Jane Street Made $39.6 Billion While AI Trading Contests Lost Money

Three facts get called “AI trading,” and they have almost nothing in common.

Fact one: Bloomberg reported that Jane Street made $39.6 billion in trading revenue in 2025, running tens of thousands of GPUs. Fact two: public contests pitting leading AI models against markets mostly lost money, and an academic study over two decades and 100+ symbols found earlier reported advantages for LLM strategies deteriorated. Fact three: the CFTC warns that scammers use “AI” to promise guaranteed returns, citing a case where more than $1.7 billion in bitcoin was taken from at least 23,000 people.

Same label. Three different universes. Here is how they actually differ.

What hedge funds do with AI

At firms with proprietary data and enormous compute, AI works in narrow, supervised roles. Man Group's AlphaGPT proposes signals, writes the code, runs backtests — and then humans review everything, after the team ran into hallucination, look-ahead bias, and multiple-testing problems. Hudson River Trading trains models on more than 100 terabytes of market data. An S&P Global study found LLM signals earned 8.4% a year vs 4.2% for traditional methods in a long-short backtest — promising, but a backtest, not live trading. Note what is missing from every serious account: nobody credible claims an unsupervised model prints money while humans sleep.

What public tests show

Bloomberg's May 2026 report on trading contests found most leading AI models lost money and traded too much. In Nof1's Alpha Arena, most model accounts lost money with real capital. LLM strategies were too conservative in bull markets (underperforming passive benchmarks) and too aggressive in bear markets (taking heavy losses). Most earlier studies covered under a year and fewer than ten stocks — the kind of thin sample that flatters any method.

What scams do with the same words

“AI” becomes a permission slip: guaranteed returns, 100% win rates, secret algorithms. The CFTC's position is blunt — AI cannot predict the future or sudden market changes, and guaranteed returns are a scam signal. The $1.7 billion case is the extreme end of a spectrum that starts with every “AI bot” ad promising autopilot profits.

The line that separates them

Supervised vs unsupervised. Human-reviewed, data-rich, cost-counted, validated AI assistance works in narrow roles. Unsupervised money machines mostly fail — and the ones sold to retail fail while charging you for the privilege. If someone sells you “what Jane Street does” for a small download fee, ask yourself what Jane Street's tens of thousands of GPUs, proprietary data, and hundred-terabyte training sets have in common with a downloadable bot. The answer is a word: “AI.” Nothing else.

Before you trust any AI trading claim, run it through the same standard: recompute from raw data, beat the baselines, survive the folds, show a public record. Jane Street's record is audited revenue. A vendor's record should at least be openable.

BuildIT Index applies that standard to any prediction file: recompute accuracy, beat baselines, check folds, count trials, return a verdict with an exit code.

Supervised vs unsupervised. Human-reviewed, data-rich, cost-counted, validated AI assistance works. Unsupervised money machines mostly fail.