A story about Hyperliquid deposits and withdrawals

On June 8, 2026, $606 million left Hyperliquid in a single day. It left the way everything leaves: through one contract on Arbitrum, in validator-signed batches, paid out to addresses that had been doing the same thing all week. Across the month, $3.95 billion went out against $842 million coming in. Half the capital the platform had accumulated in two and a half years was gone in six weeks.

The timeline had a name for that shape, and people used it.

Hyperliquid is a perpetual-futures exchange running on its own L1, and it has exactly one financial door: the Bridge2 contract at 0x2Df1c51E…63dF7, where native USDC goes in and validator-signed withdrawals are paid out. Every dollar of USDC collateral that has ever traded there crossed that address twice, once in each direction. So we read the whole thing — 6,860,110 transfers, $80.084 billion in and $79.660 billion out, from the first deposit on December 2, 2023 through August 10, 2026.

Rebuilt from the transfers alone, the bridge should be holding $423.6 million. It holds $431.4 million. The gap is $7.8 million on $80 billion of gross flow, 0.0098 percent, and it is the licence for everything that follows: who funds the largest perp DEX, who wins on it, who leaves, and what actually happened in June.

Fifty-eight percent of June's outflow was one address.

The door

For its first year the door was quiet. Through October 2024 the bridge took in between $200 million and $700 million a month and sent out almost as much, and the balance inside it never passed $700 million. It was a serious market with a small float.

Then the airdrop. HYPE's genesis distribution landed on November 29, 2024, and the deposit series breaks in half at that line. November tripled to $2.11 billion. December brought $4.88 billion, and 102,738 addresses deposited for the first time in that single month — more first-timers than the entire preceding year.

The real peak came nine months later, when the money arriving had stopped being airdrop farmers topping up. August 2025 took in $8.22 billion, September $6.57 billion, October $7.19 billion. In August, 16.7 percent of all native-USDC volume on the whole of Arbitrum was a Hyperliquid deposit or withdrawal. One dollar in six on a general-purpose chain was moving through one exchange's front door.

Net the two directions and you get the balance inside the contract, which is the closest thing this market has to a float. It compounded for twenty-two months, peaked at $5.64 billion in September 2025, drifted down through a winter of small negative months, and then lost $3.11 billion in June alone.

Whales at the window

1,283,831 addresses have used the bridge. 516,413 of them went both ways, 419,901 only ever deposited, and 347,338 only ever withdrew — a first clue that money does not necessarily come out where it went in.

The deposits follow a power law hard enough to be a punchline. 3,188 addresses deposited on fifty or more separate days. Between them they supplied $37.46 billion, which is 47 percent of every dollar ever bridged in, from 0.3 percent of the addresses. At the other end, 535,717 addresses deposited on exactly one day, never returned, and put in a median of $371. That is a car payment. It is also the typical experience of using this market.

The whole distribution says the same thing twice. The median deposit is $371 and the median withdrawal $452, while the means are $20,182 and $27,545. Transfers above $1 million are 0.36 percent of the count and 43 percent of the dollars.

One address deserves its own paragraph. 0xecb6…2b00, an EOA, active from January 21, 2024 to June 20, 2026, made 2,488 deposits totalling $8.58 billion. That is 11 percent of everything that has ever entered Hyperliquid, from one signing key. It has never received a single dollar back through the bridge. Whatever comes out for it comes out somewhere else, under other addresses, which is ordinary practice at that size and also the reason no per-address profit figure in this dataset should be read as a person's ledger.

The direct counterparties are anonymous by construction. Of the top 1,000 depositors, exactly one carries a label of any kind anywhere in our data. So we went one hop further out and looked at who funds them.

Binance sits behind the big depositors more than anyone else does. Second is not an exchange at all: $9.78 billion arrives as freshly minted USDC from the zero address, which is Circle's CCTP delivering dollars bridged from other chains. That capital does not originate on Arbitrum. It is routed there to be deposited, and Arbitrum is a corridor rather than a home. On the way out the ranking narrows: Binance first by a distance, then Coinbase.

The unlabelled majority still has a shape, because an EOA signs its own transactions and a contract never does. Sorted that way, $30.8 billion of the flow belongs to public DEX infrastructure — one pool contract alone made 19,272,166 outbound transfers to 205,332 destinations — $10.6 billion to operated desks, $9.9 billion to treasury contracts, $6.2 billion to operated wallets, and $5.0 billion to sweepers, addresses with exactly one destination that exist to forward and nothing else.

The scoreboard

To ask who wins, we took every address that deposited, withdrew, and then went quiet for sixty days or more. That is 481,523 finished accounts, and it is the only population where the arithmetic closes.

168,476 finished ahead. 309,743 finished behind. The median finished account is down $3.85, which is the most honest number in the dataset: the typical user of the biggest perp DEX in the world walked away having lost the price of a coffee. The tenth percentile is down $4,468 and the ninetieth is up $1,978.

The losses concentrate more brutally than the wins do. 7,475 addresses are behind $7.2 billion of the downside, and 6,664 addresses hold $7.4 billion of the upside. Everyone else, all 467,000 of them, are arguing over the change.

