Geely's Automotive Shift: Fewer Factories, More Efficiency and TCO

Li Shufu, the billionaire chairman of Geely Holding Group and the man who acquired Volvo Cars from Ford for $1.8 billion in 2010, has reached a conclusion that many of his peers in the global automotive industry have been slower to embrace. In his view, the world has too many car factories, and building more of them is no longer the most effective strategy. This perspective has led Geely to a significant change of course: the company intends to cease constructing new production plants, opting instead to utilize existing facilities owned by other industry players.

Geely's decision is not merely a tactical move but reflects a broader trend towards resource optimization and more prudent Total Cost of Ownership (TCO) management. In an increasingly competitive and saturated global market, the ability to maximize the efficiency of existing assets and reduce operational costs becomes a critical success factor. This approach, though applied to the automotive industry, offers relevant insights for other capital-intensive sectors, such as infrastructure for artificial intelligence and Large Language Models (LLMs).

Resource Optimization and TCO in the LLM Era

Geely's strategy of "borrowing" production capacity from others finds a conceptual parallel in the challenges companies face when deploying LLMs on-premise. For CTOs, DevOps leads, and infrastructure architects, the decision of whether or not to build new dedicated hardware infrastructure is complex. Often, the goal is to maximize the utilization of existing GPUs, optimize available VRAM, and ensure high throughput with minimal latency, all while keeping TCO under control.

Geely's approach highlights how investment in new assets (CapEx) can be avoided or reduced by prioritizing the use of resources already available in the market (OpEx or partnership agreements). In the context of LLMs, this translates into evaluating between purchasing new high-end GPUs, such as A100s or H100s, for a self-hosted deployment, and exploring solutions that allow leveraging existing computing capacity, perhaps through agreements with data centers or specialized service providers. The choice depends on factors such as data sovereignty, compliance requirements for air-gapped environments, and the need for granular control over the entire inference or fine-tuning pipeline.

Strategic Implications and Trade-offs for AI Infrastructure

Geely's shift underscores a fundamental trade-off: total control over an asset against a high initial investment, versus greater flexibility and potentially lower operational costs, but with reliance on third parties. For LLM deployments, this manifests in the choice between a fully internally managed bare metal infrastructure, which offers maximum control over security, customization, and performance optimization (e.g., batch size, quantization), and hybrid solutions or those based on external services.

Companies prioritizing data sovereignty and security often opt for on-premise or air-gapped deployments, investing in specific hardware and local stacks. However, even in these scenarios, TCO analysis is crucial. It's not just about the initial cost of GPUs and servers, but also energy costs, maintenance, specialized personnel, and technological obsolescence. Geely's strategy suggests that, in some contexts, sharing or leasing resources can be a way to mitigate these burdens, shifting the focus from pure investment to efficient capacity management.

A Mature Market Perspective

Geely's decision to abandon the construction of new factories in favor of utilizing existing third-party facilities is a sign of maturity in the automotive market, where efficiency and TCO optimization have become absolute priorities. This "asset-light" or "capital-light" mindset is increasingly relevant in the artificial intelligence landscape as well. As LLMs become a central component of corporate strategies, the ability to deploy and manage these technologies efficiently, both on-premise and in hybrid environments, will be a distinguishing factor.

For those evaluating on-premise deployments, analytical frameworks explored by AI-RADAR on /llm-onpremise exist to assess the trade-offs between control, costs, performance, and scalability. Geely's experience, though in a different sector, offers a valuable lesson: long-term sustainability does not necessarily lie in accumulating assets, but in their intelligent management and the optimization of existing capacity—a principle that applies strongly to the complex infrastructure required by Large Language Models.