At the Intelligent Electric Vehicle Development High-Level Forum (2026) held at the Beijing National Convention Center from April 11-12, the focus was on advancing the intelligent, green, integrated, and international development of new energy vehicles.
Baidu Vice President Shi Qinghua stated that while the industry previously concentrated on autonomous driving training and R&D, AI has now become a standard feature in vehicles. With intelligent cockpits and autonomous driving systems serving massive user bases, OEMs are accelerating the practical implementation of AI applications. He indicated that a fundamental shift in computing power structure is underway, moving from an era dominated by training to one dominated by inference.
He presented data showing that by 2026, the increase in computing power demand from inference tasks will reach 73%, accounting for two-thirds of the total. This represents a dramatic rise from just one-third in 2023. The overall growth in computing power over these years is measured in hundreds to thousands of times. Secondly, the cost of performing inference with equivalent performance using the GPT-3.5 model has decreased by over 200 times between 2022 and 2024. Another striking figure, based on the latest OpenAI data, reveals that global large AI model usage reached 27 trillion tokens in the previous week alone, a week-on-week increase of 18.9%. In China, this growth is even more pronounced, exceeding 30% week-on-week. "This signifies that AI is rapidly integrating into our daily lives and work across all commercial and industrial sectors at an astonishing pace, particularly within the automotive industry," he said.
Within the automotive sector, he outlined three primary activities: First, large models are being fully integrated into intelligent cockpits, enabling generative HMIs, multimodal reasoning, and proactively responsive intelligent agents. Second, the entire internal R&D process for vehicle intelligence—spanning research, production, supply chain, sales, and service—is being overhauled using AI for efficiency gains. Third, while the industry previously discussed "software-defined vehicles" and "R&D-defined software," it is now AI that is defining R&D. He remarked that programmers developing large models face a peculiar dilemma: by creating exceptionally powerful models, they effectively redefine their own roles, essentially making their previous jobs obsolete.
He pointed out that since late last year, all intelligence is fundamentally supported by computing power. The industry is now likely entering a phase of computing power scarcity, as acquiring NVIDIA GPUs has become increasingly difficult. Therefore, domestic GPU alternatives should be developed to fill this gap.
Consequently, Shi Qinghua offered three recommendations: First, companies must begin their own computing power planning, as access to hardware is now critical for both personnel and business operations. Second, they should start building their own enterprise-level large model platforms to enhance efficiency across the entire chain of research, production, supply, sales, and service. Third, and most crucially, he emphasized the importance of robust data governance. "We consistently talk about data governance. In the next phase, regardless of how advanced intelligence becomes, it will all be built upon piles of data," he concluded.
Comments