This year, AI officially transitioned from the training era to the inference era.

NVIDIA launched the Rubin/Rubin Ultra chips at the GTC conference on March 16, and previewed the Feynman architecture for 2028 in advance, focusing on inference solutions based on GPU+LPU architecture. The goal is to reduce the cost per token of AI inference and open up the inference market space.
Currently, NVIDIA's market value fully reflects the value of training. After launching the inference solution, if validation is successful, the market will assign a "training + inference" dual valuation to NVIDIA, potentially breaking through the current $5 trillion market cap range and continuing upward.
Against this backdrop, what are the opportunities for excess returns in the industry, and what opportunities can ordinary people participate in? I think this is a question many of us must consider by 2026.
I summarize the investment logic for 2026 in eight characters:
Heat dissipation
Packaging
Power consumption
Yield rate
Each sector harbors the potential for excess returns, but many opportunities are fleeting; delaying them means higher costs.
During the recent US-Iran conflict, there were very suitable opportunities for fear-based investing, but unfortunately, X cannot speak publicly and cannot share information in time. Fortunately, some opportunities are still undervalued.