And the win rate goes up with size, monotonically enough to be uncomfortable.

Read the right-hand column rather than the bars. Every size band has a negative median return, so this is not a market where big accounts print money. It is one where they lose slowly and everybody else loses fast: the median small account gives back 6.7 percent of what it deposited, the median $10M-plus account gives back 0.6 percent. Ten times the discipline, roughly, for a hundred times the size.

Two caveats belong right here rather than in a footnote. A quiet address may still be holding an open position on Hyperliquid's own L1, which this ledger cannot see. And $14.1 billion of withdrawals went to addresses that never deposited, so some winners and losers are the same trader wearing different hats. Read the shape — win rates, medians, ratios. Do not read a dollar total of user losses out of this; the ledger does not support one.

The exit that wasn't a run

Which brings us back to June, and to a search that failed first.

The obvious test for a whale exit is to sort withdrawals by size and look at the top. Do that for June 2026 and you find nothing: no single transfer above $20 million in a month that moved $3.95 billion. The month looks like 78,364 ordinary withdrawals to 43,289 ordinary recipients.

Group by counterparty instead and the month collapses into one line. 0x48594ff0…1f0b received $2,301 million, 58 percent of everything that left in June, in 330 transfers, not one of them above $10 million.

That address is a pass-through and nothing else. It was first seen on May 18, 2026. It has received $2,873.7 million all-time and sent $2,873.7 million, to the cent. It holds nothing. Seventeen addresses fund it and it forwards to fifteen, seven of which carry Coinbase hot-wallet labels; its single largest destination took $761.8 million.

Why it moved is not in the ledger, and we are not going to guess in print. A custody rotation, an OTC settlement, an institution reducing exposure and a fund changing venue all produce the same trace. What the ledger does establish is what June was not: it was not 43,000 users running for the door. Strip the router out and $1.65 billion left across everyone else, which is an elevated month, not a panic.

The run was one address with a spreadsheet, and a chunking rule under $10 million.

There is a reason nobody had seen this. The raw transfer table we started from was wrong. A re-ingest on July 19–20, 2026 duplicated withdrawal rows for November 2024 through February 2025 and part of August 2025, and dropped roughly $531 million of August 2025 deposits from the newer copy. Summed naively it claims $6.0 billion more money left Hyperliquid than ever entered, which is impossible on a contract that cannot pay out what it does not hold. We deduplicated by transaction, counterparty, amount and ingest generation, then re-derived three block windows against Arbitrum JSON-RPC to check the result — 22 transfers and $72,028.75 inside the corrupted period, exact to the cent. The reconciliation with the live bridge balance is the final check, and it is what separates this from a dashboard quoting the broken number.

The more consequential finding in June is not the router. It is what the deposit side did afterwards, which is nothing. $842 million in June, $730 million in July, $95 million in the first ten days of August. Deposits did not bounce.

Two slower measurements explain why that matters. The first is speed. Because holding periods are censored by the end of the data, we measured each quarterly cohort by the share that had left within a fixed window. Money moves through roughly twice as fast as it used to: same-day exits went from 5.8 percent of the 2024 Q1 cohort to 13.3 percent of 2025 Q4, and ninety-day exits from 21.4 to 45.7 percent.

The second is where deposits come from. First-time depositors have contributed $15.48 billion of the $80.08 billion ever bridged in. Nineteen percent. The engine was never new users; it was the same addresses topping up, and topping up again.

What's left

The bridge holds $431 million today, against $5.64 billion at the peak. Hyperliquid's share of native-USDC volume on Arbitrum has fallen from 16.7 percent last August to 2.2 percent this one. On the busiest chain-level measure available, the door is a tenth as busy as it was a year ago.

The address-level picture is bleaker still and then quietly reverses itself. Of the 102,738 addresses that arrived in the December 2024 airdrop wave, 11.4 percent were still transacting three months later, 8.6 percent at six, and 4.7 percent at twelve. Ninety-five of every hundred were gone within a year.

Their money was not.

The same pattern holds for every cohort we tracked: addresses vanish, dollars persist and often grow. The people churn and the capital consolidates into fewer, larger, more permanent hands. Whale persistence measures the same consolidation from the other side — of the top-100 depositors of 2023, seven were still top-100 in 2024; of 2024's, twenty-seven survived into 2025; of 2025's, thirty-three are still there in 2026. The top decile of this market is becoming a fixed roster. Top-ten concentration has risen every year, from 17.3 percent of deposit volume in 2023 to 31.3 percent in 2026.

The small details survive the same way. Users have spent 26.99 ETH of gas across 3.94 million deposit transactions — the cost of admission to an $80 billion market, split a million ways. 211 deposit transfers simply failed. And people have sent the bridge $5,026 of USDC.e, $20,203 of USD₮0, $221 of USDT and at least four counterfeit tokens whose ticker renders as "USDC", none of which it credits to anyone.

The door is still open. The queue outside it is gone, and the people who never left own more of what is inside than they used to.