Nvidia CEO Huang: The DeepSeek incident highlighted the massive demand for AI compute power.

NVIDIA CEO Jensen Huang spoke about the surge in demand for AI computing power in a recent interview. Specifically, he cited the DeepSeek case to emphasize the growing computational demand driven by the evolution of AI models. This article provides a detailed explanation of the current state and future of AI computing, as well as the medium-term signals investors should watch, based on Mr. Huang's remarks.

1. Current state of AI computing demand and short-term signals

1-1. Indicators of short-term demand

Mr. Huang first cites NVIDIA's POS (point-of-sale) data and forecasts as indicators of short-term demand. These directly reflect current market trends and are important metrics for the health of demand.

Furthermore, he points out that the emergence of new startups not included in forecasts is also a key factor. These startups are born from breakthroughs in new reasoning AI capabilities and artificial general intelligence (AGI) capabilities, each requiring additional computing power.

"These startups are seeking additional computing power from cloud service providers like AWS as soon as possible. This is a factor that is pushing demand even higher, in addition to the POS and forecasts we are already aware of," Mr. Huang stated.

1-2. Remarks by Amazon CEO Andy Jassy

Amazon CEO Andy Jassy has also stated that if there were more AI resources available to sell through AWS, they could sell even more. This indicates that current demand for AI computing power significantly exceeds supply.

"Andy Jassy's remarks indicate that short-term demand is extremely strong. This clearly shows how much AI computing power is in demand in the market," Mr. Huang emphasized.

2. Medium-term signals and investment in data centers

2-1. Capital investment in data centers

Mr. Huang cites the increase in capital investment in data centers as a medium-term signal. Capital investment in data centers this year has increased significantly compared to last year, indicating that demand for AI computing power is continuously growing.

"Last year's investment in data centers was very large, but this year it is on an even larger scale. With the introduction of Blackwell and the coming online of new data centers, we expect significant growth to continue," Mr. Huang explained.

2-2. The arrival of the reasoning AI era

Mr. Huang emphasizes the arrival of the reasoning AI era as a long-term perspective. This is an era where AI will have the ability to think and reason for itself before answering questions.

"Instead of generating an immediate answer to a question, AI will first reason about the question, break it down step-by-step, and perform internal searches. Then, it will create and present the optimal answer," Mr. Huang said.

The amount of computation required for this reasoning process is expected to be 100 times greater than that of conventional AI models. Reasoning AI models such as DeepSeek, ChatGPT-4, and Grok-3 are already showing this trend and will require even more computing power in the future.

3. Increasing computational demand as seen from the DeepSeek case

3-1. Misunderstandings and the truth about DeepSeek

Initial reports on DeepSeek suggested that it was achieving more with less computation, but Mr. Huang points out that this is a misunderstanding. In reality, the DeepSeek case demonstrates that the demand for AI computing power will increase significantly.

"The DeepSeek case shows that compute demand is not decreasing, but rather increasing. This means that as AI evolves, the amount of computation required will increase exponentially," explains Huang.

3-2. Three phases of AI development

Huang explains the three phases of AI development.

1. Pre-training:

   • The phase of learning basic mathematics, language, and knowledge

2. Post-training:

   • The phase of deepening learning through human feedback or AI feedback

   • In this phase, reinforcement learning and synthetic data generation are utilized, significantly increasing the computational load

3. Inference:

   • The phase of reasoning about questions and generating optimal answers

   • Even in this phase, 100 times more computation is required than in conventional AI models

"In the post-training process, many innovations are emerging, particularly related to reinforcement learning and synthetic data generation, which are contributing significantly to the increase in computational load," says Huang.

4. Points for investors to watch

Investors need to understand that the demand for AI computing power will continue to grow, and that computational demand will increase exponentially, especially with the arrival of the inference AI era.

Furthermore, increased capital investment in data centers and the emergence of new startups are also important indicators of market growth. It is important to synthesize this information to develop investment strategies related to AI computing.

NVIDIA CEO Jensen Huang's remarks suggest that the demand for AI computing power will continue to grow. In particular, with the arrival of the inference AI era, computational demand is expected to increase exponentially. Investors should carefully consider their AI computing-related investment strategies based on this information.

Related Articles