loading...
PROVENUndisclosed (Series A, €438M valuation) · 6 Jul 2026 · 4 MIN
Zero Negative Months: The Poker AI That Moved to Wall Street
Three ex-DeepMind researchers applied their poker-beating reinforcement learning to quant trading — zero negative months since launching on crypto in 2025.
EQUILIBRE TECHNOLOGIESequilibretechnologies.com ↗
LATEST ROUNDSeries A, €438M valuation, June 2026
TOTAL RAISEDUndisclosed (Series A, €438M valuation) · 1 rounds
FOUNDED2021
OPPORTUNITY SNAPSHOTadapt model
ENTRY ANGLES
Build RL-based bidding agents for programmatic advertising DSPs — one brand, one video format, 4–8 week controlled A/B test against existing strategy · Apply incomplete-information game theory to B2B procurement auctions where competitor bidding patterns directly affect clearing prices · License RL architecture to mid-tier quant shops as an alternative to statistical factor models that break on regime changes
VERTICALS
Quantitative FinanceProgrammatic AdvertisingB2B Procurement
CAPABILITIES
Reinforcement learning engineering, Real-time market data integration, Backtesting infrastructure, Risk management systems
01 /The Concept
In 2017, three DeepMind researchers published a paper that changed how AI played poker. DeepStack was the first program to beat professional human players at no-limit Texas Hold'em. Unlike prior poker AIs that precomputed strategy against memorized hands, DeepStack used reinforcement learning to reason about hidden information in real time – updating its model of the game state with every new card, against opponents it had never encountered before. Nine years later, Martin Schmid, Rudolf Kadlec, and Matej Moravcik are running a hedge fund.
EquiLibre Technologies, founded in Prague in December 2021, applies the same reinforcement learning architecture to financial markets. Through a partnership with quant firm Tower Research Capital, its agents trade billions of dollars per month on the S&P 500 and NASDAQ. The system has been live since early 2025 and has recorded zero negative months. In late June 2026, Creandum led a Series A at a €438 million valuation – the firm's single largest investment in one company, ever. The amount raised was not disclosed.
EquiLibre's core assertion is that poker and trading share structural properties, not just metaphorical ones. Both involve incomplete information about what other participants will do. Both require sequential decisions whose outcomes compound. Both operate in environments where other agents adapt to your behavior – which means historical records of what worked in the past are an imperfect guide to what will work in the future.
02 /Why It Matters
Quantitative trading has been an AI problem for thirty years, but the dominant approach has been statistical: find correlations in historical price data that predicted past returns, and bet on their continuation. The limitation is well known. Correlations hold until market regimes change. The 2022 rate shock – fifty years of correlation between bond prices and equity valuations breaking inside twelve months – is the most recent systematic example. Statistical quant models trained on the preceding decade's data had no representation of that regime change, and their performance reflected it.
Reinforcement learning does not predict from history. It learns to make decisions under uncertainty, updating its policy as the environment changes. DeepStack didn't beat professional players by memorizing how those players had bet in past games. It beat them by reasoning about what they were likely to do next, given incomplete information about their current hands. Applied to markets, EquiLibre's agents update their behavior in response to current conditions rather than historical correlations – which means a regime change is a new environment to adapt to rather than an anomaly that breaks the model.
Zero negative months across crypto and equities in live trading, through 2025 and 2026, is not a backtested metric. It is a claim about a live system performing across multiple market environments. The Tower Research Capital partnership, and Creandum's largest single investment, are independent validations of a track record the company has been careful not to publicize.
03 /Opportunities
The model EquiLibre demonstrates has almost no applications outside financial markets right now, despite the fact that the underlying architecture addresses a class of problem that appears in many industries.
Programmatic advertising auction markets have the same structural properties as trading. Each bid affects who wins the current impression. That outcome affects how competitors adjust their bids in the next auction. The clearing price for future inventory shifts based on aggregate bidding behavior. Historical CPM data captures past outcomes but not the game-theoretic dynamics that determine current outcomes when another major bidder changes strategy. Current programmatic bidding systems optimize against historical average CPMs using machine learning that cannot adapt to real-time strategy shifts from competitors.
An RL agent trained on auction market dynamics learns to compete rather than to predict. The training environment is available: real-time auction data from any DSP with API access. The reward signal is defined: margin-adjusted ROAS or CPM efficiency against a benchmark. The pilot structure is tractable: one brand, one ad format (programmatic video, where CPMs are high enough for the efficiency gain to be material), and a controlled A/B test against the existing bidding strategy over four to eight weeks. The team that runs that test and publishes the comparison has the case study that opens every major advertiser's door.
Valuation€438M / ~$500M
Series A investorCreandum (firm's largest ever single investment)
Trading partnerTower Research Capital
Markets tradedS&P 500, NASDAQ, crypto
Live sinceEarly 2025 (crypto), 2026 (equities)
Negative monthsZero since launch
Origin technologyDeepStack – first AI to beat professional poker players
FoundedDecember 2021, Prague
SEE ALSO · 2
RELATED DRILLS · 6