The Power Law Is Not a Bug, and AI Won't Fix It

The pitch I hear most often now, usually from someone who has never lost their own money on a deal, is that AI is going to make venture predictable. Feed a model enough decks and outcomes and the guesswork drops out. I've built the model. It doesn't work like that, and the reason is a piece of arithmetic the whole industry is organised around and mostly refuses to say out loud.

Why doesn't AI make venture capital predictable?

Venture is a power law, and a power law is not a distribution you can average your way through. Roughly half of all venture investments return less than the money put in. A tiny fraction — a few percent of the cheques — generate the large majority of the profit. I don't need a proprietary dataset to tell you that; the fund-level numbers have been public for years. But I did just watch it draw itself again in our own. This week we finished an outcome graph of roughly 6,900 accelerator companies — every Y Combinator batch, every Startmate cohort — and the shape is unmistakable: about one in six of the companies with a known status is dead, most of the rest are quietly ticking along, and the genuinely large outcomes concentrate into a short list of names that recur across the exits.

Most bets fail or fizzle. A handful pay for everything. That is the physics, not a market flaw awaiting a smarter algorithm.

Why the industry looks irrational

Almost everything strange about how venture operates is a rational response to that arithmetic. If you can't predict which bet is the outlier, you optimise for access to the most bets and for never being the one who passed on it. Spray-and-pray, logo-chasing, "we invest in lines not dots" — these aren't laziness, they're what you do when the payoff lives in a tail you can't forecast. Judgment on the median deal looks pointless, because the median deal is a rounding error against the one that returns the fund.

What does AI actually change in venture investing?

The economics of evaluation — not the shape of outcomes. When it costs almost nothing to read and rank a thousand companies, the binding constraint moves. It's no longer "how many can a partner physically get through," it's "how well can we order them." You can widen the top of the funnel without dropping the bar at the bottom. That's a real gain, and I don't want to undersell it. Clearing the structurally weak, surfacing the base rates, killing the noise — machines are good at that, and it frees a scarce human to look harder at fewer things.

The trap: a variance game, not a mean game

Here's the trap, and it's a subtle one: a power law is a variance game, not a mean game. You are not trying to maximise the expected value of the average deal. You are trying to maximise your exposure to the right tail. A model optimised for "most likely to succeed" — which is what almost every scoring system quietly optimises for — will systematically strip out the high-variance, non-consensus bets, because those are exactly the ones that look riskiest on the mean. Let that model gate your pipeline and it will hand you a portfolio of sensible companies that never returns a fund.

It didn't pick badly. It picked for the wrong statistic.

The outlier is the thing the machine sees worst

The company that becomes the power-law winner is usually the illegible one — the founder who reads as too early, too weird, too different, the market that isn't a market yet. My own research keeps landing on the same uncomfortable point: across 172 deal memos, the single most predictive signal of who was still standing years later was conviction — the least legible thing on the page — and the structured frameworks I'd trusted weighted it at around ten percent. Meanwhile the tidy, legible signals only take you so far; in that same work, product execution correlated with outcomes at about r²=0.77 and team quality at barely 0.31. A model trained on consensus will always mark down the non-consensus bet, because non-consensus is, by construction, under-represented in what has already worked. It regresses to the mean. The mean is not where the returns are.

So what is AI good for in a power-law game?

Not picking winners. Honesty about the odds. A well-calibrated score doesn't tell you a company will succeed — it tells you the base rate for companies that look like this one, so that when a human decides to pay up and defy that base rate, they know exactly what they're overriding and why. That's the correct relationship: the machine holds the probabilities steady; the person supplies the conviction to bet against them when it counts. Calibration you can only earn by scoring a lot, waiting, and counting the failures — the slow work no prompt shortcuts.

So the division of labour I'd actually stake a firm on: let the machine raise the floor, and keep the ceiling human. Clear the noise, calibrate the odds, widen the funnel — then spend conviction, the scarce and illegible thing, on the tail where the money has always lived.

AI won't make venture predictable. Anyone selling you that is selling you a filter tuned to the median in a game decided at the extremes. What it can do — if you build it honestly, on the failures as well as the wins — is make you honest about the odds.

The odds were never the hard part. The bet still is.

Frequently asked questions

Does AI make venture capital returns predictable?

No. Venture returns follow a power law — roughly half of all investments return less than the money put in, and a small fraction of the cheques generate the large majority of the profit. That isn't an inefficiency a cleverer model arbitrages away; it's the physics of backing things that mostly fail so you can own the few that don't. AI changes how cheaply you can read and rank companies, not the shape of that distribution.

What is the difference between accuracy and calibration in scoring?

Accuracy is whether the eventual winners were ranked above the eventual losers. Calibration is whether a score of 70% actually happens roughly 70% of the time. A score can be accurately ordered and still badly calibrated — and calibration is the property you can only earn by scoring many companies, waiting for real outcomes, and counting the results, including the failures.

Why does an AI scorer struggle with outlier startups?

Because the outlier is, by definition, the low-probability, non-consensus bet that looks wrong against everything that has already worked. A model trained on past winners regresses to the mean and marks the outlier down — yet in a power law those rare, illegible bets are exactly where the returns concentrate.