EIP-7915 Adaptive Mean-Reverting Blob Pricing (Adaptive mean reversion blob pricing)

No call decision · Fees & gas · EIP status Stagnant · created 2025-03-23

Changes the fee update mechanism so that when the blob fee is below the long-run average, the fee rises quickly if demand increases and falls slowly if demand decreases. The long-run average is stored in the header as an exponential moving average.

Why it is needed

Blob demand has a property where it does not shrink or grow much even when the fee changes (fee inelasticity). As a result, even a small shift in demand causes large swings in the blob base fee. If blob usage falls short of the target for an extended period, the fee drops to the minimum of 1 wei, and afterward, once demand recovers, nearly full blocks must continue for a long time before the equilibrium price is restored. In the meantime, users experience bidding competition under a first-price auction scheme, where they pay exactly what they bid to get into a block, which degrades usability and causes resource consumption to fluctuate unevenly, becoming an obstacle to throughput scaling. In addition, the appropriate fee range keeps changing depending on the ETH price, circulating supply, and protocol throughput scaling, so a fixed floor may become inappropriate over time. This EIP therefore uses the long-run average fee as a reference point and asymmetrically adjusts the update speed the further the fee falls below the average, so that the fee returns quickly to a desirable equilibrium even as demand changes.

What changes

Today

After EIP-7915

Who is affected

In depth

Calculating the Mean Reversion Weight w

The weight is defined as w = ((f̄g/fg + D)/(D+1))^(1/8), where D is MEAN_REVERSION_DEADBAND = 64. In the implementation, if the current fee is higher than the EMA, w is set to the upsampling constant Up = 2**16, effectively making it 1; if lower, w is calculated as the ratio of the EMA to the current fee. If the ratio exceeds MEAN_REVERSION_FULLBAND = 10**6, it is clamped at that value to prevent the fee schedule from running away and overflowing, and to allow internal computation to be handled as uint64. The 8th root is obtained by applying integer_squareroot() three times; since this function currently exists only in the consensus layer, it must be newly implemented in the execution layer.

Change to the Excess Blob Gas Update Formula

Changes EIP-4844's update formula e(n) = e(n-1) + c - t to e(n) = e(n-1) + c·w - t/w. In code, this is parent.excess_blob_gas + parent.blob_gas_used * w // Up - TARGET_BLOB_GAS_PER_BLOCK * Up // w. When w is greater than 1, actual usage is treated as inflated and the target is treated as reduced, so the lower the fee is below the average, the more the fee is maintained even with fewer blobs, while with more blobs the fee rises quickly. The Rationale also notes an optional approach of adding a term c(w-1), equal to a constant c worth of one to two blobs, to further constrain the operating region.

EMA Calculation of the Long-Run Average Fee and the Header Extension

The long-run average is updated every slot by a new function calc_base_fee_per_blob_gas_ema() as f̄(n) = f̄(n-1) + (f(n-1) - f̄(n-1))/m, where m is EMA_DECAY = 2**18. The half-life of this setting is about one month. The value is stored in the header as a uint field base_fee_per_blob_gas_ema, and in the first block after the fork, the parent value is treated as 4*10**9 (4 gwei). An assert is added to execution layer block validity checks to confirm that this field matches calc_base_fee_per_blob_gas_ema(block.parent.header). The source also gives an example in which raising EMA_DECAY to 2**20 (a half-life of about four months) guarantees f̄g > 5*10**8 even one year after activation.

Why Mean Reversion Applies Only in the Downward Direction

Mean reversion operates only when fg < f̄g. Bidirectional mean reversion might also have benefits, but the case where the fee is below the average matters far more, and the EIP-7691 update schedule used above the average already has a larger maximum downward step at zero blobs than its maximum upward step. A symmetric schedule near a ratio of 1 was also considered but not adopted, because updates for the same number of blobs above and below target would not be reciprocals of each other, causing equilibrium to form at throughput slightly below TARGET_BLOB_GAS_PER_BLOCK and price slightly above it. Instead, as the ratio decreases, there is a smooth transition to a symmetric schedule in the intermediate range. Because the change in fee with respect to the number of blobs remains smooth across all ratios, neither users nor validators have an incentive to strategically insert a particular number of blobs.

Linear-Domain EMA vs. the Log-Domain Alternative

The EMA is computed in the linear domain to preserve additivity and interpretability. Averaging in excess gas (the log domain) would let minor fluctuations in the low-fee region have excessive influence, creating feedback that is hard to explain. However, the source also describes an alternative in which the EMA computation itself is done in the linear domain but stored in the log-domain representation excess_blob_gas_ema. In this case, get_base_fee_per_blob_gas() and its inverse (requiring a new fake_log() function) are used to move between domains, and the weight is computed in the log domain in the form w = (ēg - eg + D)/(D+1). The source explains that the fee schedules of the two approaches are very similar, so the choice is mainly a matter of architectural preference.

Security considerations

When the current fee is dozens or hundreds of times below the average and blocks are full, the blob base fee doubles with every block. As a result, a one-block delay in inclusion can cause a user to pay twice as much, which could affect inclusion strategy and fee market strategy. However, the source assesses that the practical impact will be limited, since this aggressive fee schedule only operates in the region where the fee is far below the average.

Current status

This EIP was proposed on March 23, 2025, and its current status is Stagnant. It is classified as a proposal where discussion has stalled, so its content may change or discussion may resume later, and no decision to include it in a fork has been made at a core developer call so far.

Original EIP · ← EIPs