Geopolitics, AI, and Level 4 Autonomy: The Datacenter Battle

By Hrishikesh S , Senior Specialist, Autonomous Driving and E/E Technology, S&P Global Mobility

The shockwaves from DeepSeek’s unveiling may seem confined to the tech sector, but they have far-reaching implications for the automotive industry — especially the future of Level 4 autonomy.

On the surface, DeepSeek is a stark reminder that innovation is relentless. No matter the barriers — geopolitical restrictions, supply chain disruptions or market dominance by a few players — technology finds a way to adapt and evolve.

DeepSeek is more than just a new entrant in AI; it challenges conventional wisdom about how compute power is developed, optimized and deployed.

DeepSeek: A wake-up call to innovate

DeepSeek introduced the DeepSeek R1 AI model, built on DeepSeek V3, offering comparable performance to GPT-4 and Gemini at a lower cost. Leveraging mixture of experts (MoE) architecture, it activates only relevant "experts" for each task, optimizing efficiency and reducing GPU reliance. This contrasts with monolithic models such as GPT-4, which handle all tasks universally.

Another key aspect of DeepSeek’s innovation is its reinforcement learning (RL) approach, which differs from the standard methods used in other AI training frameworks. Unlike models that require human feedback to fine-tune responses, DeepSeek R1 uses Group Relative Policy Optimization (GRPO), a simplified reinforcement learning technique that eliminates the need for complex reward functions.

DeepSeek is more than just a new entrant in AI; it challenges conventional wisdom about how compute power is developed, optimized and deployed.

DeepSeek prioritizes transparency by making its reasoning process open source, unlike OpenAI’s proprietary chain of thought methods. The model was trained on Nvidia H800 graphics processing units (GPUs), which DeepSeek deems sufficient, though its lack of access to Nvidia’s most advanced chips raises concerns about future development scalability.

For the automotive industry, where AI-driven decision-making is essential to autonomy, this shift is more than theoretical. As alternative compute ecosystems emerge, the entire architecture behind Level 4 autonomous vehicle datacenters and cloud-based training could be reshaped. The question is no longer whether AI breakthroughs will impact the industry, but how quickly and through which pathways.

Potential impact on Level 4 autonomy: Rethinking cloud and datacenter infrastructure

In the development of AI systems for autonomous vehicles, the training phase and inference phase have two distinct hardware requirements. The training phase often involves developing AI models with billions of parameters; this necessitates substantial computational power, typically provided by Nvidia’s GPUs in datacenters. This phase focuses on processing vast datasets to optimize model performance.

The inference phase pertains to the deployment of trained models in vehicles, with the aim of reducing the model’s complexity — often by one to two orders of magnitude — to ensure efficient real-time processing. This reduction allows the use of specialized system on chip (SoC) solutions, such as Mobileye EyeQ 6H and Nvidia Thor SoCs, which are designed to streamline inference tasks within the constraints of the automotive environment.

As alternative compute ecosystems emerge, the entire architecture behind Level 4 autonomous vehicle datacenters and cloud-based training could be reshaped.

The key question in evaluating DeepSeek’s success is whether it presents a viable alternative to Nvidia’s powerful GPUs in supporting advanced AI in datacenters, including those required for the Level 4 autonomous driving training phase. While DeepSeek’s MoE and RL innovations have demonstrated efficiency gains in AI training and inference, its architecture was designed for general AI applications, rather than for the highly specialized needs of autonomous vehicle perception, decision-making and planning.

Autonomous vehicles require models that generate high-probability outputs and meet stringent precision and reliability standards, where "correctness" is not just an optimization goal but a functional necessity. This means that although DeepSeek showcases how AI can be trained with fewer computational resources, its direct applicability to autonomous vehicle backend architectures remains uncertain.

While DeepSeek’s breakthroughs in MoE and RL suggest that alternative AI models can be trained without relying on leading edge GPUs, it remains to be seen whether these models can be adapted to meet the strict performance and safety requirements of Level 4 autonomous vehicle technology. Even with advancements from companies such as DeepSeek in AI model development, it is essential to assess how these models can be adapted to vehicle architectures, considering the specific performance and memory requirements of automotive SoCs.

Although DeepSeek showcases how AI can be trained with fewer computational resources, its direct applicability to autonomous vehicle backend architectures remains uncertain.

For now, the autonomous vehicle industry remains reliant on domain-optimized AI models built specifically for automotive applications. Looking ahead, DeepSeek and other mainland Chinese companies may face growing difficulties competing with frontier models due to their limitations in semiconductor access.

By subscribing to AutoTechInsight (ATI), you can quickly gain intel on market developments and technology trends, dive into granular forecasts, and seamlessly drive analytics to support challenging decision making.

Click here to subscribe.