Google DeepMind is using the MMORPG EVE Online as a new research environment for AI agents. On 6 May 2026, DeepMind and Fenris Creations (formerly CCP Games) announced a partnership under which DeepMind acquired a minority stake in the Icelandic game studio. The research builds on 15 years of game research, ranging from simple Atari titles through chess and Go to complex real-time strategy games.

From Atari Pixels to Persistent Worlds

DeepMind was founded in 2010 in part by Demis Hassabis, himself a former game developer. Video games served as a central test environment from the outset. In 2015, a small team published the Deep Q-Network (DQN) in Nature, a deep neural network that learned 49 Atari 2600 games directly from raw pixels – without game-specific engineering. The work catalysed the modern era of deep reinforcement learning.

In 2016, AlphaGo defeated world champion Lee Sedol at Go, an achievement experts thought was a decade away. Move 37, such an unexpected move that professional commentators initially mistook it for a mistake, overturned centuries-old wisdom in Go. Subsequent projects like AlphaGo Zero, which learned exclusively from self-generated play, and AlphaZero, which applied a single algorithm to chess, shogi and Go, generalised the approach. MuZero later learned to play without knowing the game rules.

In 2019, AlphaStar reached Grandmaster level in StarCraft II, navigating real-time complexity and incomplete information. The same foundations also flowed into AlphaFold, the protein-folding system, which was awarded the 2024 Nobel Prize in Chemistry.

Paradigm Shift: Agents Rather Than Score Hunters

The more recent research direction shifts focus from systems that master a single game with a clear score to agents that can move through arbitrary game worlds. SIMA (Scalable Instructable Multiworld Agent) perceives only what a player sees on screen and acts through ordinary keyboard and mouse controls. The agent has no access to game code and must infer the game state from screen output – the same constraint under which human players operate.

In November 2025, SIMA 2, with a Gemini model at its core, expanded the repertoire of the first version beyond its 600 language-based skills. DeepMind reports that SIMA 2 closes a significant portion of the gap to human task-completion performance in training environments, including games it was not trained on (such as ASKA and a Minecraft research implementation). SIMA 2 can reason about objectives, converse with users and improve through self-directed play. However, the agent still struggles with very long-term tasks and operates with limited memory (constrained by the context window for low latency). It is only available as a limited research preview to a small group of academics and game developers.

EVE Online as the "Final Boss" for AI

EVE Online is a persistent world that has been running continuously since 2003. The single-shard universe holds thousands of players in a shared economy with genuine supply and demand dynamics and player-built political structures. Speaking to Engadget, Hilmar Veigar Pétursson, CEO of Fenris Creations, said: "We jokingly say that the final boss for AI in games would obviously be EVE Online."

Unlike most previous research environments, EVE has no defined match, no clear objective and no point at which everything is reset. Research with DeepMind begins with an offline version of EVE Online running on local servers. The live Tranquility Universe is unaffected. The agents see only what a human sees, not the underlying data structures – the same observational constraint as in the Atari DQN work and SIMA.

Corporate Structure and Partnership

The DeepMind partnership coincided with a comprehensive transformation at CCP Games. On 30 April 2026, a management buyout from Pearl Abyss was completed. Hilmar Veigar Pétursson and the team regained ownership of the company after eight years under the South Korean publisher. Birgir Már Ragnarsson, CCP co-founder, assumed the position of Board Chair.

On 6 May 2026, the now-independent Icelandic studio announced a rebrand to Fenris Creations – a callback to CCP's first published game from 1997. On the same day, DeepMind's minority stake and research partnership were announced. The companies plan to explore AI-powered gameplay experiences.

Further Partnerships

Alongside Fenris Creations, DeepMind maintains deep partnerships with Hello Games, Coffee Stain Studios and Foulball Hangover. The EVE Online partnership is regarded as the next logical step in DeepMind's historical research trajectory: Atari games → chess and Go → EVE Online.