If the AI race were decided in the court of public opinion, China would already have the trophy. A new report by London-based consultancy Public First reveals that in 11 of the 15 countries surveyed, a clear majority now believes China has overtaken the United States in AI capability and innovation. Beneath the headline numbers lies a paradox: the world acknowledges China’s perceived technological edge but remains unwilling to trust the supposed winner.

The survey, part of the global conversation around AI perception, doesn’t measure performance metrics but rather a factor equally critical for those designing infrastructure: sentiment. And it is precisely the gap between perception and trust that triggers strategic reflections, especially for organizations evaluating where and how to deploy their large language model workloads.

Perception and trust: the diverging tracks of global AI

Public First’s findings serve as a temperature check on a climate where technological leadership is judged not by objective benchmarks but by media coverage, consumer application diffusion, and the speed at which Chinese models climb open-source leaderboards. Yet the same respondents who applaud China’s edge are not ready to entrust their data, code, and decision-making processes to an ecosystem perceived as opaque and subject to state control.

The short circuit is clear: if the country embodying theoretical AI progress is also the one enterprises fear depending on, architectural choices become a form of geopolitical insurance. This is where the on-premise deployment argument gains weight. Hosting inference and fine-tuning within one’s own corporate boundaries, on self-hosted hardware, is no longer just a matter of latency or customization but a lever to guarantee data sovereignty and reduce exposure to supply chains whose regulatory center of gravity eludes Western compliance frameworks.

The infrastructure ripple effect

For CTOs building the AI factory, Public First’s data is not a casual poll but a signal reinforcing the move toward architectures combining open models, aggressive quantization, and on-premise servers. This isn’t about demonizing China as a technology supplier – many of the most efficient models originate there – but recognizing that stakeholder trust and regulations like GDPR push organizations toward stacks where data never leaves the corporate perimeter.

This tension finds fertile ground in the galaxy of tools enabling on-premise LLM serving without sacrificing competitive performance. Frameworks such as vLLM or orchestration systems on Kubernetes allow local inference to scale, while the availability of consumer GPUs with increasingly generous VRAM (think RTX 4090 with 24 GB or multi-GPU setups) makes self-hosting feasible even beyond large data centers. The thorniest trade-off remains TCO: acquiring dedicated accelerators involves significant CapEx, but over time it can balance recurring cloud costs, especially when high token volumes or continuous data flow audits come into play.

A map for on-premise evaluation

Those venturing into an on-premise infrastructure assessment for LLMs find a landscape more mature than expected. The variables to weigh are many: workload type (training, fine-tuning, or inference-only), acceptable quantization levels, deployment context (air-gapped, edge, or hybrid), and the need to retain full data control. AI-RADAR covers these topics extensively through its on-premise strategy section, analyzing scenarios where companies can build autonomous stacks without sacrificing output quality.

The absence of a blindly trusted technology leader, highlighted by the Public First report, is not a condemnation to paralysis but an incentive to diversify. Rather than leaning on a single cloud provider tied to one jurisdiction, organizations can combine models from different labs, fine-tune them on proprietary data, and serve them internally, maintaining governance over the entire data lifecycle.

The bigger picture: toward jurisdictionally separated AI

The fracture between perceived leadership and trust is not merely an analytical curiosity. It’s an engine already reshaping the geography of AI infrastructure. European companies, for instance, are accelerating private cluster builds for Large Language Models, driven not by fear of Chinese performance but by the desire to avoid dependency on an ecosystem whose regulatory alignment remains uncertain. This trend will likely lead to a landscape of “multiple sovereignties,” where each major economic bloc equips itself with independent computational capacity, with models trained and served within its own borders.

At its core, the Public First finding tells us that technological primacy alone cannot build a trust ecosystem. And for those doing business with AI, trust is the most expensive infrastructure to replace.