Google’s premier AI lab is buying into one of gaming’s most infamously complex virtual universes, betting that the chaos of EVE Online can teach machines how to think long-term in worlds that never sit still.

A decade-long courtship between AI and games

Long before this deal, Google DeepMind was using games as a laboratory. From mastering Go to tearing through Atari classics and outplaying pros in StarCraft II, the company has consistently treated virtual worlds as “the perfect proving ground for AI,” as CEO Demis Hassabis has put it.¹

That philosophy set the stage for a more ambitious target: not a neatly bounded board game or a symmetric RTS, but a sprawling, player-driven MMO where politics, betrayal, and emergent economics are the real endgame.

May 6, 2026: EVE’s makers buy freedom, DeepMind buys in

On May 6, 2026, the studio behind EVE Online quietly pulled off two transformations at once.

First, CCP Games’ management bought the company out from former owner Pearl Abyss for $120 million, regaining independence and rebranding as Fenris Creations. The newly independent entity “will continue to operate as normal without any restructuring or layoffs,” according to the company.¹

Second, Google’s DeepMind division took a minority stake in the freshly rebranded studio. DeepMind is investing “in the millions” of dollars, Fenris CEO Hilmar Veigar Pétursson told Bloomberg, in a deal framed not as a simple cash injection but as a formal “research partnership.”²

The structure is blunt: DeepMind gets equity and a bespoke AI testbed; Fenris gets capital, prestige, and direct access to one of the world’s leading AI research groups.

Building an AI lab inside New Eden

At the core of the partnership is a custom, closed version of EVE. DeepMind will “work with an offline version of EVE Online running on a local server to test and evaluate models in a controlled setting,” Fenris has said.² Ars Technica describes this instance as a “specially designed offline version of the game running on a local server, without directly impacting the experience for online players.”¹

In other words: no AI-driven market crash is going to vaporize your hard-earned ISK—at least not on Tranquility.

DeepMind and Fenris describe EVE as “a uniquely rich environment for study,” particularly for AI systems that need “long-horizon planning, memory, and continual learning.”¹ That’s a polite way of saying that New Eden is a nightmare of interlocking systems where decisions echo weeks or months later.

DeepMind isn’t just there to run basic pathfinding experiments. The lab wants to see how its models handle “complex, dynamic, player-driven systems” that behave more like messy real economies than clean simulations.¹ Recent DeepMind work has leaned on “virtual world” models as stepping stones for AI agents meant to operate in physical reality; EVE is now the latest, and arguably wildest, of those sandboxes.¹

The fanboy CEO and the “extraordinary” game

Hassabis has never been shy about his love of games, and this deal reads as equal parts scientific ambition and gamer wish-fulfillment. “I’ve always been passionate about games and they’ve played a big part in @GoogleDeepMind’s history, as the perfect proving ground for AI,” he wrote, announcing the partnership publicly. “Thrilled to announce this research partnership with @FenrisCreations - @EveOnline is one of the most extraordinary games ever built and has...”@demishassabis on X3. I've always been passionate about games and they've played a big part in @GoogleDeepMind’s history, as the perfect proving ground for AI. Thrilled to announce this research partnership with @FenrisCreations - @EveOnline is one of the most extraordinary games ever built and has...³

Inside Fenris, the framing is even more poetic. In an open letter to players, Pétursson argued that “EVE is one of the few environments where questions about intelligence can be explored inside something that already behaves like a living world.” Studying EVE, he said, will let DeepMind’s models tackle “difficult problems, long timelines [and] strange possibilities.”¹

It’s not subtle: both sides are pitching EVE as less a game and more a synthetic civilization — one that can be probed without risking real-world damage.

Not just lab rats: promises of new gameplay

The partnership isn’t framed purely as research. Both companies say they “will also explore new gameplay experiences enabled by these technologies.”¹²

That phrase is doing a lot of work. On the optimistic reading, it hints at smarter NPCs, richer economies, and dynamic content-generation tuned to the game’s famously hardcore sandbox. Imagine AI agents that can participate meaningfully in corp politics or industrial logistics without feeling like lifeless bots.

