Google DeepMind Takes Minority Stake in EVE Online Studio to Train AI on 20 Years of Player Data

Google DeepMind has taken a minority stake in Fenris Creations, the studio formerly known as CCP Games, folding the investment into a $120 million management buyout that separated the Icelandic developer from Korean publisher Pearl Abyss. The deal turns EVE Online, a space MMORPG running continuously since 2003, into a research environment for studying how AI handles decisions that play out over years, not seconds.

Fenris Creations CEO Hilmar Veigar Pétursson described the arrangement as a study of intelligence within dynamic, player-driven ecosystems, according to Crypto Briefing. DeepMind's stake is a minority position, meaning the studio keeps creative control of the game. It's a financing and research deal, not an acquisition.

Why EVE Online

EVE Online runs a single-shard universe spanning more than 7,000 star systems, according to both Crypto Briefing and Cryptonomist. Real players have been trading, warring, and forming alliances inside it for over 20 years, generating a behavioral dataset that's hard to fake and expensive to build from scratch.

Unlike a lab experiment with fixed variables, EVE's economy evolved organically out of thousands of players pursuing competing strategies over two decades, Cryptonomist reported. DeepMind wants to know if AI agents can operate inside that kind of mess the way a human does.

The actual work won't touch live servers. DeepMind and Fenris Creations plan to run AI models inside an offline version of EVE Online on local servers, a sandbox where researchers can watch how models handle long-horizon planning, memory, and continual learning without risking anything for actual players, according to Crypto Briefing.

Fifteen years of game research behind it

DeepMind published a retrospective on August 21, 2026, tracing the lineage that led to this partnership, according to Unite.AI. It starts with the Deep Q-Network, which learned 49 Atari 2600 games from raw pixels and was documented in a 2015 Nature paper. From there: AlphaGo beat world champion Lee Sedol in 2016, AlphaGo Zero and AlphaZero dropped human training data entirely and generalized across chess, shogi, and Go, MuZero learned to play without even knowing the rules, and AlphaStar hit Grandmaster level in StarCraft II in 2019. DeepMind CEO Demis Hassabis has pointed to that track record as evidence gaming is a legitimate proving ground for AI, not a sideshow.

That same research line, per DeepMind's own account, fed into AlphaFold, the protein-structure system that won the 2024 Nobel Prize in Chemistry. DeepMind has a real history of turning game research into scientific payoff.

The current frontier is SIMA, the Scalable Instructable Multiworld Agent, which watches a screen like a human player and acts through ordinary keyboard and mouse controls, with zero access to underlying game code. SIMA 2, released in November 2025 with a Gemini model at its core, moved past a repertoire of over 600 language-following skills into reasoning about goals and improving through self-play, according to Unite.AI. DeepMind says SIMA 2 closes a meaningful chunk of the gap to human performance even in games it was never trained on.

DeepMind is also upfront about the limits. SIMA 2 still struggles with very long-horizon tasks and works off a short memory constrained by the context window needed for real-time interaction. It's available only as a limited research preview to a small group of academics and game developers, not the public.

Four specific things DeepMind wants to test

Unite.AI's report lists the capabilities DeepMind expects the EVE environment to exercise: continual learning without forgetting old skills, memory that extends beyond the model's current context window, and reasoning across timeframes that stretch into years rather than minutes or hours, matching how EVE's political alliances and market cycles actually unfold.

Most AI systems today, including large language models, lose track of context past a certain window and don't retain lessons from one session to the next without being explicitly retrained. Successfully deploying an agent that plans and remembers across an EVE Online timeline the way a veteran player does would represent a meaningful technical advance.

What's not happening

DeepMind has been explicit, per Crypto Briefing, that there's no AI-powered game feature shipping next quarter. This is foundational research. Players logging into EVE Online today won't notice anything different, because the testing happens on an offline, local-server version of the game, not the live universe.

The open question is what DeepMind actually does with what it learns. The lab hasn't said whether findings from the EVE research will feed into consumer products, other Gemini-based agents, or stay purely academic. Fenris Creations, for its part, gets a research-focused investor and continued independence from Pearl Abyss, a studio structure it only regained through the $120 million buyout that made this whole partnership possible in the first place.