Nvidia, the well-known technology company, reported that it made over $19 billion in profit in the last quarter. However, this impressive figure didn’t fully reassure investors about the company’s future growth. During a call with investors, analysts asked Nvidia’s CEO, Jensen Huang, how the company would perform if tech firms started adopting new techniques to enhance their AI systems.
One of the new methods discussed was called “test-time scaling,” which is a technique used in OpenAI’s latest AI model. The basic idea behind this method is that AI models can provide better answers if they are given more time and computing resources to process the questions. This means that when a user submits a question, the AI can take longer and use more computing power to come up with a response, which could lead to more accurate results.
During the earnings call, Huang was questioned about whether he noticed AI developers moving towards these new methods and how well Nvidia’s older computer chips would perform in this new environment. He responded positively, suggesting that test-time scaling and the o1 model could play a significant role in Nvidia’s future. He described this development as “one of the most exciting” advancements and referred to it as a “new scaling law.” Huang assured investors that Nvidia is well-prepared to adapt to these changes.
His comments echoed what Microsoft CEO Satya Nadella mentioned at a recent event, indicating that the o1 model represents a new approach to enhancing AI systems. This shift is particularly important for the chip industry because it increases the focus on AI inference, which is the process of using AI models to generate responses after they’ve been trained. Although Nvidia’s chips are widely recognized as the best for training AI models, there are several well-funded startups, like Groq and Cerebras, that are developing very fast chips specifically for AI inference. This means that Nvidia may face more competition in this area.
Despite some reports suggesting that progress in creating generative AI models is slowing down, Huang reassured analysts that developers are still actively improving their models by adding more computing power and data during the initial training phase, known as pretraining. He emphasized that the scaling of foundational models is still ongoing, although he acknowledged that simply adding more resources might not be sufficient on its own.
Investors were likely relieved to hear this, especially since Nvidia’s stock price has skyrocketed by more than 180% in 2024, largely due to the demand for the AI chips that companies like OpenAI, Google, and Meta use to train their models. However, some experts, including partners from the venture capital firm Andreessen Horowitz, have pointed out that the benefits of these new methods may already be starting to diminish.
Huang explained that most of Nvidia’s current computing tasks revolve around the pretraining of AI models rather than inference. He believes that as more people begin to use AI models, the demand for AI inference will grow significantly. He also highlighted that Nvidia is currently the largest platform for AI inference globally, and the company’s size and reliability give it a substantial advantage over smaller startups.
Read Also: Lightning Aims To Make AI Management Easy!
He concluded by expressing hope that one day, AI inference will become widespread, which would signify a major success for the technology. Huang stated, “Everyone knows that if they innovate on top of CUDA and Nvidia’s architecture, they can innovate more quickly, and they know that everything should work.” This underscores Nvidia’s commitment to providing a strong foundation for AI development.