On the more cynical reading, it raises familiar alarm bells: will EVE players become unpaid test subjects for monetizable AI features? Will in-game behaviors end up feeding training pipelines that benefit Alphabet far more than the community that generated the data?

Fenris and DeepMind stress that experimentation will happen in a sealed, offline shard, suggesting a hard separation between player reality and research sandbox.¹² But the moment “new gameplay experiences” are on the table, the wall between lab and live service starts to look semi-permeable by design.

A timeline of entangled incentives

Prior decade: DeepMind builds its reputation on game benchmarks — Go, Atari, StarCraft II — proving its reinforcement learning models can smash human high scores and defeat pros.

Lead-up to 2026: The lab shifts focus to “virtual world” training grounds as proxies for real-world complexity, seeking environments where agents must coordinate, negotiate, and plan across long time scales.¹

Early 2026: CCP management negotiates its way out of Pearl Abyss, lining up the $120 million needed to reclaim independence.¹

May 6, 2026: The rebrand to Fenris Creations is unveiled alongside assurances of no layoffs or internal restructuring, signaling continuity for EVE’s notoriously change-averse community.¹

The same day, DeepMind announces its minority stake — “in the millions” — and the research partnership centered on an offline EVE instance used to “test and evaluate models in a controlled setting.”² Hassabis amplifies the news on X, calling EVE “one of the most extraordinary games ever built.”³

From that point, EVE’s development roadmap and DeepMind’s research agenda are formally linked. The lab gains a complex, persistent, multiplayer environment. Fenris gains cash, prestige, and a new axis along which to pitch EVE’s future.

All the competing narratives

DeepMind’s angle: EVE as a stress test for real-world AI

For DeepMind, this looks like the next logical testbed: if agents can navigate a snarled web of alliances, markets, and warfare in EVE, maybe they can handle less contrived, real-world decision spaces.

The language about “long-horizon planning, memory, and continual learning” is straight out of cutting-edge AI research.¹ EVE’s player-driven systems supply precisely the kind of non-stationary environment current models struggle with.

Fenris’s angle: independence with a flagship partner

For Fenris, the story is survival and swagger. Buying independence for $120 million is a bold move; pairing that announcement with a multi-million-dollar minority investment from DeepMind tells investors and players alike that this isn’t a studio limping out of a publisher, but one graduating to a new league.¹²

The promise of “no restructuring or layoffs” is a clear message to a community used to seeing buyouts followed by cuts.¹ The rebrand to Fenris Creations, combined with an AI heavyweight as partner, signals a long-term bet on EVE as both game and research artifact.

Players’ angle: intrigued, wary, and very online

EVE’s community has built a culture around emergent complexity; many will instinctively like the idea that their universe is now a frontier for AI research instead of just another live service.

But this is also a community famous for paranoia and metagaming. Even with the offline server firewall, some will worry about:

• Data use: How much historical player behavior informs DeepMind’s models?
• Balance and opacity: Will AI-informed tweaks make the sandbox feel more manipulated?
• Ownership: Who benefits if AI systems trained on EVE data are later deployed in finance, logistics, or defense?

Those questions don’t have satisfying answers yet, and neither Fenris nor DeepMind has been eager to spell out the long-term IP and data-sharing contours in public.

The bigger picture: when virtual worlds become infrastructure

Strip away the branding and this deal is about something larger: live service games graduating into infrastructure for AI research. EVE is now not just entertainment, but a mid-level rung in a ladder that runs from Atari ROMs to robots in warehouses.

If the experiment works, expect more of this: equity-for-access deals where AI labs plug into rich, persistent online worlds to harvest exactly the kind of emergent, adversarial complexity their models desperately need.

EVE has always been a place where players test each other’s intelligence and nerve. Now, quietly and in an offline mirror universe, it’s about to become a place where some of the world’s most advanced AI systems are tested too.