Self-Hosted AI Deployment in Stockholm, Sweden

We deployed a self-hosted AI platform over 10 weeks — integrated with the group's game server infrastructure across Stockholm data centres. (1) Gaming-optimised GPU infrastructure. Hardware: 48 NVIDIA H100 80GB GPUs across 6 server nodes, deployed across the group's two Stockholm-area game server facilities (facilities already hosting game servers for European players — co-located AI reducing network latency to zero for AI-assisted gameplay decisions). Game operations cluster: 32 GPUs (4 nodes) handling real-time inference — matchmaking, anti-cheat, dynamic difficulty, and monetisation decisions. Sub-100ms latency requirement met through co-location with game servers. Development and training cluster: 16 GPUs (2 nodes) for model training, fine-tuning, and content generation. Separated from production to prevent training workloads from affecting real-time game experience. Network: InfiniBand between GPU nodes. Connected to game server clusters via dedicated 100Gbps Ethernet. Game-to-AI inference path: under 5ms network latency (co-located infrastructure).

Storage: 60TB NVMe for model weights and active game data. Connected to the group's game data lake (2.8PB — player behaviour data, game telemetry, asset libraries). Auto-scaling: inference workload scaling with player population — GPU allocation automatically adjusting between games based on concurrent player counts (Kubernetes-based orchestration with NVIDIA GPU Operator). (2) Game economy and monetisation AI. Economy management: Llama 3 70B fine-tuned on the group's game economy data — learning the relationship between economy parameters (currency generation rates, item drop rates, marketplace pricing, crafting costs) and player behaviour (engagement, spending, progression, satisfaction). The model monitoring live game economies in real time: detecting inflation (currency accumulating faster than intended — indicating exploit or design flaw), identifying economy imbalances (specific items becoming too common or too rare — disrupting game balance), and recommending parameter adjustments (economy tuning suggestions for game designers, presented through an internal dashboard).

Economy AI reducing emergency economy interventions from 14 per quarter to 3 (developer team intervening only for AI-flagged critical issues rather than reacting to player complaints). Monetisation optimisation: models recommending: which items to feature in the in-game store for each player segment (based on play style, progression stage, and historical purchase patterns), optimal pricing (dynamic pricing within ranges approved by the game design team — not gouging but matching willingness-to-pay), and timing (when during a play session to present offers — after positive experiences, not during frustrating moments). Monetisation AI increasing per-player revenue 18 percent while maintaining player satisfaction scores (measured through in-game surveys and engagement metrics) — demonstrating that well-targeted offers improved rather than degraded player experience. (3) Anti-cheat AI: protecting competitive integrity.

Behavioural analysis: Mistral 7B fine-tuned on the group's cheat detection dataset (84,000 confirmed cheat cases with behavioural profiles) — detecting: aim assistance (statistical analysis of aiming patterns — accuracy, snap-to-target speed, reaction time distributions that differed from human capabilities), speed and movement manipulation (movement patterns exceeding normal parameters — wall clipping, speed increases, flight), economy exploitation (resource generation patterns indicating automated farming, duplication exploits, or marketplace manipulation), and account sharing/selling (behavioural pattern changes indicating different humans using the same account). Real-time processing: analysing player behaviour during gameplay at sub-50ms latency. The model scoring every active player continuously, flagging suspicious behaviour for investigation. Key advantage of self-hosted: the anti-cheat model was completely opaque to external analysis. Cheat developers could not observe the model's behaviour through API probing (as they could with cloud-hosted models) because the model ran entirely within the game server infrastructure with no external interface.

Results: cheat detection rate from 72 percent to 94 percent (22 percentage point improvement). False positive rate maintained at 0.3 percent (falsely banning legitimate players being extremely damaging to reputation). Time to detect: from average 48 hours (manual review of reports) to average 12 minutes (real-time behavioural detection). Competitive game player satisfaction increased 34 percent post-implementation (measured through player surveys about match quality). (4) Content generation AI: accelerating game development. Level generation: fine-tuned models generating game levels, dungeons, and environments based on the group's design specifications and art style.

Game designers providing high-level briefs ("a medium-difficulty dungeon for level 30-35 players with a frost theme and a boss encounter") and the AI generating: spatial layouts (room configurations, corridors, environmental features), encounter placement (enemy types, positions, patrol routes, difficulty scaling), reward distribution (loot tables, resource nodes, secret areas), and visual styling (texture selection, lighting parameters, environmental effects matching the game's art direction). Human designers reviewed and refined AI-generated content — the AI producing "first drafts" that designers iterated on rather than building from scratch. Content generation productivity: level designers producing 3.4x more content per sprint (measured in completed, quality-approved game content). NPC dialogue: Llama 3 70B fine-tuned on the group's narrative archives (4.2 million lines of dialogue across 14 games) — generating NPC dialogue that matched each game's narrative tone and world-building.

NPCs conducting dynamic conversations with players rather than repeating scripted lines — responding to player questions, commenting on game events, and revealing lore organically. Player feedback: 76 percent of surveyed players rating AI-generated NPC dialogue as "good" or "excellent" quality — compared to 82 percent for human-written dialogue (close to parity). Asset generation: AI generating supplementary game assets — textures, simple 3D props, and sound effects that supplemented human-created core assets. Asset generation reducing art team workload on repetitive assets by 28 percent, allowing artists to focus on hero assets and creative direction. (5) Player privacy and children's data protection. Data minimisation: AI models designed to process behavioural patterns rather than personal identifiers — player behaviour represented as feature vectors rather than linked to personal data. This approach satisfying GDPR data minimisation principles while enabling effective AI.

Children's data: for games with significant underage player populations, enhanced protections: no monetisation AI targeting for accounts identified as under-16 (parental controls and age verification), content moderation AI more conservative for children's game modes, and behavioural data processing limited to game experience optimisation (not monetisation or profiling). IMY compliance: self-hosted deployment with no third-party data sharing providing the strongest privacy position for IMY scrutiny. Complete data processing records maintained for every AI application, demonstrating purpose limitation and proportionality.