Hundreds of thousands of machines are now racing to price every corner of reality down to the millisecond. That should make these markets so accurate they stop being worth playing. Kalshi's fee revenue suggests the opposite.
In a single month, a Polymarket wallet turned about $300 into more than $400,000. It never predicted anything.
The wallet traded 15-minute contracts on Bitcoin, Ethereum and Solana, and its entire strategy was to notice that Binance and Coinbase already knew the price before Polymarket's order book did. The window it traded through was measured in milliseconds. No thesis about the world, no view on where crypto was heading, just a machine standing between two clocks that were slightly out of sync.
That wallet is not the exception. A review of Polymarket's public leaderboard found that 14 of the 20 most profitable wallets belong to bots. Between April 2024 and April 2025, arbitrage traders pulled an estimated $40 million out of the platform, almost none of it for being right about anything.
Hence the hope in the headline. If a rapidly growing crowd of machines is going to spend the next several years grinding every mispricing on earth down to nothing, the honest conclusion is that these markets should eventually become so accurate that there is no reason for a human being to place a bet in one. That is the good version. I would like it very much.
I do not think the fee revenue is going to cooperate.
The important detail in that $400,000 wallet is what it was not doing. It was not modelling Solana's fundamentals. Latency arbitrage is a plumbing trade: when information moves, prices on different venues update at different speeds, and whoever reacts first collects the difference. The skill is in the wiring.
The lengths people go to are genuinely funny. CNBC reported on August 1 that one full-time prediction-market trader bought an over-the-air antenna so he could watch Super Bowl commercials live, because streaming services run a few seconds behind broadcast, and a few seconds is an eternity when there is a market open on which ad airs next. Somewhere between that antenna and a co-located server farm sits the entire business model of modern prediction markets.
David Minarsch, co-founder of Valory AG, put the situation plainly: "You have human participants in prediction markets alongside many machines. So humans are already in a battle with machines."
He is describing a battle that is already over. The retail trader reading a news alert on a phone is competing with software that read the same alert, priced it, and finished trading on it before the phone finished vibrating.
This is the part of the thesis that holds up, and it holds up on ordinary market logic rather than anything exotic about AI.
Arbitrage is self-consuming. Every machine that shows up to harvest a 3% gap between Polymarket and Kalshi makes that gap smaller and shorter-lived for the machine that shows up next month. Spreads that used to sit open for seconds now close in milliseconds. There are already more than 100,000 active markets across the two platforms, and the marginal cost of pointing a language model at another one keeps falling.
Run that forward and you get a live, public, machine-maintained probability on almost everything, updated continuously, free to read, and very difficult to beat. The informational product these companies sell becomes a commodity, and then becomes furniture.
The forecast itself would get genuinely good. That part I believe.
Efficiency destroys the trader's edge. It does nothing whatsoever to the operator's edge, because the operator was never betting. Kalshi and Polymarket do not need the price to be wrong, and they do not need you to lose. They need volume, and they take a cut of it.
Kalshi has generated roughly $1.15 billion in cumulative fee revenue since launch, with about $850 million of that in 2026 alone. On peak World Cup trading days, daily fee revenue reportedly topped $13 million. None of that required a single market to be mispriced.
A perfectly efficient market is, from the operator's chair, the ideal product. It is maximally liquid, maximally defensible, embarrassment-proof, and it still charges rent on every transaction that crosses it. The vig does not care whether the price is right.
The machines grinding out that last inefficiency are finishing the casino, not dismantling it.
| What grew | From | To |
|---|---|---|
| Combined monthly volume, both platforms | Under $5B Sept 2025 | About $24B April 2026 |
| Kalshi monthly volume | $16.8B May 2026 | $31.5B June 2026 |
| Kalshi valuation | $11B Dec 2025 | $22B May 2026 |
For scale, legal US sportsbooks averaged around $14 billion in monthly wagers through 2025. In June 2026, these two platforms did $44.8 billion between them.
There is a second thing efficiency threatens, and it is worth more to these companies than any trading edge.
Both platforms are sold on a civic premise: that a market is the most honest forecasting instrument ever built, that it aggregates dispersed knowledge better than pundits or polls, and that betting on outcomes is a public good because it produces an unbiased signal. It is a good argument. I have made versions of it myself.
Pew Research measured what people actually trade. From July 2024 through April 2026:
During the 2024 US election, political contracts were 90% of Kalshi's volume. Today they are 4%. The forecasting instrument that was going to out-predict the polls turned out to be a sportsbook that had a very good November once.
And this is where near-perfect machine pricing genuinely does damage. If a public, free, continuously updated probability exists for every question that matters, the "we produce a valuable social signal" defense stops being a reason for the exchange to exist. The signal exists with or without a retail customer taking the other side of it. What is left in the building is the vig and the sports.
The legitimacy story is not a marketing problem. It is the legal foundation these companies stand on, and it is under active attack right now.
Kalshi's position is that its event contracts are federally regulated derivatives on a CFTC-designated market, which puts them outside state gambling law entirely. States disagree, loudly:
Read that list next to the Pew numbers and the shape of the fight is clear. The federal argument protects a price-discovery instrument. The actual business is 80% sports. Every month that gap widens, the states' case gets easier to make and the "truth machine" defense gets thinner.
So if these companies die, the cause of death on the certificate will be a court ruling or a rewritten CFTC rule, not a bot achieving perfect efficiency. The machines just make the argument harder to win by making the informational excuse redundant.
The uncomfortable part is that the profitable version does not require anything to go wrong. It only requires things to continue.
Nobody in this business needs the public to get worse at probability. The current arrangement already works: people who enjoy betting will bet, the machines will keep the prices honest, and a percentage of an enormous number is an enormous number. Kalshi doubled its valuation from $11 billion to $22 billion in five months and was reportedly in talks at roughly double that again seven weeks later. That is not the funding pattern of a business anyone expects to be arbitraged out of existence.
I do not know which way this goes, and I want to be honest that the evidence available right now points the wrong way for the hopeful version. The volume is climbing, the fees are climbing, the valuations are climbing, and the only genuine threat on the board is a set of court cases whose outcome nobody can call. Prediction markets themselves would be the natural place to look up the odds on that, which is a joke that works better than I would like.
What I am reasonably confident about is the mechanism, because the mechanism is just arithmetic: the machines will keep getting faster, the prices will keep getting better, and none of that touches the rake. Whether that ends with these apps becoming a boring public utility for probabilities, or a very large sportsbook with an unusually good origin story, is not a question the math answers. It is a question about what regulators decide gambling means, and that fight is only nine states old.
Nothing here is investment advice, and I hold no position in any of it. Figures are as reported on the dates shown and this is a fast-moving story.
Kalshi says forecasting instrument. The volume says 80% sports. That gap is most of what I write about, across AI, markets and whatever else is currently explaining itself.
Unsubscribe anytime. Your email stays private.