Autonomous AI Hedge Fund: GT AI Fund by the Numbers

GT AI Fund is an autonomous hedge fund. AI agents run a real-money vault on Hyperliquid with no human in the loop, decide every six hours, publish their reasoning as they go, and clear a deterministic risk check before any order reaches the market. This is what two months of it measures out to.

Everything below is a ratio rather than a headline number, because ratios are the part that survives a change in fund size. The measurements come from Hyperliquid’s own public data on 25 August 2026, covering 23 June to 25 August, and every one of them can be re-derived from the venue without asking us. The strategies themselves run on GT App, the same platform any other trader uses.

What is GT AI Fund?

A fund where the portfolio managers are AI agents and the compliance officer is code. Each agent gets a mandate, a budget and the full action space of the trading platform: open a position, size it, set the stop and the target, add to it, close it. It decides on a six-hour cycle and writes down why. Between cycles a watchdog checks every open position every five minutes and can close one without waiting for the next decision.

Nothing is discretionary and nothing is advisory. A risk layer sits between the agent and the exchange and rejects any order that breaks a limit, so a trade that would exceed leverage, concentration or stop-width never reaches the market at all. The capital is real, held in a Hyperliquid vault rather than in our custody, open to anyone who wants to deposit, with depositors keeping 90% of profit and withdrawing after a one-day lock. Alongside it the fund runs several paper books that never touch real money. Those are the laboratory, and they are labelled as such everywhere they appear.

How has it performed?

Two months is a sample, not a track record, and the honest summary is that the fund makes money and also loses it, on a schedule nobody controls.

Measure	Since 23 June 2026
Return, deposit-adjusted	+6.5%
APR reported on Hyperliquid’s vault page	18.2%
Closed round trips	263
Win rate	63.5%
Deepest drawdown in realised P&L	1.3% of the book
Median holding time	9.8 hours
Markets traded	23 perpetuals

The two return figures use different methods. The first is the fund’s own deposit-adjusted calculation over the period; the second is Hyperliquid’s own annualised figure for the vault. Neither is a forecast.

The path was not a straight line, and the sharpest three days of it are a useful illustration. Between 20 and 22 August the book ran 24 trades at a 92% hit rate. Over the following day it ran 34 at 44% and gave back half the gain. Same agents, same rules, same week. Almost every position in that stretch was long, so when the tape turned, the whole book turned with it, which is the single largest source of correlated loss in the fund’s history and the reason the current work is on exposure limits rather than on entries.

What does the trade record actually look like?

Rules-based trading has a signature, and it shows up cleanly in the distribution of outcomes. Wins cluster where the take-profit and the exit logic live. Losses cluster where the stop sits. There is very little in between.

263 closed round trips by return on position. Losses concentrate between −2% and −4%, wins between +1% and +3%.

That shape sets the arithmetic the fund has to beat. Its average win is smaller than its average loss, so a majority of winners is not optional, it is structural:

break-even win rate = average loss ÷ (average win + average loss)

expectancy per trade = win rate × average win − (1 − win rate) × average loss

Geometry	Value
Average win	+2.06% of position
Average loss	−3.17% of position
Payoff ratio	0.65 : 1
Break-even win rate	60.6%
Actual win rate	63.5%
Expectancy per closed trade	+0.15% of position
Closed round trips per day, last two weeks	about 8

The margin is three points of win rate. That is real and it is thin, and it is exactly why the work described two sections down never stops. It also explains why exit quality matters more here than entry quality: at this payoff shape, moving the average win up by half a point is worth more than another point of hit rate.

How is risk enforced before a trade reaches the market?

In code, before the order is sent, not in the prompt and not after the fact. An order that fails a check is rejected outright rather than trimmed, and the agent is told why so its next decision accounts for it.

Limit	Setting
Stop-loss	Mandatory on every position, never wider than 10%
Payoff shape	A stop may not exceed three times its paired take-profit
Leverage	3× maximum on real capital, a fraction of what the venue allows
Concentration	No single position above 15% of NAV
Liquidity reserve	At least 15% of the vault in free USDC at all times
Drawdown halt	An agent 20% below its allocation stops trading and may only reduce risk

The liquidity floor is worth one line of context: it is a hard rule here, while the average hedge fund holds about 15% of NAV in free cash and the median holds under 7%. It exists so that a withdrawal request and a margin call can never compete for the same dollar.

How does the fund decide what ships?

By measuring candidate rules against the book’s own closed trades, and by shipping them switched off.

Some of it survives. A revised stop rule, where the stop width comes from the pair’s own volatility and the position is then sized backwards from a fixed budget of risk, returned 15.3% on the risk it carried against 6.6% for the stops the agents had chosen, measured across 216 closed deals at identical risk. It also cut the worst single trade by roughly two thirds. It shipped disabled and runs on a paper book first, because a replay proves the arithmetic and only a live desk proves the mechanism.

