Google DeepMind Takes Stake in EVE Online's Studio, Rebranded Fenris Creations, in $120M AI Research Deal

When AI researchers talk about the capabilities that still separate today’s large language models from more general intelligence, three problems come up repeatedly: planning across long time horizons, maintaining and updating persistent memory, and learning continuously without forgetting previous knowledge. EVE Online, a space-based massively multiplayer game with a 23-year-old player-driven economy, turns out to be an unusually rich environment to study all three.

On May 6, 2026, the Icelandic studio that developed EVE Online made a pair of announcements that simultaneously ended one chapter of its history and began another. The company formerly known as CCP Games split from its South Korean parent Pearl Abyss, became a privately held independent company, adopted the new name Fenris Creations, and revealed that Google had made a minority equity investment as part of a dedicated research partnership with Google DeepMind. The total transaction value is $120 million, comprising both cash and non-cash consideration.

Why EVE Online?

EVE Online is, by any reasonable measure, one of the most complex persistent simulations ever built. Since its 2003 launch, the game has hosted a player-driven economy with its own supply and demand dynamics, a political system involving alliances of thousands of players, long-running military campaigns planned over months, financial instruments, espionage, and emergent social phenomena that have occasionally attracted coverage from academic economists and social scientists.

For AI researchers, that complexity translates into a specific kind of training and evaluation value that most environments lack. In a standard game benchmark — chess, Go, Atari, even StarCraft — agents operate under rules that are fully specified, time horizons are relatively short, and the agent’s goal is unambiguous. EVE Online offers something different: an open-ended environment where the “right” strategy depends on context established weeks or months earlier, where memory of past events is essential to rational action, and where novel situations arise continuously from the interactions of hundreds of thousands of players.

Google DeepMind describes EVE Online as “a one-of-a-kind simulation for testing general-purpose artificial intelligence in a safe sandbox environment.” The research partnership will focus specifically on three capability clusters: long-horizon planning (making optimal decisions when consequences unfold over days or weeks rather than milliseconds), persistent memory (retaining and retrieving relevant context across arbitrarily long episodes), and continual learning (acquiring new skills and knowledge without degrading existing capabilities — the so-called “catastrophic forgetting” problem).

These are not incremental improvements on existing systems. They are, arguably, the capabilities most directly relevant to the vision of AI agents that can work autonomously on complex, multi-week tasks — which is where the commercial frontier of AI deployment is currently headed.

Research Methodology

To study these capabilities in a controlled scientific environment, Google DeepMind will work with an offline version of EVE Online running on a local server, isolated from the live game and its 23 years of accumulated player culture. This approach allows the research team to run experiments at arbitrary speed, reset states, and vary parameters in ways that would be impossible in a live multiplayer environment.

The offline environment also addresses the ethical complexity of using an active player community as unwitting research subjects. By isolating the research infrastructure from the live game, Fenris Creations and Google DeepMind can conduct rigorous scientific work without disrupting the experience of existing players.

More details about the research agenda will be shared publicly at EVE Fanfest 2026, scheduled to begin May 14 in Reykjavik. Adrian Bolton, a member of Google DeepMind’s founding team, is confirmed to appear on stage — a signal of how seriously the organization views this collaboration’s research potential.

The Corporate Transformation

The Google DeepMind partnership was announced alongside a significant corporate restructuring that separated Fenris Creations from Pearl Abyss, the South Korean gaming conglomerate that had acquired CCP Games in 2018 for approximately $425 million. While the acquisition had provided financial stability during a turbulent period for the MMO genre, the separation appears to reflect a mutual agreement that Fenris Creations’ future lies in a direction — AI research collaboration and platform development — that diverges meaningfully from Pearl Abyss’s core gaming portfolio.

Fenris Creations retains the same leadership team, staff, and operational structure. The rebrand reflects a deliberate attempt to signal a new strategic identity: not just a game studio, but a research organization that happens to have built and maintained one of the world’s most sophisticated persistent simulations.

The company reported strong operational performance going into the transition — EVE Online recorded a record-breaking November 2025 and one of the strongest quarters in its more than 20-year history, providing financial confidence for the transition to independence.

Broader Context: Why DeepMind Is Here

Google DeepMind’s interest in gaming environments is not new. The organization’s early breakout work involved training AI on Atari games, and AlphaGo and AlphaStar later demonstrated superhuman performance in Go and StarCraft respectively. But those earlier projects were focused primarily on demonstrating competitive performance in well-defined strategic games.

The EVE Online collaboration represents a shift in research intent. Rather than optimizing within a defined ruleset, DeepMind is now interested in open-ended environments where the challenge is more similar to real-world tasks: understanding context accumulated over a long history, setting and pursuing goals that evolve over time, and reasoning about a world that is continuously changed by the actions of many other agents.

That shift mirrors the broader evolution of AI research priorities. As frontier models become increasingly capable at in-context tasks, the research community is investing more effort in capabilities that don’t emerge naturally from scale alone — specifically, the kind of temporal depth and memory that distinguishes a general-purpose AI assistant from a capable but episodic one.

For Fenris Creations, the partnership offers something beyond capital: association with one of the world’s top AI research organizations, access to techniques that could directly improve EVE Online’s own AI-driven systems, and a strategic identity that positions the company at the intersection of two industries rather than in the increasingly difficult center of one.

Implications for AI Research

The collaboration raises intriguing questions about the future role of complex simulations in AI development. If EVE Online proves useful for advancing long-horizon planning and memory research, it could inspire a broader category of “complex simulation environments” as research tools — not games designed for fun, but persistent worlds designed to exhibit the properties most relevant to understanding and developing general-purpose AI.

That would be a meaningful development for the field. One of the chronic criticisms of current AI benchmarks is that they are too narrow, too well-specified, and too disconnected from the messiness of real-world tasks to adequately measure genuine capability progress. An environment that has generated genuine emergent complexity over two decades of player interaction is, almost by definition, something different.