The new commodity of the AI economy is not the model, and it isn't even the token. It's the probability. A wire report out of New York this week — picked up in The Hindu and across a dozen outlets — describes a startup scene where the thing being manufactured and sold is no longer a forecast. It's a calibrated number about a future event, good enough to trade on. The numbers in the story are the hook: Theo Sommer and Veniamin Veselovsky, the co-founders of Preseen, used AI tools to turn $35 into $1.9 million in seven months on Kalshi, the American prediction market. The method is the more important fact: they're not selling predictions. They're selling the probability of an event, with the ability to cross-check the assumptions across interconnected outcomes — the price of oil against a Middle East conflict, a chart position against a release cycle. Read that again, because it's the whole business model in one line. The product is a number. The number is what moves. The market is where the number gets tested. And the buyers, in Veselovsky's words, are the hedge funds, "just because if they think they can make money off something, they can move pretty fast." That's not a forecasting story. That's the price mechanism of the next decade, arriving with a receipt.
What the Wire Actually Reports
The facts, in order of how much they should matter to the reader:
- The capability claim: AI forecasting models are already on par with the best "superforecasters" — the gifted individuals whose estimates are markedly better than ordinary people's — thanks to troves of data from public and private sources. In early September, two AI models took the top spots in the Metaculus Cup, a competition among prediction champions, both human and non-human. The cup is the proof that the parity is not a sales pitch. It's a measured result, in a public, scored, public event.
- The $35 story: Preseen's founders used AI tools to turn $35 into $1.9 million in seven months on Kalshi. The number is the headline, but the structure is the point: the firm's edge is not a single model. It's a cross-check across interconnected events, so the probability of one thing moves with the probability of the things that cause it. That's the difference between a forecast and a model of a causal web.
- The human in the loop is scared, and says so: FutureSearch's in-house superforecaster, Tom Liptay, is aware, per co-founder Dan Schwarz, that "he's not secure in his job" in the medium term. Schwarz: "He knows that these things are going to surpass him." The most candid line in the story is not a forecast. It's a forecaster describing his own replacement.
- The market is already automating: Kalshi and Polymarket, the two prediction platforms now used by millions of Americans, are, per the report, being "inundated by AI use." And the next layer is already here: the most advanced models no longer just predict events. They formulate their own questions about the future and assign probabilities to the outcomes of those questions. The model is starting to choose the market.
- The retail lock-out is the warning: Veselovsky's line is the one that should get quoted in the next regulation paper: "It might soon be too late for retail investors, who don't have access to expensive AI tools, to profit on those types of consumer-oriented trading sites, because they will soon just be dominated by AI forecasting systems." A consumer-facing market, dominated by a system the consumer can't afford. That's the shape of the asymmetry, and it's the one that has no natural limit.
The Probability as a Commodity
Here's the economic structure underneath the story, because it's the part that will outlast any single model. A probability is a commodity with three properties that make it the natural output of the AI era: it's cheap to produce in quantity, it's calibrated by the market the moment it's traded, and it's the input to a decision that someone else is going to make. The model doesn't need to be right. It needs to be right more often than the market's price. The $35-to-$1.9M result is not a claim that the model knows the future. It's a claim that the model's probability is, in aggregate, better calibrated than the price the market was charging. That's a trading edge, and trading edges don't need to be permanent. They need to be present. And they are present, across a sector, with a buyer base that is already the fastest money on the street.
The extension past finance is where the commodity story gets real. The report names the next use cases in one breath: government policy, corporate decision-making, scientific research, insurance risk assessment. Mantic is already providing a large American company with likely scenarios of regulatory developments in its sector, and has worked for the British government. FutureSearch is being asked to help decide which medical treatments are worth funding. The line between "forecasting a stock" and "forecasting a policy" is the line between a financial product and a public one, and the report is telling you the line has already been crossed on the corporate side. The probability commodity is now a policy input. That's not a metaphor. That's a supply chain.
What It Means
1. The "who's the best forecaster" question just died, and the "who's the best calibrator" question took its place. For two decades the answer to "can AI forecast?" was a moving target of benchmark scores. The Metaculus Cup is the last honest scoreboard, and the result is that the target has stopped moving: the models are on par with the best humans, and the humans know it. The next question is not whether the model is better. It's whether the model's probability is better calibrated than the market's price, on the specific event, in the specific window. That's a trading question, and it's the question the $35 story answers with a receipt. The forecasting industry just became a calibration industry, and the calibration industry is a market.
