With the AI boom triggered by ChatGPT, the consensus in the industry has become that "larger-scale AI models are better," driving a competition among tech giants like Microsoft (MSFT.O), Google (GOOGL.O), Amazon (AMZN.O), and Meta Platforms (META.O) in chip procurement, with NVIDIA (NVDA.O) being the biggest beneficiary due to its outstanding performance in AI training GPUs. However, this competition may soon change, with the industry facing multiple obstacles encountered in the pursuit of larger AI models.
NVIDIA's Leading Position and Bottleneck Challenges
NVIDIA's GPUs dominate AI model training due to their ability to perform efficient parallel computing. Currently, the main metric for AI capability is the number of model parameters, and more parameters mean the need for more GPUs. However, there has been skepticism in the industry regarding the benefits of scaling model sizes. Waseem Alshikh, co-founder of Writer, pointed out, "After exceeding one trillion parameters, the returns tend to diminish." Microsoft CEO Satya Nadella also stated at a recent Ignite conference that skepticism around scaling AI models may stimulate more innovation.
Nevertheless, industry leaders in AI, such as OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei, strongly oppose these doubts, believing that the potential for AI expansion has not yet reached its limit.
Data Bottlenecks and the Future Path of AI
Thomas Wolf, Chief Science Officer at Hugging Face, pointed out that the lack of high-quality training data may be the biggest challenge facing AI development. "We exhausted the internet as a resource for training data months ago." This limitation may prompt a future shift towards smaller models based on company or personal data, rather than the large models currently dominated by cloud giants.
Meta's Chief AI Officer Yann LeCun emphasized that developing models with memory, planning, and reasoning capabilities is key to achieving true artificial general intelligence (AGI), rather than solely relying on larger chips.
The Rise of Inference and Opportunities for New Competitors
The focus of AI is gradually shifting from training to inference (the process of generating answers or results), bringing new dynamics to the chip market. Inference computing may not rely on NVIDIA's GPUs as much as training does, and AMD (AMD.O), Intel (INTC.O), Amazon's custom chips, and startups may all carve out a share in this field. Microsoft's Eric Boyd believes that, in addition to model size, technological improvements during the inference process are equally critical.
NVIDIA has recognized the potential of inference, mentioning in its recent financial report that inference business accounts for 40% of data center revenue and is growing rapidly. Its newly launched NVL72 server system has shown a 30-fold performance improvement in inference, demonstrating strong competitiveness in this field.
Diverse Winners in a New Phase
The AI competition shifting from training to inference means that industry opportunities will become more decentralized. Although NVIDIA remains the winner in the short term, as the importance of inference rises, AMD, Intel, and other competitors may gradually erode NVIDIA's market share. For investors, the focus at this stage is no longer solely on supporting larger model training, but rather on preparing for a series of new winners that may emerge when using models.
Article forwarded from: Jinshi Data
