A provider outage costs AI-assisted searcher teams release days; on this evidence, it rarely costs them fills.
Where Wu puts AI in the trading process
Wu asks where MEV (maximal extractable value) bots depend on AI services: during development or during trading. Timing separates the possibilities. Ethereum produces a block every twelve seconds, Base every two, and builders arrange transactions in milliseconds. A cloud language-model call takes seconds, too long for a winning bot's execution loop. Assistants can help operators write and fork-test new executor contracts, the private contracts that later carry out verified arbitrages, sandwiches or liquidations. They can also support background monitoring and parameter updates that bots apply locally. Dependence during development would mean fewer new executors during outages, with existing bots continuing to trade.
Wu tests this with AIDA, a registry of AI service outages assembled from 8,189 raw status, postmortem, archive and forum reports. These resolve into 5,868 incidents at 30 providers. Seventeen providers are foundation-model vendors; they account for 709 developer-relevant disruptions lasting 60 minutes to 24 hours, or 2,586.5 exposure hours. Primary sources confirm 94.8% of their incidents. The Ethereum ledger covers 5 September 2024 to 12 October 2026, supplying 5,041,034 verified strategy transactions from 8,516 addresses. The baseline includes 8,997 addresses: verified MEV senders and the separate keys used to deploy their executors. On Base, the additional panel contains 1,099,593,191 pool swaps and 29,120 private executors.
After an eight-hour Anthropic or OpenAI outage, the number of addresses releasing a new executor drops to 0.67 [0.49, 0.92] of normal on day zero. On day two, it is 0.66 [0.51, 0.86]. Verified arbitrage during all-vendor outage hours runs at 0.99 [0.96, 1.02]. These are activity measures. We found no profit measure beyond the 21.07 basis point gross surplus in one illustrated cycle.
Exact strategy rules, down to the integer
Wu defines an arbitrage as a closed cycle with at least two swaps across at least two pools. Hops must match exactly in integers and follow log order; the closing hop must return at least the initial input. For a sandwich, the back-run input must be within 90 to 110 percent of the front-run output. Liquidations require decoded Aave V2, Aave V3 and Compound V2 events with strictly positive amounts.
For deployments, Wu uses a daily Poisson pseudo-maximum-likelihood (PPML) count regression, relating counts to outage hours at lags zero to three. It includes day-of-week and month fixed effects, with week-clustered errors and a 28-day block bootstrap. Wu also audits the nine-signature swap filter against one unfiltered day. Entirely missed transactions amount to 1.4% of the selected count. Another 4.1% of selected transactions include a leg on an unlisted pool that the decoder cannot read.
Where would two implementers disagree?
The choice of outages comes first. Anthropic and OpenAI supply the daily panel. In the hourly Ethereum model, executor creations during OpenAI hours are 0.68 [0.48, 0.94] of normal. Pool all vendors and the ratio becomes 0.96 [0.67, 1.38]. Base operator-signed creations show a similar split: 0.76 [0.59, 0.99] in OpenAI hours, against 0.94 [0.82, 1.09] across all vendors. Pooling removes the hourly effect. Base top-level creations also decline during GitHub Copilot disruptions, to 0.73 [0.56, 0.96]. The effect appears for named tools and washes out across the 17-vendor pool.
The counting unit changes the answer too. Uncapped creations fall to 0.78 [0.62, 0.98], a decline Wu traces to a handful of addresses stopping bulk non-trading batches. Cap creations at three per address-day and the estimate becomes 0.97 [0.89, 1.06]. On day zero, all persistent contracts from the same operators run at 0.94 [0.84, 1.05]; the executor contrast against them is 0.72 [0.51, 1.00], p=0.048. Changing the operator definition has less effect. Single-caller bots yield 0.62, private bots 0.65 and all DEX deployers 0.64. Using direct sender wallets alone gives 0.73 [0.53, 1.01]. Its block-bootstrap interval, [0.46, 0.99], excludes one.
Later behaviour determines a contract's role, leaving late contracts less time to qualify as executors. Wu compares 2026 with 2025 and estimates 0.95 [0.53, 1.69]. From 0.53 to 1.69, the interval leaves room for a sizable difference.
The inference method also matters. Over four-day windows, the cumulative effect is -15.3% [-30.6, -0.1] with week clusters; the block-bootstrap interval is [-27.0, 3.5]. The day-three rebound reaches 1.23, with an interval of [0.90, 1.69] that includes one (p=0.197). The dip-and-catch-up account therefore depends on the day-zero and day-two declines. On Base, the affected margin changes with the design. Executor-deploying operators fall to 0.88 [0.78, 0.98] at lag two, yet run at 0.98 [0.84, 1.14] during OpenAI hours. Top-level creations go the other way, reaching 1.04 [0.91, 1.19] at lag two.
We would bet on the day-zero Ethereum dip replicating.
Live execution holds up, with two downward exceptions
Across the 156 cloud-screened nearby-day episodes, arbitrage runs at 0.95 [0.88, 1.04], and sandwiches at 0.98 [0.92, 1.03]. Compare the same weekdays and arbitrage is above normal, at 1.12 [1.03, 1.22]. Constructor-executed trades are single-use contracts that trade during their own deployment. In the daily panel, they reach 1.11 [1.00, 1.23] on disruption day zero.
Base arbitrage executor transactions supply the first downward exception: 0.95 [0.91, 0.99] during OpenAI hours. The decline is 5%, and its interval excludes one. The summary claim of uninterrupted execution depends on the all-vendor estimate of 1.00.
Liquidations are the second exception. They fall to 0.25 [0.11, 0.61] in the cloud-screened baseline (156 episodes), versus 0.51 [0.18, 1.51] in the extended sample (198). Wu writes "we report the shortfall as unresolved" and offers opportunity lumpiness, uncaptured infrastructure overlap and cohort aging as possible explanations. Smaller liquidators' share changes by +4.22 points [-5.10, 14.32], so the losses show no tilt toward incumbents.
More arbitrage addresses, yet 83% of trades stay with the top decile
Active arbitrage addresses increased from 914 in 2024Q4 to 3,749 in 2026Q3. Although the top-50 share dropped from 80.32% to 50.25%, the top decile still executes 82.79% of trades. Sandwich addresses declined from 207 to 133, and their top-50 share was 98.14%. From 2025Q1 to 2026Q3, deploying addresses increased from 69 to 241 per quarter. Creations per deployer moved the other way, from 44.46 to 7.57.
Wu acknowledges that the chain records no vendor usage. The cross-strategy pattern is therefore indirect evidence; Wu presents the outage design as the direct test. That test shows developers depend on assistants. A one-third release decline on outage days fits the idea that assistants helped drive the fourfold rise in addresses, without isolating their contribution. Wu offers Uniswap V4 growth as a competing explanation, then argues that venue growth would not account for the sandwich exit. That reply addresses sandwiches. Whether arbitrage entry follows lower coding costs or new pools remains open.
Why we did not rerun the outage tests
Our data consists of price bars. It contains no contract-creation traces, internal factory creations or operator attribution, the dependent variables needed here. Anyone attempting a replication should begin with Wu's released AIDA dataset and replication package, which reconstruct the estimates from transaction-level records.
Desks should arrange a vendor fallback for executor releases. Arbitrage and sandwich fills showed no outage dip; the liquidation shortfall remains unexplained.