The news comes from Taipei, where WeMo has reaffirmed its course: no owned vehicle fleets, no monolithic hardware, but an asset-light model and an open platform. At first glance, an announcement that speaks only of scooter sharing and Mobility-as-a-Service. Yet, digging beneath the surface, the reasoning touches chords familiar to those building stacks for Large Language Models on their own premises.

WeMo’s move

WeMo, founded by Jeffrey Wu and led by CEO David Liu, has decided to scale Taiwan’s MaaS ecosystem without owning the entire fleet or sealing the technology behind a proprietary wall. The idea is simple: open APIs, connect operators, let the value multiply without loading fixed capital costs. A two-sided architecture — financial lightness and technical openness — that WeMo expects will drive growth over the next decade.

Asset-light and on-premise: unexpected parallels

Those involved in on-premise deployment of language models know the tension between buying specialized hardware — GPUs with dozens of GB of VRAM, NVLink systems, servers with significant power consumption — and relying on cloud services. WeMo’s bet echoes the debate between doing everything in-house (self-hosted) and consuming external APIs. Owning the infrastructure gives data control, predictable latency, and vendor independence, but it shifts updates, maintenance, and TCO onto the organization. In contrast, an asset-light model for AI as a Service moves complexity to the provider, at the cost of reduced sovereignty.

Openness as a sovereignty lever

The second pillar, the open platform, has a direct equivalent in the universe of LLM frameworks. WeMo exposes standard interfaces to integrate third-party services; in AI, open-source solutions like vLLM, Ollama, or Kubernetes allow orchestrating inference without depending on a single proprietary stack. Interoperability reduces lock-in, eases GDPR compliance, and leaves room for customizations such as fine-tuning on internal datasets. Yet governing an open platform is not free: it requires integration, monitoring, and security skills that, on owned hardware, become heavier.

Open questions for local deployment

The WeMo case offers no technical answers but poses the same questions every on-premise AI team should face: how much control do we want over our data? What total cost of ownership are we willing to bear, between initial CapEx and ongoing OpEx? The answer is not uniform, and the asset-light or self-hosted direction depends on risk profile, regulatory sensitivity, and usage scale. At a time when local inference becomes accessible thanks to quantization techniques and more affordable hardware, WeMo’s example suggests that an ecosystem’s value lies not only in hardware, but in the ability to make the right pieces talk to each other, without ever losing control of one’s course.