Crypto AI Vaults: 3 of 6 Beat Bitcoin, 1 Fell 30% — What That Actually Proves

Crypto is where AI trading claims can actually be checked, because Hyperliquid vaults publish their numbers on-chain. No screenshots, no self-reported PnL — just public data. So what does the checkable corner of AI trading actually show?

As of the 2 November 2026 verified check: of 6 Hyperliquid vaults whose leaders say they are AI-run, 3 beat Bitcoin over the same dates, 2 lost money, and 1 fell more than 30% from a peak. Among the vaults passing strict safety and drawdown gates, results ranged from +26.5% (with a −24.2% max drawdown) to −0.1% (with a −10.8% drawdown) — and the top-ranked vault overall, not AI-labelled, returned +24.1% with an −11.9% drawdown while trailing Bitcoin by 4.2 points over the same period.

First, “AI-run” predicts nothing

The label is the operator's claim, not a performance category. AI-labelled vaults appear across the whole range — beating Bitcoin, losing money, and crashing. If the label predicted returns, the table would not look like this.

Second, beating Bitcoin is the right benchmark — and most don't

Holding Bitcoin costs nothing, needs no skill, and takes no drawdown decisions. Any vault that trails it over the same dates charged you complexity, fees (typically a 10% profit share for the leader), and drawdown risk for less than doing nothing. Several vaults in the data did exactly that.

Third, drawdown is the story the return hides

A vault up 26.5% that fell 24.2% from a peak is a different proposition from a vault up 8% that fell 3%. The first can wipe out a depositor who arrived at the wrong time; the second is boring in the best way. Rankings by return alone are how people buy the top of a drawdown.

How to evaluate any vault in two minutes

(1) Return vs Bitcoin over the same dates. (2) Max drawdown from peak. (3) Leader's profit share and stake. (4) How long it has been live (a month of numbers is an anecdote, not a record). (5) Whether deposits are open and what happens in a drawdown — who bears it, you or the leader?

The checkable corner of AI trading does not vindicate the hype or confirm the dismissal. It shows variance: some AI-labelled vaults beat the benchmark, some lose, one fell hard. That variance is the honest answer to “do AI vaults work?” — sometimes, for some operators, with drawdowns that demand respect. Anyone selling you a simpler answer is selling you something else.

BuildIT Index applies the same standard to backtests: recompute from raw data, beat baselines, check folds, count trials, model costs. A vault that cannot survive those checks on its own history is a bet on luck, not skill.

If the label predicted returns, the table would not look like this.