Some of it does not survive, and that half is more useful. We built a layer that weighted the agents against each other and vetoed trades it disliked. It passed its acceptance gates twice. Then we ran the same gates against a placebo, a rule that simply held every position at a third of the size, and the placebo passed too, with a better number. The layer was retired that week. A separate research book that put a full roster of agents through 897 trades over 53 days was closed the same way, on its own result, with the one strategy that paid carried forward and the rest dropped.

This is the part that is hard to copy. Anyone can point the same public models at a market. The asset is the decision journal underneath: every agent’s every decision, the risk verdict it got, and the outcome, on live trades rather than in a backtest.

What is a fund like this worth to an exchange?

An exchange is paid on flow, so the metric that matters there is not return but turnover: how many times the book changes hands per unit of time, and what the venue collects on it.

Book turns per week. The dashed line is the pace of the last thirty days.

annualised turnover = (notional ÷ average equity) × (365 ÷ days)

fee yield to the venue = annualised turnover × blended fee rate

Flow measure	Value
Book turnover, last 30 days	26.9×
Annualised turnover	about 327×
Blended fee rate collected by the venue	3.7 basis points
Fee yield to the venue	about 12% of AUM per year
Share of fills that crossed the spread	63%
Venue’s take against the book’s net profit	0.43 : 1

Two of those deserve a sentence. A blended 3.7 basis points is not a rate card, it is what the venue actually collected across every fill, with 63% of them paying the taker side and no rebate arrangement behind any of it. And the last row says that for every dollar this strategy kept after costs, the exchange earned forty-three cents, with none of the risk and none of the drawdown.

Normalise the turnover to capital and it scales cleanly on paper:

Per unit of capital	Notional per month	Fees to the venue per year
Per $1M under management	about $27M	about $120K
Per $10M under management	about $270M	about $1.2M

The caveat is not optional. That ratio was measured on a small book placing small clips on a nine-hour horizon, and short-horizon strategies are precisely the ones where market impact eats the edge first. We would expect turnover to fall as size rises, so treat the lower row as arithmetic rather than as a promise. What does not change with size is the shape of the flow.

Share of traded notional by market. BTC and ETH together are 12.5% of it.

Three properties make this flow unusual for a venue. It is always on, arriving on a fixed cycle in small clips with no opinion about the hour or the day of the week. It pays taker on most of it. And it lands mainly outside the majors, in the mid-cap books where depth is thinner and fee tiers are higher. Because the strategy lives in a vault, the capital that follows it arrives as deposits on the venue rather than sitting elsewhere waiting for a signal.

What we do not claim is a trader audience. This is one book, run by software, and its worth to an exchange is the turnover ratio applied to whatever capital sits behind it.

Can any of this be verified?

All of it, and not from us. On Hyperliquid the vault is listed under the name Axiom Digital, which is the name to look for. Its equity, its depositor list, its monthly return and its complete trade history sit on Hyperliquid’s own vault page, and every order the agents have ever placed is on the on-chain explorer with a hash. The fund’s own dashboard is a read-only mirror of those surfaces plus the one thing the chain does not carry: the agents’ reasoning, published each cycle, including the decisions that went wrong.

That is the deliberate part of the design. A fund that publishes only its winners is a marketing exercise. A fund whose entire book is on a public ledger cannot be one.

Frequently asked questions

Is this real money or a simulation? Real money on a live venue, in a Hyperliquid vault, since 23 June 2026. GT AI Fund also runs paper books for research, and those are labelled as such wherever they appear.

Who makes the trading decisions? AI agents, on a six-hour cycle, with a five-minute watchdog between cycles. A deterministic risk layer approves or rejects every order before it reaches the exchange. No human takes a trading decision.

Why report ratios instead of absolute returns? Because absolute numbers on a young book say more about its size than about its process. Turnover, payoff ratio, break-even win rate and expectancy are the properties that carry over when the size changes.

What is the break-even win rate and why does it matter? It is the hit rate a strategy needs just to stay flat, set by the ratio of its average win to its average loss. This book needs 60.6% and gets 63.5%, which is a thin and honest margin rather than a claim of an edge.

What happens in a drawdown? An agent 20% below its allocation stops trading and may only reduce risk. Every position carries a hard stop that is checked before the order is placed, and no stop may be wider than 10%.

How do depositors and the fund split profit? Standard Hyperliquid vault terms: depositors keep 90%, the vault leader takes 10%, and deposits unlock after a one-day lock.

Can the same book run on another exchange? The execution platform already connects to centralised venues, so the strategy itself is portable. The vault wrapper, with its on-chain deposits and public trade history, is specific to Hyperliquid.

Measurement note: fills, turnover, fee rate, win rate, payoff geometry and holding times were computed on 25 August 2026 from Hyperliquid’s public fills endpoint for the vault address, covering 23 June to 25 August 2026. Vault equity, depositor count and reported APR come from the same API’s vault detail endpoint on the same date. The vault’s size and every position in it are public on the venue’s own page, linked above.

GT AI Fund: An Autonomous Hedge Fund, Measured