From $5 million to $40,000: what an AI-run trading firm still needs a human for

Brian Kelly estimates his old trading team cost about $5 million a year. He says his AI setup now costs $30,000 to $40,000 a year. CNBC, September 8, 2026.

That is less than 1% of the old figure. It is the kind of number that travels fast, so it deserves a slow read.

What is actually being reported

Kelly, a former crypto hedge fund manager and "Fast Money" trader, closed his previous fund in early 2025 and began testing AI later that year. His new firm, Bracket22, trades crypto, stocks and commodities using only his own money.

CNBC describes three named agents. Steffi handles technical analysis. Desmond handles quantitative strategies. Houston acts as mission control and combines their outputs. Kelly makes the final decision himself.

Both cost figures are his estimates. The $5 million covered a team of seven or eight, including office space and bonuses. The newer figure covers the agents, the compute and what he says it takes to replicate a hedge fund with AI.

Three things we noticed

The observations below are our reading, not CNBC's.

The design is more interesting than the savings. Kelly keeps his agents isolated so he gets unbiased views. Think of analysts who write their notes before the meeting, so the first opinion in the room does not shape the rest.

That raises a useful question: if agents are kept apart but built on similar models or data, how independent are they? The Bank of England's Sarah Breeden made a version of this point at market level in June, warning that agents trained in similar ways on similar data could "amplify volatility in stress." The Next Web.

The missing number is the return. The report covers cost and productivity. It does not report trading performance. Kelly says he is "at least 10 times more productive", but productivity measures output, not whether that output makes money.

A cheaper desk making the same decisions is a better business. A cheaper desk making worse ones is not. This story cannot tell us which one we are looking at. And because the capital is his own, there are no outside investors asking for a track record.

The job that stayed human is the one nobody can hand off yet. Agents can produce views around the clock. Someone still has to own the position.

Regulators are asking the same thing. At a forum on October 28, the CFTC's chair asked: "Who's liable if the agent goes rogue?" It was a question, not a rule. CoinDesk.

There is a quieter risk too. CNBC's piece notes a Goldman Sachs partner warning that AI could erode bankers' reasoning skills. If a person's main job becomes approving what agents suggest, does that judgment stay sharp?

What we would ask any AI-run desk

• What are the returns after costs, over what period and in what market conditions?
• What happens when the agents disagree?
• Who or what can stop trading, and how quickly?
• Can you reconstruct why a specific trade was made?

None of this is a criticism of Bracket22. These are simply questions the public reporting does not answer, and they apply to anyone building this way.

The cost is the headline. Control is the question.

What would you need to see before you trusted a desk like this?

References

• CNBC: How one hedge-fund manager built his firm to be powered entirely by AI agents, September 8, 2026. Source for all Bracket22 and Brian Kelly details, including the cost estimates, and for the Goldman Sachs partner's warning.
• The Next Web: Bank of England's Breeden warns AI agents could trigger market meltdowns, June 30, 2026.
• CoinDesk: CFTC chief poses questions about policing AI, October 28, 2026.

FinRAG.tech has no connection to Bracket22 or Brian Kelly. The figures are Kelly's estimates as reported by CNBC and have not been independently verified by us. This article is commentary, not financial advice.