The Money Leaves. You Do Not Get To Know When It Lands.

Somebody sends you a screenshot at four in the afternoon. The place holding the company's dollars pays 3.9%. The place in the screenshot pays 4.4%. You run it against the four million you are sitting on, get twenty thousand dollars a year, and decide to move.

That is the decision everyone argues about. The part nobody argues about is that moving it is two separate instructions, sent one after the other. You take the money out of the first place, then you put it into the second. In between, it is in neither. It earns nothing, and if anything interrupts you after the first instruction and before the second, what you have is a pile of dollars in a wallet and a plan you were halfway through.

Everybody who has done it knows that gap exists. Nobody has published how long it lasts, so we measured it.

Ninety days of Ethereum, ending this week. Seven of the places companies actually park dollars: Aave, Spark, Compound's two stablecoin markets, Morpho, and the two savings vaults behind Sky's dollar. Every withdrawal of ten thousand dollars or more, paired to the next deposit of at least eighty percent of it somewhere else, by the same account, inside a week.

Nine hundred and thirty-two of those moves were made straight from somebody's own keys, across a hundred and seventy-eight wallets. Not one of them landed in a single transaction. Not one among the eighty-eight wallets that did it exactly once in three months, and not one among the four that did it dozens of times each.

Two hundred and seventy-four moves in the wider population did land in a single transaction, and every one of them came from a contract. Twelve of them, and we can name them: Sky's own allocator moving its own treasury, IPOR's two automated dollar optimizers, a trade settlement contract, a couple of vaults, a handful of unnamed proxies. Software, moving money that somebody else put under its control.

If you only move money a couple of times a quarter, the typical move is quick. The middle one takes fourteen and a half minutes, and a quarter of them are done inside three.

Then it splits. Six in ten finish inside the hour. The rest take a median of two and a third days, and one in ten is still going after four and a half. Nothing in the data separates the two groups: same venues, same sizes of money, same kind of wallet. A fifty thousand dollar hop from Aave to Compound looks identical, at the moment you send the first instruction, whether it is going to finish before you get a coffee or on Thursday.

The middle move in that group is a hundred thousand dollars. At four percent a year, fourteen and a half minutes of it sitting in neither place costs eleven cents. Thirty-two hours on the same hundred thousand is fifteen dollars, and the longest gap in the whole ninety days, seven full days, is seventy-seven. Against the twenty thousand a year that started the conversation, none of it registers.

What the company loses in the gap is a position. For those fourteen minutes, or those two and a half days, it has an instruction that went out and an instruction that has not. The number in the treasury report is wrong. Whoever countersigns the second half has to still be at their desk. If the destination fills up, or the rate you moved for moves back, the money is sitting in a wallet with nowhere it was meant to be. And you cannot tell your board the move takes fifteen minutes, because a quarter of the time it takes more than a day.

They exist and they work. DeFi Saver's Loan Shifter moves a position between Aave, Compound and other venues in one go, and its page says so plainly: "All transactions in Loan Shifter are happening atomically and are powered by flashloans that enable this process to happen in 1-tx and that your funds are never at risk during the transfer. Either the entire shift succeeds, or nothing happens." A flash loan is money borrowed and paid back inside the same transaction, which is what lets the whole move either happen or not happen at all. Morpho ships a migration flow that brings a position over from elsewhere. Any custodian will do it on the phone.

DeFi Saver's page quotes a quarter of a percent, which on that hundred thousand dollar move is two hundred and fifty dollars. That is more than the interest lost at every gap length we measured, including the seven-day one. A company paying it is paying to always have a position, because the interest alone would never justify it.

Every one of these tools makes the move land whole the same way: the money goes through a contract somebody else controls, along a route built in advance. So it is only whole if the pair you want is a pair they built, and the wallets we measured moved between seven venues in twenty-one different directions. And it needs a standing permission over your account. Of 703 wallets running no DeFi Saver automation at all, 695 still gave its executor the right to run code as them, over eighty million dollars of assets, for a median of 512 days after the automation was switched off. The twelve accounts that moved money in one transaction are all built that way: every one is a contract holding standing power over somebody's funds.

Plug is programmable Ethereum transactions without smart contracts. The move becomes one program you sign once: take the dollars out of the first venue, read what the destination is actually paying and whether it has room at the moment it runs, and put the dollars in. It lands whole or it does not happen, the same property DeFi Saver sells, and afterwards nothing is standing, because nothing was deployed and nobody was given the keys.

Every step is a real action in the catalog today: withdraw and supply on Aave, Spark, Compound and Morpho, deposit and withdraw on the savings vaults, a swap when the dollar at each end is a different dollar, and the reads that let the destination be worked out instead of typed in.

Of the 932 moves made from somebody's own keys, a quarter outlasted the working day they started in, and the interest lost on the typical one was eleven cents. The rate difference decides whether to move. How long the company goes without a position is decided by how the money travels.

Ethereum mainnet, blocks 25,338,879 to 25,986,879, which is 648,001 blocks, about ninety days, ending 15 September 2026. Withdrawals are only counted up to block 25,936,479, so every one has a full seven days to be matched; a gap longer than seven days is not counted as a move at all. Seven venues: Aave v3's main pool, Spark, Compound v3's USDC and USDT markets, Morpho Blue, sDAI and sUSDS. We read all 4,536,007 block-ranges we set out to read, with nothing left unread and nothing unreadable.

Only stablecoin reserves count: USDC, USDT, DAI and USDS, each one's own decimal precision read live off the token, all priced at a flat dollar. Withdrawals below ten thousand dollars are dropped, because the raw stream is mostly dust and sweeps that would own the median: 183,092 of them, worth $254m together. A further 5,533 withdrawals sent the money to a different address than the one withdrawing it and are excluded too.

That leaves 93,605 eligible withdrawals, of which 7,695 were matched to a deposit elsewhere. The other 85,910 did not arrive at another venue we watched inside seven days. That is not abandonment: the money may have gone to an exchange, paid a bill, or bridged to another chain. A pair is a withdrawal followed by a large enough deposit elsewhere by the same account, so some of what we counted as one move was really two decisions a day apart; the claim is that the money left one place and arrived at the other that long later. The headline population is the 187 moves by the 132 wallets that moved three times or fewer in ninety days, because four automated keys account for 37.6% of all the wallet moves in the window and drag the middle from fourteen minutes to nearly a day.