Correction — 17 July 2026: I misread Xatu's derived UNSUPPORTED collector label as a response from the execution client. The snooper applies that label to a null or non-array result, while V2 permits null for missing or pruned blobs, syncing, and other inability to serve the complete request. The client-support, version-upgrade and performance conclusions below are not supported by these rows. I also attributed V2 to Pectra; it is specified in the Osaka execution fork paired with Fulu. The full correction reproduces the same calls across both capture paths.

One month after Pectra went live, roughly 44% of all Nethermind validator nodes are rejecting a blob retrieval call that was introduced in that very fork. Every other major execution client has mostly fixed this. Nethermind hasn't moved.

This is the data on what's happening, why it matters, and where the fault line sits.

Most validators proposing a block right now will earn about 0.011 ETH. That's the median. But the mean is 0.050 ETH — more than 4× higher. The reason the mean is so detached from the median tells you everything about how staking rewards actually work.

Over the 30 days ending February 26, 2026, there were 200,963 MEV-Boost blocks on mainnet. I deduped them by taking the highest bid per slot across all relays. Here's what I found.

Each blob on Ethereum costs the same regardless of how much data you actually put inside it. The slot doesn't know if you packed all 131,072 bytes or left 99% empty. You pay either way.

That flat pricing creates a question that hasn't been answered cleanly: how efficiently are rollups actually using the space they're buying? The answer turns out to span an enormous range — from 100% fill all the way down to a rollup that's posting the same completely empty blob, over and over, every block.

Correction, 2026-07-16: I misclassified these withdrawals. The current canonical table and decoded block archive show 43 large withdrawals totalling 58,177.93 ETH, not 52 and 76,322 ETH. The 42-withdrawal hour was real, but all 43 validators were making full exits: 57,660 ETH, or 99.11% of the total, was their original deposit principal. Pectra also activated on May 7, 2025, not January 25, 2026, and the 1.16% credential-adoption estimate used an invalid row denominator. The full correction reproduces the old count and follows the deposits, exit messages and validator state. I have left the original text below unchanged.

Something unusual happened on February 21, 2026. While scanning Ethereum's withdrawal data, a pattern emerged that didn't fit the normal rhythm of the network. 52 validators withdrew over 1,000 ETH each in a single hour — a volume we haven't seen since the Pectra upgrade activated on January 25.

When people talk about the Ethereum Virtual Machine, they reach for the "world computer" metaphor — a globally shared processor executing smart contract code. That framing implies computation: arithmetic, cryptography, logic. In practice, the EVM spends more than half its gas budget on something far more mundane: reading and writing persistent state.

Every week, roughly 1,440 gigagas of EVM execution passes through the mainnet. More than half — 56.7% — goes to exactly two opcodes.

Block builder win rate is the number everyone tracks. Builder X won 18% of blocks last week. Builder Y's market share is up. But win rate hides something: most of those blocks are worth almost nothing. The real competition isn't for volume. It's for the blocks worth 0.05, 0.2, even 1 ETH in builder payments — the slots that account for a disproportionate share of all MEV value.

Looking at seven days of mainnet MEV-Boost data (~45,000 deduplicated slots), two completely different builder ecosystems are visible.

Correction, 2026-08-13: I grouped a per-observer timing table into 200ms buckets, ran count(), and called the result slots. The old histogram therefore counted 7,438,781 observer rows, roughly 148 rows per corrected block, even though only 50,400 slots physically fit in the seven-day window. Rebuilding from exact canonical block roots at one-block-per-slot grain gives 32,694 / 11,502 / 6,006 blocks in Waves 1/2/3. The three waves and the Wave 2 versus Wave 3 accuracy cliff survive; the old y-axis, ~73,000-slot claim, 10,945 Wave 2 count, and 28× penalty ratio do not. Full correction and queries.

When a validator is chosen to propose a block, it has a choice: publish the moment the block is ready, or wait for MEV-Boost bids to arrive and raise the payout. Most discussions frame this as a binary — you either participate in the timing game or you don't.

The data says it's more complicated. There are three distinct groups, and the middle one has mostly gone unnoticed.

The MEV relay system starts building blocks for a slot before that slot exists. Builders are running their engines eight seconds ahead of the clock, constantly revising bids as new transactions hit the mempool. What nobody seems to have charted is what happens at exactly two seconds before a slot starts: the entire auction collapses to near-zero and then rebuilds from scratch.

PeerDAS — EIP-7594's data availability sampling system — has been live on Ethereum mainnet for over 30 days. All 128 column subnets are active, and the data is arriving: 10,956 out of 10,958 slots in the last 48 hours had every single column propagate within 12 seconds. That's 99.98% completeness. The protocol is working.

But there's something nobody seems to have noticed: column index 0 arrives 156 milliseconds faster than column index 101. The correlation between column index and median propagation time is 0.82. And it's been this way, consistently, for seven consecutive days.

Every proposer using MEV-Boost faces the same tradeoff: wait longer to capture more value, but at some point your block arrives too late for attesters to see it before they commit their vote. The timing game is well-understood in theory. What hasn't been measured is exactly where the cliff is — and how steep the drop really is.

The cliff is at 3.0 seconds. What happens after it is sharper than you'd expect.