-side overhang from emissions unlocks. The sequencing asset shows tighter depth and passive bid stacking at lower levels. That is not a sentiment reading. That is positioning.
There is a second blind spot. Most people evaluating DA tokens look at total blobspace sold and assume that equals revenue. It does not. A large share of blobspace on alt-DA layers is being purchased with the DA layer's own token via incentive programs โ paid in the asset, denominated in the asset, and recycled back into the market through validator sales. That is not revenue. That is a customer acquisition cost financed by dilution, and it shows up as a liability the moment the emissions taper. I built the same tracking discipline in 2024 when I wired a dashboard to the Grayscale GBTC and BlackRock IBIT wallets and found a $50 million accumulation pattern ahead of the Q4 rally. The signal was in the flow direction and the wallet clustering, not in the headline numbers. The same discipline applies here: track who pays, in what asset, and where that asset goes next.
What would actually change my mind? Two things. First, a sustained rise in blobs posted per block above the target of three across a full month with compression ratios held constant. That would mean demand is real and structural. Second, a material increase in proving costs that makes DA the relatively cheaper path โ which is unlikely, because proving costs are falling faster than DA costs. Neither condition is close to being met.
So here is the forward-looking frame. The trade is not long DA and it is not short DA. The trade is recognizing that the DA layer answered a question that the market stopped asking, and that the capital currently sitting in that category is mispriced against a compressing cost line. Watch the blob base fee. Watch the blobs-per-block count. Watch the emissions schedule on any alt-DA token.
Three data points, all public, all visible before the quarterly report. The question is whether you are still pricing the narrative, or whether you have started pricing the ledger.