Poker software runs calculations that would take a human brain centuries to complete. The machines do this in seconds. Artificial intelligence systems have beaten the best players on the planet, and the technology behind those victories now trickles down into training tools, security systems, and gameplay itself. Online poker in 2025 looks nothing like it did ten years ago, and the next decade will bring changes that are harder to predict.
How AI Learned to Beat Professionals
Carnegie Mellon University built Libratus using more than 15 million core hours of computation on the Bridges supercomputer at the Pittsburgh Supercomputing Center. The system played heads-up no-limit Texas Hold’em against professional players in 2017 and finished $1,766,250 ahead in chips. Its win rate of 14.7 big blinds per 100 hands is considered extremely high and statistically meaningful.
The technique behind Libratus is called counterfactual regret minimization. Oskari Tammelin introduced the CFR+ variant in 2014, and Carnegie Mellon’s team applied it at scale. The AI does not rely on reading opponents or picking up tells. It calculates the mathematically sound response to any situation, adjusting its strategy over billions of simulated hands.
Two years later, Carnegie Mellon partnered with Facebook AI Research to build Pluribus, a system capable of playing six-player no-limit Hold’em. This was a harder problem. With more opponents, the number of possible game states grows exponentially.
Pluribus averaged $5 per hand against professionals and earned roughly $1,000 per hour during testing. The program trained in eight days on a 64-core server, costing $144 in cloud computing resources. The developers refused to release the source code because the system could be used to cheat. Libratus received the Marvin Minsky Medal for Outstanding Achievements in AI. Pluribus appeared on the cover of Science Magazine.
Decentralized Shuffling and Cryptographic Verification
When players play online poker on blockchain-based platforms, the shuffle process becomes distributed across all participants at the table. CoinPoker applies KECCAK-256 cryptography, the same protocol securing the Ethereum network, to ensure no single entity controls the randomization. Each seat contributes to shuffling a portion of the deck. After every hand, players can verify the exact sequence of cards dealt. CoinPoker has offered one million CHP tokens to anyone who can find flaws in their RNG system, and that bounty remains unclaimed.
Training Tools Built on Solver Technology
The AI breakthroughs at research universities have filtered into consumer products. GTO Wizard generates optimal strategies for games up to 200 big blinds deep with customizable bet sizes. The system produces solutions in an average of 3 seconds per street. When tested against Slumbot, a strong abstraction-based poker AI, GTO Wizard won at a rate of 19.4 big blinds per 100 hands over 150,000 hands.
In April 2025, GTO Wizard switched from Nash Equilibrium calculations to Quantal Response Equilibrium. Average flop exploitability dropped from 0.17% to 0.12% of the pot, a 25% improvement. The company now works directly with poker networks to detect cheating using game theory analysis.
Browser-based tools like GTO Wizard charge between $30 and $100 per month. Desktop solvers such as PioSolver and MonkerSolver sell lifetime licenses for $250 to $1,000.
The Deep Learning Approach
The University of Alberta’s Computer Poker Research Group took a different path. Their system, DeepStack, defeated professionals in December 2016. Science published the findings.
DeepStack does not pre-compute a fixed strategy. It recalculates its approach at each decision point using neural networks trained through deep learning. The system runs on a standard gaming laptop with an Nvidia graphics card, making decisions in about 3 seconds.
Heads-up no-limit Hold’em contains more possible situations than there are atoms in the observable universe. DeepStack handles this by building intuition through training rather than storing every scenario in memory.
Large Language Models at the Table
Researchers have started testing large language models on poker problems. PokerBench, accepted to the AAAI 2025 conference, contains 11,000 scenarios covering pre-flop and post-flop play. Trained poker players helped develop the benchmark.
GPT-4 scored 53.55% accuracy, the highest among the tested models. This is poor performance compared to other benchmarks where large language models excel. After fine-tuning on poker data, some models reached 78.26% accuracy.
PokerGPT addresses a different problem. Traditional CFR calculations become computationally expensive in multiplayer games because the game tree grows exponentially. PokerGPT attempts to provide end-to-end solutions without running full CFR iterations.
Virtual Reality Tables
PokerStars developed Vegas Infinite with Lucky VR. The platform connects PlayStation VR2, Meta Quest, and Steam VR users in the same virtual poker rooms. PC players without headsets can also join.
Quest Pro support adds face and eye tracking, allowing avatars to display player expressions. Advanced haptics simulate the feel of chips and cards. Mixed reality passthrough lets players see their physical surroundings while seated at a virtual table.
The technology combines elements of live and online poker. Players can read body language and interact verbally while benefiting from the convenience of playing from home.
Security and Cheating Detection
PokerStars employs a 50-person Game Integrity Team that reviews flagged accounts alongside automated systems. The combination achieves a 95% proactive detection rate, meaning most cheaters are caught before other players report them.
Machine learning models look for patterns that humans would miss. A player whose timing is too consistent, whose bet sizing follows predictable formulas, or whose results are statistically improbable gets flagged. Bots that try to mimic human randomness still leave traces.
AI links seemingly unrelated accounts by analyzing device fingerprints, IP behavior, play style, and timing data. A single person running multiple accounts often reveals connections through these markers.
888poker refunded $250,000 to affected players in 2024 after identifying bots and real-time assistance programs.
Market Size and Platform Growth
The online poker market was estimated at $3.86 billion in 2024. Projections from Grand View Research suggest it will reach $6.90 billion by 2030, growing at 10.2% annually. Texas Hold’em holds over 62% market share.
Other analysts offer higher estimates. Some project growth from $7.98 billion in 2024 to $37.19 billion by 2030, a compound annual growth rate of 29.24%.
More than 500 active platforms serve over 100 million players worldwide. Mobile gaming accounts for 70% of all online poker traffic.
What Comes Next
The gap between AI capability and human performance continues to widen in controlled settings. But the real changes happen when these technologies become accessible to regular players. Training tools that cost hundreds of dollars a month will get cheaper. Security systems will become harder to fool. Virtual reality hardware will improve and drop in price.
Blockchain verification removes trust from the equation entirely. Players no longer need to assume a site deals fairly. They can check.
The research keeps moving forward. Large language models may eventually play competent poker without fine-tuning. New equilibrium concepts may replace Nash as the standard for solver calculations. The technology available to players in 2030 will make current tools look primitive.
