Growing up, I didn't know how good Su Ma was, mistaking trash for treasure.

When DEEPSEEK came out, U.S. stocks' AI sector plummeted, AMD even dropped to 80, and in just a few months, it’s already at 300, which is truly speechless.
But there's no way around it; missing out is because I didn't understand.
If you don't understand, learn and practice.
In the past three years, most of the AI computing power has been spent on "training"—OpenAI training GPT-4, Anthropic training Claude, Google training Gemini, these are all training.
Training is characterized by being one-time, concentrated, and peak-intensive.
But every time you ask ChatGPT a question, every time you use Claude to write a piece of code, every time you generate an image with Midjourney—you are not consuming training compute power, but inference compute power.
Inference is characterized by being continuous, dispersed, and long-tail.
Once training is completed, the model is launched. After launch, it responds to hundreds of millions of user requests 24 hours a day.
After three months, the training compute power consumed is no longer visible in the ledger—all of it is inference.
This turning point's magnitude can be illustrated with a comparison.
In 2023, inference accounted for about 20% of AI compute expenditure; in 2024, this ratio climbs to 50%, and by 2026, it has already exceeded 55%, and continues to rise.
Some more aggressive predictions believe that by 2030, inference will account for 70-80%.
Note that this is not because training demand is shrinking—the absolute expenditure on training is still increasing—but because inference is growing much faster than training.
On this big slope of inference, NVIDIA truly sits at the top.
In NVIDIA's fiscal year 2026 (up to January 2026), data center revenue reached $194 billion, and two years earlier, this figure was less than $50 billion.
Such growth has never appeared in semiconductor history.
The CUDA ecosystem has five million developers, twenty years of accumulation, with both training and inference sides benefiting simultaneously—this is a true monopoly.
NVIDIA, as the leader, AMD as the second, and the third are Google TPU, Amazon Trainium, Meta MTIA—these self-developed ASICs—this is the current structure of the game.
What is AMD's position on this table? It is the second chair.
This chair is very important—without the second chair, the first chair has no bargaining power.
But the second chair is not the first chair.
So the real question becomes two sub-questions:
First, can AMD hold this second chair steadily for ten years?
Second, how much is this chair worth if it is held steadily?
In addition, AMD has a severely underestimated angle: the real story behind Meta's 170k-dollar MI300X.
"AMD Research Report: Looking back over 10 years, is $300 expensive?"