When Christopher Nolan talks about artificial intelligence, he does so with the precision of a filmmaker accustomed to constructing complex narrative architectures. In a recent interview, he called AI an obvious ‘Trojan horse’. ‘Everybody knows the Greeks are inside,’ he added. That sentence isn’t just a sound bite: it’s the perfect synthesis of an anxiety rippling through the tech sector, one that has a lot to do with who controls the foundations on which AI rests.

The myth of the Trojan horse tells of an apparently harmless gift that conceals soldiers ready to slip out at night and open the city gates. Nolan isn’t suggesting that algorithms are evil in themselves; his narrative screen is subtler. The danger isn’t the code, but the box that contains it. And in 2025, that box has a precise name: centralized infrastructure, often cloud-based, managed by a handful of major vendors.

Take Large Language Models. Most companies use them via APIs from external services: OpenAI, Google, Anthropic. It’s convenient, fast, and doesn’t require GPUs. But it’s also a gigantic wooden horse, because those models live on servers you don’t own, process data you don’t fully control, and are trained with logics you cannot inspect. The Greeks – in this case, dependency on third parties and the loss of technological sovereignty – are already inside the walls, and everyone knows it.

Nolan’s point, interpreted through an AI-RADAR lens, becomes structural: artificial intelligence is not an inherent threat, but it becomes one when its adoption passes through an architecture that strips the user of audit capabilities, granular control, and genuine customization. It’s no coincidence that the organizations most sensitive to data sovereignty – defense, healthcare, finance – are accelerating precisely on on-premise deployment: bringing models into their own data centers, with local inference and no traffic outside. It’s the way to check that the horse is empty, or at least to decide when and how the hoplites come out.

Who benefits from this reading? Not only large enterprises with substantial hardware budgets. Medium-sized businesses, thanks to quantization, fine-tuning on consumer hardware, and frameworks like vLLM or Ollama, can now run LLMs locally at manageable cost. The TCO of a server with a 48 GB VRAM GPU, when compared to monthly API bills that explode as tokens increase, starts to show its cost-effectiveness over the medium term. And above all, it eliminates the risk that someone uses your horse to enter your home.

Of course, self-hosted doesn’t mean immunity. Models can contain biases, software can have vulnerabilities, and maintenance requires in-house skills. But shifting the center of gravity from cloud to on-premise is the most direct way to turn AI from an insidious gift into a tool you own. As Nolan might say: it’s not the twist ending, it’s the premise. We knew it all along.