Sinéad O’Sullivan is a former senior researcher at Harvard Business School, professor in aerospace engineering and now works at CFM, a multi-strategy quant hedge fund.
If you’ve never played poker in your life, there’s a great formula to guide you on exactly how much to bet on any given hand.
It’s called the Kelly Criterion, and it says the size of your wager should be set by the size of your edge — not by the size of the pot, the Richard Mille on the wrist of the other players, or the number of dirty martinis flowing through your bloodstream. So if you bet more than your edge supports, you do not get richer. You get wiped out.
The late Jim Simons delivered this as a masterclass. His hedge fund Renaissance Technologies “discovered” — he was a mathematician, after all — that the correct amount of assets under management for its fabled Medallion fund was about $10bn, roughly one fifteen-thousandth of the global equity market.
And despite investors being desperate for a piece of the Medallion pie, it has been capped around there for two decades, paying profits out every six months to stop the fund growing. In 2003 it even expelled every external investor rather than manage more money, and now just consists of employees.
Which brings us to three former DeepMind researchers who taught a computer to play no-limit hold ‘em, partnered with high-frequency trading firm Tower Research, and the venture capital firm that has just valued their Prague-based AI lab at $500mn — partly on the grounds that the equities market is really, really humongous.
As TechCruch reported late last month:
The common denominator between poker and Wall Street is that they are well suited for reinforcement learning, an AI training technique where self-learning models are incentivized by rewards. According to Martin Schmid, EquiLibre CEO, “The nice thing about trading and markets is that the scoring is super simple: how much money did the agent make?”
This isn’t just game money. In partnership with quant firm Tower Research Capital, EquiLibre’s algorithms have been trading billions in daily volume across the S&P 500 and Nasdaq. The startup claims its agents have been doing well since their rollout on crypto markets in 2025, and now on stock exchanges, with “a perfect record of zero negative months since inception,” meaning they have finished each month with their investments up overall.
By applying its AI to quant hedge funds, the startup is in a field where automation is commonplace and, if successful, improvements can quickly turn into cash. That made the startup appealing to Creandum, [its vice president Cameron] Sellers said.
“The potential total addressable market of trading in the financial markets is one of the biggest on earth, and there are countless funds over the years that have generated quantums of profit that make most venture-backed successes look small,” Sellers said. But he noted that EquiLibre explicitly defines itself as “a lab first, not a finance firm.”
Which begs the question: Is AI taking over quant funds? Maybe VCs think so. After all, from the outside a quant fund looks a lot like a software company: there is code and data, there are pedigreed engineers and mathematicians, and there are algorithms that get backtested and deployed.
And there are what look like SaaS unit economics; no marginal cost to using the algorithm on one dollar versus billions, in the same way that there’s no marginal cost to serving one more user. All of this makes for a beautiful P&L graph, up and to the right, that would make any Series B founder envious.
So — what is a quant fund, if not a tech company? Quite a bit more, actually.
It’s a book of clients that takes years to build and a pile of assets that can walk out the door on a redemption notice, all sitting atop a collection of trading strategies which, the more successful they become, erode non-linearly and grotesquely fast in value.
Which is precisely, symmetrically, exactly backwards from what a venture capitalist wants to buy.
A VC wants deals that appreciate non-linearly: more users make the product better, which attracts even more users, which creates the infamous moat. And copycats don’t matter in moat-land, because a competitor can copy your technology but not your network. That is the entire spiritual basis of the asset class — buying something that gets better the more the world uses it.
On the contrary, a trading strategy gets weaker the more the world uses it. Its existence is predicated on a price being wrong, and it is paid for correcting that price. Eventually, that mispricing closes; the more that use the edge, its arbitrage disappears into ubiquity by other, dumber people running your same idea on slower hardware.
And if you try to scale more capital through it — which is exactly what venture capital does to scale an investment — it degrades even faster, because your positions grow relative to available liquidity and your own orders start moving the prices you were hoping to trade against. Gah!
