Huang Renxun Put the water bottle on the table and lean back on the chair again.
His leather jacket made a slight noise as it rubbed against the leather of the chair's back.
"You know what surprised me the most?
It's not like you dare to buy it when it's more than 120 US Dollar.
Before you buy, you have already prepared 13D declaration, Section 16 compliance, and even how to communicate with me - everything is ready."
He picked up the water bottle but didn't drink it.
I just used my palms to feel the coolness left by the condensed water droplets on the bottle wall as they slid down.
"Then, during the most difficult time, you carried the long positions of two companies by yourself.
Facing the joint sniper attack from Wall Street, the strongest Short-Seller Alliance, I still have the energy to sit down and analyze calmly."
He closed the manual and pushed it back in front of Zhao Chen.
Then he stood up and walked to the whiteboard.
He picked up a black marker and wrote two numbers next to the curve--
One is the lowest range of Nvidia stock price in November 2018, and the other is the latest stock price in June 2020.
Between the two numbers, he drew a long arrow and wrote three words next to the arrow:
Long-termism.
Huang Renxun After writing the three words "long-termism" on the whiteboard,
Put the marker back into the pen tray, turn around and lean on the edge of the table.
Outside the window California the sunlight is pouring down from the top of the triangular glass curtain wall,
Cut out a bright parallelogram light spot on the conference table,
The edge just fell on the cover of the Henglong Research Institute manual in front of Zhao Chen.
"AI Computing Power demand has changed from a niche topic to a global consensus in the past two years."
He removed his hand from his chest and picked up the black marker again.
Draw a larger arrow on the right side of the whiteboard in an area that is not yet full.
The starting point of the arrow says "December 2018"--
In that month, Nvidia's stock price fell to the bottom, SoftBank liquidated its position, and Goldman Sachs removed Nvidia from the list of believers.
He left a blank space at the end of the arrow and did not fill in the year.
"In December 2018, the total computing power demand for global AI training tasks was probably increasing several times every year.
This number seems abstract to many people.
But in our industry, it's a very specific physical quantity --
How many GPU hours are consumed in neural network training every day?
How many servers are running CUDA code in each cloud service provider's computer room?
At that time, we made an internal prediction: in the next ten years, this number will increase millions of times."
He tightened the cap of his pen, turned around,
"Now I want to hear your judgment.
It's not Henglong Research Institute's Model—it's you. Where do you think AI will go next?"
Zhao Chen stood up and walked to the whiteboard, taking the black marker from Huang Renxun's hand.
A few words were written next to the end point of the arrow drawn by Huang Renxun:
"Artificial General Intelligence(AGI)."
"The transition from specialized artificial intelligence to Artificial General Intelligence.
In the past, AI could only do one thing - playing Go could not translate languages.
Those who translate languages cannot write code, and those who write code cannot recognize medical images.
Each task requires a separate training of Model,
Each Model requires massive annotation data.
But in the past year or two, this paradigm has begun to loosen.
The appearance of Large Language Model proves one thing:
As long as the parameter scale is large enough and the training data is broad enough,
The same Model can master multiple skills such as translation, writing, programming, logical reasoning, etc. at the same time.
This transfer of capabilities is not gradual, but emergent -
After the parameter size exceeds a certain threshold, Model suddenly acquires capabilities that were never explicitly taught in the training data."
Huang Renxun picked up the marker in the whiteboard slot and put it down again.
His fingers rubbed back and forth on the pen cap, as if he were pressing the pins of a transistor that had been repeatedly checked.
"Emergence - the word you used is critical. We are also discussing the same phenomenon internally.
After the parameter scale exceeds a certain level, the behavior of Model can no longer be predicted linearly.
What does this mean?"
"This means that the growth curve of computing power demand is no longer linear, but super-exponential.
If a Model can only translate to write code,
The parameter scale needs to be expanded from hundreds of millions to hundreds of billions, and then from being able to write code to being able to do scientific research--
For example, independently propose a physical hypothesis, design experiments to verify it, and then revise the theory based on the experimental results--
The required parameter scale may be expanded by several orders of magnitude on the existing basis.
This is not a ten-fold or a hundred-fold increase, this is a ten thousand-fold or a hundred thousand-fold increase."
Zhao Chen drew a slightly steeper curve on the whiteboard, parallel to the growth line Huang Renxun drew before but rising higher.
"Both your internal Model and Henglong's independent Model predict that computing power demand will grow millions of times over the next decade,
But this is a deduction under the existing AI paradigm.
If AGI does come true, millions of times may be just the starting point."
Huang Renxun stared at that curve for a long time.
Meeting Room was so quiet that the only sounds left were the very low air flow from the air conditioner outlet and the faint hum of the fan from the distant server room.
He suddenly turned around, walked to the whiteboard, picked up the red marker, and drew a steeper line next to the curve drawn by Zhao Chen.
The two lines are like two stretched bows, almost shooting out of the edge of the whiteboard.
"This is our internal Model, which has only been shown to the core technical team.
After AGI is implemented, we have no answer to what shape the computing power demand curve will change from super-exponential.
But one thing I'm sure of:
No matter what shape this curve eventually becomes, Nvidia will not just make one chip Suppliers.
The future of AI is not a hardware race - hardware is a means, not an end.
The real value of Nvidia is that we are building a Accelerated Computing Platform that covers the full stack of AI.
From the underlying GPU architecture and network interconnection to the middle-layer CUDA software ecosystem and Inference Engine,
Then to the upper level AI application framework for specific industries.
The hardware provides computing power, the software ecosystem allows computing power to be called upon, and the upper-layer framework allows computing power to truly solve industry problems.
Every turning point on this curve requires the entire technology stack to be upgraded at the same time--
The chip architecture needs to be redesigned, the network interconnection needs to be rewired, and the software ecosystem needs to be recompiled."
He put the red marker back into the whiteboard slot, turned around,
Those eyes that have been immersed in the ocean of GPU transistors for more than 20 years no longer have the same sense of separation when reviewing the prototype test data.
This is not the relief after confirming that Zhao Chen is a "shareholder who does not need to explain".
Rather, it is a deeper excitement ignited by the possibility of the future itself.
"I don't need Wall Street to tell me how much computing power demand will grow over the next ten years.
I'm the one making predictions myself.
But the problem with Wall Street is that they only look at the next quarter.
When the quarterly results fluctuated slightly, they asked Data Center whether its growth rate had peaked.
When the gaming business slowed down a bit, they asked if the AI story could no longer be told.
Their imaginations about the future are locked into a three-monthly reporting cycle.
But you are not.
Those public data you got when Nvidia's market value fell to the bottom,
The data in these quarterly reports are the same underlying logic -
CUDA ecosystem is expanding, and the computing power requirements for AI training tasks are accelerating.
Data Center is replacing games as the core revenue engine of Nvidia.
These trends will not be interrupted because of a quarter's performance fluctuations, nor will they slow down because of Wall Street's doubts."