The Grand Theft Auto modding scene has just delivered a little gem: a mod lets you boot up GTA 3 and Vice City without ever leaving San Andreas. The cities of Liberty City and Vice City, along with their missions and early-2000s vibe, become explorable islands inside the 2004 installment. No remake, no separate emulator—just the same executable, the same installation, one game containing three.

Behind the nostalgia hacking lies a consolidation principle that translates into tangible benefits in on-premise inference. When a team runs multiple language models on local infrastructure, fragmentation is the first enemy: each separate instance consumes VRAM, CPU cycles, and, most of all, energy. Consolidating workloads on a single node—just like the San Andreas engine running maps and scripts from other titles—reduces redundancy and allows sharing resources that would otherwise sit underutilized.

The modding community shows that legacy software and different environments can coexist with remarkably low overhead, provided the container is mature enough. For teams deploying LLMs on-site, the parallel is immediate: multi-GPU servers linked by NVLink or NVSwitch can host several quantized models, each with its own context window and numerical precision, without requiring dedicated machines. This plays directly into TCO: less hardware to buy, less cooling, fewer orchestration licenses to manage.

Data sovereignty gains ground when the entire stack sits in a single corporate rack, much like the gamer who has everything they need inside their own PC without connecting to external services. The modder’s experiment, however distant from enterprise racks, tells a structural story: the temptation to always separate workloads for order or security reasons clashes with the efficiency of consolidation, and the latter often wins. A lesson that on-premise AI teams know well—and one that now finds a mirror even in a twenty-year-old video game.