2. The retail lock-out is the asymmetry with no natural ceiling, and it's the one a regulator will have to decide on. A consumer-facing prediction market, used by millions, dominated by a system the consumer can't afford, with an edge that compounds with every model improvement, is not a market. It's a toll road with a free lane that's about to be closed. The question is not whether the AI systems will dominate. The report says they will. The question is whether the platform, the state, or the consumer gets to decide when. And the answer is going to be shaped by the fact that the platform is already "inundated by AI use," which means the disclosure, the matching, and the pricing are all now running through a system the retail user is not the customer of. The consumer is the input.
3. The "model asks its own questions" line is the one that will define the next two years of the space. The report's claim that the most advanced models no longer just predict events but formulate their own questions and assign probabilities to their outcomes is the quiet, most important sentence in the story. A model that chooses its own questions is a model that defines its own market. It's the difference between a forecasting tool and a forecasting economy, and the difference is going to show up in the product: the model will sell the question and the answer together, because the value of a probability is the value of the question it's attached to. The model that picks the question picks the trade. The next forecasting product is not a number. It's a menu of numbers, and the menu is the product.
4. The "scared because Russia or China has better systems" line is the geopolitical read the whole story is hiding. Veselovsky's closing worry — that maybe Russia or China has better systems — is not a footnote. It's the strategic read of the commodity. A probability that is better calibrated than the rival's is a decision that is better made than the rival's. In finance it's a trading edge. In policy it's a lead time. In military, in medicine, in the regulation of the other side's sector, it's a force multiplier that doesn't show up in the arms inventory. The forecasting startup is the first place in the AI economy where the output is a decision, not a capability, and a decision is a thing a state can act on before the rival state acts on it. The $35 is the receipt. The decision is the weapon. And the two capitals are already in the room.
🔥 Hot Takes
1. The $35-to-$1.9M number is not the story, and anyone who reads it as a payday story is going to miss the structure that made it possible. The structure is: a calibrated probability, cross-checked across a causal web, sold to the fastest money on the street, in a market that prices the probability back to you the moment you trade it. That structure is a factory for edges, and the $35 is the cost of the first edge in a factory that can run forever. The next $35 is going to be someone else's, and the next $1.9M is going to be theirs. The edge is the structure. The structure is the commodity. The commodity is the probability. The receipt is just the proof of concept.
2. "He's not secure in his job" is the most important sentence in the report, and it's a forecaster saying it about himself. The superforecaster is the last human in the loop that the industry still pays for, and the industry's own in-house expert is the one who's named the end date. That's not a morale line. That's a schedule. The calibration work that the human was doing, the work of being the thing the market prices against, is now being done by the model, and the market is moving to price the model. The next step is the market pricing the model's menu of questions, and the step after that is the market the model chose for itself. The human forecaster is the last witness to the shape of the old trade, and he's the one who said the new shape is already here.
3. The retail lock-out is going to be the first AI-market regulation that isn't about the model, and that's what makes it the one that matters. Every AI regulation so far has been about the model: the guardrails, the testing, the alignment, the safety. The prediction-market lock-out is the first AI regulation that's going to be about the market: the disclosure, the matching, the pricing, and who gets to be the customer when the customer is the input. That's a different object, and it's the object that decides whether the next decade of AI finance is a market or a toll road. The report already told you the toll is coming. The regulation has to decide who pays it, and the only people who can pay it are the ones the system isn't built to charge. That's the asymmetry with no ceiling, and it's the one a regulator will have to pick up, because the market won't pick it up for free.
The Bottom Line
The AI economy's new commodity is the probability, and the report is the first honest inventory of the factory that makes it. Two models took the top of the Metaculus Cup. A $35 bet became $1.9M on Kalshi. The in-house human forecaster is the first to say his job is on a clock. The fastest buyers are the hedge funds, and the next buyers are the governments and the insurers and the medical funders. The models are starting to pick their own questions, which is the line between a forecasting tool and a forecasting economy. And the retail investor, who built the consumer-facing market, is the one the system is about to stop being built for. The probability is the commodity. The commodity is the decision. And the decision is the thing that moves first. The $35 is the receipt. The decision is the weapon. The market is where the two meet, and the market is already full.