The news, reported by AFP, is as sparse as it is telling: the United States is chasing AI talent abroad, while China is strengthening its domestic pipeline. On one side, H-1B visas and recruitment drives at universities worldwide; on the other, a state apparatus that nurtures technical training from the school years onward. The immediate reading is geopolitical competition. But for those dealing with on-premise deployments, self-hosted LLMs, and the real costs of inference, the contest is more concrete and far-reaching.

Hardware is everywhere: NVIDIA GPUs with tens of gigabytes of VRAM, multi-card servers linked via NVLink, fast storage for billion-parameter checkpoints. Yet without people who can configure a serving framework like vLLM, apply the right level of quantization without quality loss, orchestrate containers on Kubernetes in air-gapped environments, and fine-tune on proprietary data, that hardware remains inert silicon. The difference comes from engineers who can translate data residency requirements into reliable on-premise architectures, reducing latency and keeping TCO under control.

China’s approach grasps this link. Strengthening the domestic pipeline not only produces researchers for academic competition but creates a fabric of systems administrators and ML engineers who can bring inference to local servers, away from public clouds subject to foreign jurisdictions. In sectors like defense, healthcare, and finance, where compliance demands that data never leave national borders, having professionals who work on self-hosted stacks becomes a strategic asset. It is a form of technological sovereignty that depends on hardware, yes, but also on the skills to manage it.

The United States, by courting foreign talent, aims to plug immediate gaps. That’s a quick fix, but it risks dependency on visas and fragile geopolitical dynamics. Moreover, it does not automatically spread those skills throughout the domestic productive system. Without adequate generational renewal, entire organizations that could benefit from on-premise deployment – from public administrations to mid-sized enterprises – risk remaining on the sidelines, hamstrung by the technical challenge of managing models without resorting to cloud services.

There is a structural lesson here for anyone designing local AI strategies. The cost of a server with eight A100 cards is not just the hardware outlay: it’s also the cost of finding, hiring, and retaining someone to run it. A single quantization optimization by an experienced technician can halve the required VRAM, reducing infrastructure expenditure. A talent shortage, conversely, pushes toward ready-to-use cloud solutions, ceding data control and increasing latency. Thus, the question “who trains the experts” directly affects the economic feasibility and sovereignty of deployments.

AFP’s ostensibly political news signals that we have entered a phase where human capital is once again the decisive multiplier of compute power. This is not a war to conquer the best papers, but to put LLMs into production on one’s own infrastructure. And for those choosing the on-premise path, that means training people is every bit as important as buying GPUs.