Moreover, software has economies of scale, while trading has diseconomies of scale; consider that the marginal user is free, while the marginal dollar is taxed, and the tax rate rises with every dollar that went before it.
So why would Creandum write its largest-ever cheque into an AI lab playing quant shop, with none of its own AUM, no clients and no traders? Alphaville asked Cameron Sellers, the firm’s vice-president, who agreed that no single strategy guarantees success over the long term — Creandum would never underwrite the deal on that basis, he said. The firms that compound over decades, Sellers argues, do it by continuously producing new methods rather than defending one or two insights.
Correct. Which suggests Creandum isn’t claiming the lab has found a magical, everlasting strategy, but something closer to a factory that keeps generating fresh edges as old ones decay. So is it a factory?
The evidence Sellers offers is that today’s model beat last year’s model by more than 10x — on back tests. Which might be true, but there are two major problems here. Firstly, the distance between a back test and a live P&L is the most infamously expensive gap in systematic trading. Secondly, Sellers is saying that the same lineage is merely getting better. What EquiLibre needs to show is evidence of edge replacement, not edge refinement.
Martin Schmid, EquiLibre’s chief executive, doesn’t dispute the curve problem at all. In fact, he agrees with it. VC investing in hedge funds doesn’t make sense, he told Alphaville — “we are not a hedge fund or trading company, we are a tech and AI company.” The venture money is spent on compute, he argues, to create models which flip the curve the right way up.
Trading, then, is only the first vertical. The durable IP is learning how to deploy AI that act in the real world — where, as Schmid puts it, the frontier labs’ models “just talk to people”. He gives the comparison of Google, which didn’t foresee building YouTube or Waymo — it first built a business that printed money, and the money bought the optionality. Most new AI labs, he says, are upside down. They research first, and revenue never. EquiLibre is trying to reverse that.
This is refreshingly candid. But the Google analogy can be seen another way — Google’s moonshots were funded by a cash business that compounded, while EquiLibre proposes to fund its lab from the one kind of cash business that depletes with use.
The trading record, however, is still doing the marketing. EquiLibre says that it has “a perfect record of zero negative months since inception”. But a quant fund should be judged on what happens when things go pear-shaped. Crowded trades suddenly unwinding or funding markets becoming choppy are typical environments that separate a real edge from a strong tailwind.
For example, the famous August 2007 “quant quake” hammered even systematic traders with years of positive returns. Some quickly recovered because they had built a durable, adaptable machine; others perished because they hadn’t.
EquiLibre’s zero negative months tell us nothing about this. Its inception was a 2025 rollout on crypto, which was later extended to equities, through a stretch in which most things went mostly up. An unbroken run of profitable months — deploying a possibly modest amount of dollars — in a rising tide is not evidence of resilience, but perhaps the absence of serious testing.
Now, back to the Kelly Criterion. The criterion doesn’t just tell you how much to bet, but what a bet is worth.
EquiLibre’s edge is worth the capital it can absorb before it degrades, multiplied by what it earns over the signal’s usually short life, discounted for the fact that the capital being deployed isn’t even EquiLibre’s — it’s Tower Research’s. Even being generous with all three and you probably get an edge worth significantly less than the $500mn valuation that Creandum’s investment round implied.
The confusion is possibly down to a mislabelling of EquiLibre. AI hasn’t suddenly turned trading into a software business. Quants have been using machine learning on markets since before it was even called machine learning, and long before the current AI craze. For a quant fund, AI is a tool that makes the researchers faster and the science better. Better tools don’t replace researchers; they raise the value of a good one. The constraint in quant is not time, it’s ideas.
The thing it cannot do is change the math that dictates every strategy that quant researchers create — that it is the edge that sets the bet, and the capacity to scale it that sets the ceiling.
A quant fund is not a tech company, and a tech company is not a quant fund. One is priced on what compounds, while the other lives and dies by the alpha that gets depleted away. Investors should be careful about confusing the two.
