EVE Online Has an AI Helper. Here Is What DeepMind Is Actually Building.
On May 6, the studio behind EVE Online — the 23-year-old space MMO famous for player wars that reshape its in-game economy — became an independent company called Fenris Creations and simultaneously announced that Google DeepMind had taken a minority stake in it (Bloomberg; Fenris Creations press release). The financial terms of DeepMind's stake were not disclosed. What is known: the Reykjavik-based studio had just spent $120 million buying back its independence from Korean publisher Pearl Abyss — a figure confirmed in Pearl Abyss regulatory filings. That $120 million was Fenris's buyback cost, not DeepMind's investment amount.
The framing from most outlets: DeepMind is using EVE Online as a new AI training ground. That's true, as far as it goes. But there's a wrinkle nobody else is talking about. EVE Online already has an AI feature. We found it, downloaded it, and spent time with it. What it does — and what it conspicuously does not do — tells you exactly how far DeepMind's ambitions extend beyond what the studio has already shipped.
In February 2026, Fenris launched an AI-powered assistance feature for EVE Online, trained on more than 5.8 million messages posted in the game's Rookie Help channel (PC Gamer). The feature is designed to answer new player questions without producing game content, creating art, or replacing creative work — a narrow, carefully scoped task. It is not designed to plan across months, remember interactions over time, or learn continuously from gameplay. It answers questions.
The partnership announcement describes a research agenda centered on "long-horizon planning, memory, and continual learning" — using an offline version of EVE Online running on a local server to "test and evaluate models in a controlled setting" (Ars Technica). DeepMind CEO Demis Hassabis described games as having been "at the heart of many of Google DeepMind's breakthroughs — like Atari DQN, AlphaGo, AlphaStar and SIMA — because they're the perfect training ground for developing and testing AI algorithms."
The stated ambition is to study "intelligence in complex, dynamic, player-driven systems." Hilmar Veigar Petursson, Fenris CEO, put it this way: "EVE is one of the few environments where questions about intelligence can be explored inside something that already behaves like a living world" (Ars Technica).
That is a genuinely different aspiration from a chatbot that answers new-player questions.
EVE Online's appeal as a research environment is the game's economy — a player-driven market spanning 20 years of transactions, scams, political manipulation, and strategic alliances. If you want to study how AI systems behave in environments where humans scheme, hoard, betray, and cooperate at scale, EVE has more documented history than any lab could synthesize.
The offline test server is key: DeepMind will run experiments without touching the live game, which is currently performing at record levels — November 2025 was the best month in the game's history, and Q4 was the second-best quarter in its more than 20-year run (Fenris Creations press release). Fenris is profitable, with over $70 million in 2025 revenue. This is not a studio DeepMind is rescuing. It is one that bought its own freedom and then negotiated a research partnership on its own terms.
The AI industry's hunger for better training environments is well-documented. The highest-quality internet data is increasingly locked behind paywalls or exhausted. Synthetic data helps but lacks the adversarial richness of real human behavior over time. Game worlds — particularly ones with decades of documented player behavior — are a candidate solution.
If the EVE Offline testbed produces publishable results demonstrating that AI systems can improve at long-horizon planning through persistent virtual environment training, expect other labs to seek similar arrangements with other studios. The MMO-as-research-infrastructure model would become a new category of AI partnership. More details will be shared at EVE Fanfest 2026, starting May 14.
The existing assistance feature — scoped, careful, limited — is the baseline. What DeepMind is describing is something substantially harder. Whether it ships, and what it looks like when it does, is the actual story.