TL;DR
Offline Testing: Google DeepMind and Fenris Creations said the first EVE experiments will stay off the live Tranquility server. Deal Structure: Fenris said the studio’s separation from Pearl Abyss was valued at $120 million, while Google also took a minority stake. Player Stake: EVE’s economy, alliances, and wars give researchers a resettable model-testing environment without putting the live game at risk.
Google DeepMind and Fenris Creations say their player-driven systems research will use controlled offline versions of EVE, turning the MMO’s markets, alliances, and wars into an AI test bed while keeping Tranquility outside the first experiments.
Controlled offline copies of EVE, not the live Tranquility server, are expected to host the first tests. For players, keeping the work off the public game separates the current partnership from any live feature rollout and gives the companies room to probe model behavior without asking subscribers to absorb the risk.
Against that backdrop, Fenris pairs the research move with a wider studio reset. Under the new name, the EVE developer goes independent.
Company materials citing regulatory filings put the separation from Pearl Abyss at $120 million in cash and non-cash consideration.
Google also invested in Fenris Creations as part of the same transition package.
That minority stake gives the relationship a financial dimension without turning the studio into a captive research arm.
Fenris also said the transition does not involve restructuring or layoffs. Keeping the existing teams in place supports the narrower promise around the deal: EVE’s live operation continues while model experiments run in a separate lane.
Why DeepMind wants EVE’s offline sandbox
Fenris and DeepMind are positioning the work as a research partnership. EVE’s market, alliance, and conflict systems only work as a controlled AI environment if the partners can pressure-test them without exposing the live shard to unstable model behavior.
Fenris used unusually direct language when it drew that line.
“To be very clear: this initial research will take place in controlled, offline versions of EVE that are not connected to Tranquility. But it does open a door to work that feels very true to EVE: difficult problems, long timelines, strange possibilities, and people willing to explore what comes next.”
Fenris Creations, company announcement
An isolated EVE server gives DeepMind a more useful testing loop than a one-off live event would. Researchers can test and evaluate models against a market shock, an alliance rupture, or a resource squeeze, then reset the same environment and compare how the models respond on a second or third pass.
Within that sandbox, many of the interesting questions in EVE are not short tactical puzzles. Scarcity, diplomacy, retaliation, and coalition behavior play out over time, so a resettable copy lets researchers isolate one condition, rerun it, and see whether a model adapts or simply memorizes a path.
Google DeepMind will work with an offline version of EVE Online running on a local server to test and evaluate models in a controlled setting. Live-server safeguards matter here: unstable negotiation, exploit-seeking behavior, or poor long-horizon planning can be observed without spilling into the world that active subscribers are paying to inhabit.
Long-horizon planning, memory, and continual learning are central to the research agenda Fenris outlined for the partnership. EVE gives those ideas a harsher proving ground than a lab benchmark because the same world can combine scarcity, delayed consequences, rival factions, and noisy information. In a resettable copy, researchers can rerun one disruption after another and see whether a model remembers prior conditions, changes strategy over time, or falls into the same mistake again.
No player-facing product, release schedule, or deployment lane has been attached to the current work. EVE remains the setting for the tests, not the first place where users are meant to encounter the results.
An MMO with entrenched corporations, trade routes, and wars creates more believable stress than a polished demo. Complexity on that scale would become a liability if the first model runs could alter the live economy or push costs onto players who did not opt into the experiment.
Offline separation also protects the evidence quality of the research. A live-service world is full of hidden variables. A bounded copy gives the teams a cleaner way to compare one run with another and decide whether a result came from the model, the scenario, or the surrounding game state.
Fenris also gains time from that setup. Teams can watch how models handle repeated conflicts, market imbalances, and resource pressure before deciding whether any part of the work belongs near a player-facing product discussion. Company planners also get a cleaner read on which scenarios generate useful evidence and which ones only create noise.
The rebrand, buyout, and Google stake
CCP Games has rebranded itself as Fenris Creations, closing a studio identity that had been tied to EVE for years.
Under the new ownership structure, Fenris has also returned to independent operation after separating from Pearl Abyss. Combined with Google’s investment, that shift puts new capital and new accountability around the same game world that will host the offline experiments.
Google’s minority position changes the business reading of the story. Google is not only borrowing a game world for model evaluation, it now has money inside the studio running that world.
Fenris used the official valuation to put scale on the break. A separation measured in nine figures tells readers that the ownership change itself is a material transaction rather than a cosmetic rename.
Operating continuity is another part of the pitch. EVE only works as a research environment if the same live-service teams keep the world stable while the offline work proceeds beside it. For Fenris, the no-layoff language supports the business case for keeping the live game intact while the research lane develops.
Pearl Abyss acquired CCP Games in 2018, so the 2026 split closes an ownership chapter that shaped the studio for much of the past decade.
Why EVE is different from a simple game demo
EVE offers a kind of complexity that is hard to stage on demand. Inside the MMO, player memory, economic depth, and political friction make long-horizon agent behavior worth testing.
Its economy, wars, and emergent narratives give researchers a system where incentives collide across long stretches rather than in a single scripted match. Planning, adaptation, and multi-actor conflict can be compared under recurring conditions instead of under a short puzzle loop.
An offline copy of the game changes the technical risk profile at the same time. Researchers can reset the environment, inject a targeted scenario, and measure how models react to the same shortage or conflict more than once without leaving permanent damage behind for players. That turns EVE from a famous MMO into a repeatable instrument.
Against the 2026 partnership push, still-CEO Hilmar Veigar Pétursson’s 2025 performance remarks give the studio a steadier operating backdrop, even if they do not validate the strategy by themselves. He described 2025 as “strongest results in years”, including a record-breaking November and one of EVE Online’s stronger quarters, which suggests Fenris is entering the arrangement from relative stability.
Results from 2025 also matter because they suggest the studio did not need to rush a risky live deployment to create a new story for investors. A healthier operating backdrop gives Fenris more freedom to keep the research lane narrow, separate, and patient.
Earlier EVE research and the next checkpoint
Before this deal, outside researchers were already using New Eden for work that sat somewhere between science collaboration and game experimentation. As a 2017 EVE Online science project, that earlier effort remains background, not proof for the DeepMind partnership, but it shows that the present test-bed framing did not appear out of nowhere.
Readers already have three hard facts to work with: the first experiments stay off Tranquility, Google has capital inside Fenris, and the ownership reset is under way. More practical unknowns remain: what model classes are involved, what safeguards might become public, and whether any player-facing boundary will ever move closer to the live game.
At EVE Fanfest 2026, which is scheduled to begin on May 14, Fenris faces a narrower task: naming the safeguard that keeps DeepMind’s model work off Tranquility. Clearer detail on model evaluation, local servers, and live-game boundaries would give players a better way to judge whether the partnership is staying inside the limits it has now laid out.