At the GTC conference, why did Jensen Huang say that computing power demand will increase by 100 times!?

Last night's big drop, is the market no longer buying Jensen Huang's story?

Last night, the market did not show any respect, and NVIDIA's stock price continued to plummet.

What’s even more disrespectful is that NVIDIA's overnight drop occurred right after Jensen Huang began his keynote speech.

I think the main reasons are:

First, the overall market is weak, after all, the Nasdaq fell nearly 2%.

Second, NVIDIA had previously rebounded by more than ten points, and the positive news from GTC had already been fully priced in.

Third, the market tends to have a short-term mindset and has not recognized that this show should focus on the long-term development of AI.

Old Huang: Three Stages of AI Technology Evolution

The development path of AI: from the current Generative AI to the upcoming era of Agentic AI, and finally to Physical AI.

Generative AI: This refers to LLM, where you ask and I answer, and NVIDIA's market value of 3 trillion is built on this.

Agentic AI: This adds "proactivity" and "decision-making ability" to Generative AI, capable of understanding tasks, breaking down steps, and interacting with other systems. This is all backed by a significant increase in computing power. (Old Huang mentioned at the conference: The combination of Blackwell Ultra and Dynamo significantly improves AI inference efficiency (40 times that of Hopper).)

Physical AI: This is AI in the "real world," which not only thinks and plans but also controls machines in the real world, such as robots and self-driving cars. In the internet era, everything can be connected; in the Physical AI era, everything can be AI.

The relationship among the three is: Generative AI is about "creating content" → Agentic AI is about "making decisions" → Physical AI is about "doing real things." From virtual to semi-virtual and semi-real, and then to fully real, the demand for computing power is certainly not on the same scale.

As for the claim of a 100-fold increase in computing power, it may have some exaggeration, but the explosive growth is undeniable:

Last year (2024), the demand for AI computing power was mainly focused on training large language models (like ChatGPT), and Jensen Huang believed that the models at that time were still relatively "simple." This year (2025), with the explosion of inference AI, multimodal models, and physical applications, the demand for computing power has upgraded from "pre-training + simple inference" to "complex inference + real-time interaction + physical control," leading to a leap in demand levels. He specifically mentioned in a CNBC interview that the computational load of the inference process could be 100 times that of last year.

Changes in capital expenditures among giants, Old Huang: Data center capital expenditures will reach $1 trillion by 2028 Jensen Huang said that capital expenditure on data centers will exceed $1 trillion by 2028.

(Huang mentioned at the conference: The top four cloud service providers (CSPs) in the United States, the so-called hyperscalers, purchased 1.3 million NVIDIA Hopper architecture chips, and in 2025, they will purchase 3.6 million Blackwell architecture chips.)

What I observe is that the capital expenditures of several major U.S. companies have significantly increased since 2023, and after the Q4 2024 earnings report, capital expenditures also greatly exceeded expectations: You can check my previous analysis

According to reports from Microsoft, Google, Amazon, and Meta, the total capital expenditure of the four giants in 2024 is $246 billion, a 63% surge compared to $151 billion in 2023. This year's total capital expenditure may reach $335 billion, an increase of 36.1%.

Several domestic giants show the same trend:

The increase in capital expenditure can be simply understood as investment for AI. This is just the situation of a few companies; there are many more enterprises and governments globally. The $1 trillion by 2028 may be inflated, but the trend is the same. Most of this will flow into NVIDIA's pockets, as it monopolizes over 90% of AI computing power.

Product Line:

At every conference or on NVIDIA's official website, there is also a product roadmap. There's not much to say, so I'll just post a picture for everyone to see.

View on Stock Price

In the short term, it will fluctuate with the market, but in the long term, it remains the best AI target. Long-term investors can continue to hold for three to five years without issue.

$NVIDIA(NVDA.US)