"Compute is GDP"—Jensen Huang on the next decade of AI at Morgan Stanley TMT

In 1998, NVIDIA went public with a $48 million IPO. Its market capitalization was $300 million. It is said that the most frequent question investors asked during the Morgan Stanley roadshow was, "When will you go bankrupt?"

27 years later, NVIDIA recorded $46 billion in net profit in the most recent quarter alone. Jensen Huang reflected on that trajectory while on stage at the Morgan Stanley TMT conference, discussing the current state of AI and the next decade. This article summarizes the key points.

1. Why NVIDIA is "full-stack"—33 years of accumulation

1-1. From video games to AI, the philosophy remains unchanged

Huang describes NVIDIA's strength with the single word "full-stack." It is a structure that completes everything in-house, from chip design and library/software integration to supporting customer deployment.

The starting point is computer graphics and the video game industry. "Without NVIDIA, today's game engines would not exist. We have been an algorithm company for 33 years," Huang stated. There are hundreds of millions of GeForce gamers, and some of them discovered CUDA and became AI researchers—it is a famous story that Ilya Sutskever and Alex Krizhevsky began using CUDA on a GeForce GTX 580, which played a part in the deep learning breakthrough.

1-2. "Annual innovation" created by the full-stack

"If you don't own the entire stack, it's impossible to innovate every year. Because we own the entire stack—from CPU, GPU, and NVLink to Spectrum X—we can change everything every year."

While competitors are "stitching together cats and dogs," NVIDIA can achieve annual updates through integrated design.

2. The three inflection points of AI—Generation, Inference, and Agents

2-1. First inflection point: Generative AI (ChatGPT)

GPT-3 was in a state of "being visible but not seen" for several months. The moment someone wrote an API wrapper and turned it into ChatGPT, the world realized it. However, generative AI had the problem of "hallucinations"—it generated content without context or evidence.

2-2. Second inflection point: Inference AI (O1)

With the arrival of O1, "evidence-based generation" became possible. The ability for self-reflection and self-correction was added, increasing the reliability of information. As a result, token generation volume increased by about 100 times compared to the first generation, model size increased by 10 times—and the computational load increased by approximately 1,000 times. Furthermore, with the explosive increase in the number of users, demand reached millions of times higher.

2-3. Third inflection point: Agent AI (Open Claw)

"OpenClaw is probably the most important software release to date," Huang said. It took Linux 30 years to reach this level of downloads, but OpenClaw surpassed it in three weeks. It has become the most downloaded open-source software in history.

The nature of prompts has changed. Previously, they were question-based, such as "What is...?" or "When is...?" Now, they are action-based, such as "Create," "Do," or "Write." "The previous prompt was a query. The current prompt is a task."

Within NVIDIA, multiple agents (which Huang called "claws") are constantly running in the background, writing code and developing tools. As a result, the amount of compute required within the company has surged.

3. The concept of an "AI Factory"—the difference from data centers

3-1. Factories produce tokens

For several years, Huang has taken issue with the term "data center." "It is not a facility for storing data, but a factory for producing tokens. Factories are profitable. That is why everyone wants to build one."

And now, the correlation between compute and token production/revenue has been proven. "We can say with certainty now—corporate revenue is directly linked to compute. If OpenAI could increase its compute right now, its revenue would increase."

3-2. Tokens/Watt 10x better than the competition

In a thorough benchmark by Semi Analysis, NVIDIA was certified as the "Inference King." It is said to outperform the competition by one order of magnitude (10x) in tokens/second, tokens/watt, and tokens/dollar.

"If your factory has 1 gigawatt and your tokens/watt are 10 times better than the competition, your revenue will also be 10 times higher. For the first time, we have entered an era where the choice of computer architecture within a factory must pass a CEO review."

3-3. Constraints are a tailwind

Regarding infrastructure constraints such as power, permitting, and a shortage of electricians, Huang stated, "I love it." "In a world with constraints, you have no choice but to choose the best. You cannot spend $50 billion just to try things out. Because you know that revenue is directly linked to the following year, you have no choice but to make the right decision."

He says NVIDIA's strong balance sheet is a strategic asset for securing the supply chain. Because they have secured everything in advance—memory, wafers, packaging, and cables—when Satya asks them to build a gigawatt-scale AI factory, they can answer, "No problem."

4. OpenAI investment, MSL, and Physical AI—the next wave of demand

4-1. $30 billion investment in OpenAI confirmed

After clearly stating that "there is no $100 billion investment opportunity," Huang said, "We have confirmed a $30 billion investment in OpenAI. They are heading toward an IPO. This may be the last opportunity to invest before they go public."

He also revealed that OpenAI's computing capacity is being expanded from Azure to OCI and AWS. He noted that capacity shortages are a constraint on revenue, and they are rushing to expand supply.

4-2. MSL—Meta's new AI lab

While they have a long-standing relationship with Meta, he says MSL (Meta Superintelligence Lab) is "net new" demand. "It requires millions of GPUs, and that is on top of Meta's existing demand."

4-3. Physical AI will define the next decade

"Two years from now, we won't be talking about agentic AI because we will already be using it normally. By then, we will be talking about physical AI."

NVIDIA has grown Cosmos into the world's most downloaded physical AI model. Alpamayo for autonomous driving and Gr00t N2 for humanoid robotics also have the highest download counts in the world. La-Proteina for digital biology is also being deployed.

"NVIDIA is defining the frontier of physical AI. We are releasing it completely open-source so that every industry can utilize this capability."

They also announced the establishment of a joint innovation lab with pharmaceutical giant Eli Lilly.

Summary—The proposition that "Compute = Revenue = GDP"

If Huang's message could be summarized in one phrase, it would be "Compute is GDP." Huang stated that this is not a metaphor, but a fact already manifesting in a verifiable way.

The entire software industry (currently a $2 trillion sector) will become token consumers, all CapEx from internet companies can be converted to AI, and every nation will require AI—if this structure holds, the ceiling for demand is still far off.

"No country will say in the future, 'We don't need intelligence.' If you need intelligence, you need compute. And compute will become GDP—that is certain."

Recommended Articles

Next Big Wave—Growth Stocks, Seeds of Ideas, and Deep Dives into Trends