Yesterday, AMD CEO Lisa Su made a judgment in an interview: the real big cycle of AI has arrived, which is in the 10th grade and driven by the demand side, not the concept hype of 2021-2023 The direct background that triggered her judgment is: Anthropic has just announced plans to invest $5 billion in building AI infrastructure in New York and Texas, but Su Zifeng has stated that data center computing power demand will be conservative at 60%+CAGR over the next 3-5 years, and this is just an appetizer. Why does she dare to speak so confidently? Because she is now witnessing with her own eyes: Those super large customers (including several AI native unicorns and players of OpenAI's scale) who originally used cloud services and NVIDIA solutions for the entire stack have begun to carefully evaluate and import AMD's solutions. Note that it is not because AMD is cheaper, but rather because of its lower overall TCO (Total Cost of Ownership), smoother upgrades, and more flexible architecture. These customers are not talking about 'trying water next year', but about signing infrastructure expansion roadmaps for 3-5 years or even longer. AMD's next-generation data center products, such as the Helios rack level solution, are all aimed at "being able to withstand 10 years of continuous climbing" in terms of performance, power consumption, cost, and delivery flexibility. She specifically mentioned a detail: insisting on retaining a considerable proportion of domestic manufacturing capabilities in the United States (even if mainly TSMC OEM), which is clearly driven by supply chain resilience and geopolitical security considerations. She repeatedly emphasized a core point: This wave of computing power frenzy is not about piling parameters, brushing rankings, burning money for financing, but about truly serving more users, generating real productivity gains and revenue. In other words, the demand side finally has a sustainable business closed loop, rather than a pure subsidy game. Several important trends can be inferred from her statement: 1. The concentration of AI infrastructure will significantly decrease In the past few years, money and computing power have been concentrated in the hands of a few Hyperscalers and several large model companies. In the future, with the improvement of single watt performance, the decrease of TCO, and more flexible delivery forms, more and more medium-sized players, industry customers, and even sovereign AI projects will build or co build their own clusters, and the market will be significantly more diversified. The competitive landscape is being rewritten, but not zero sum Nvidia's leading advantage still exists, especially in training, but in scenarios such as inference, large-scale hybrid deployment, and long-term maintainability, AMD's cost-effectiveness has begun to impress customers. The cake is big enough for everyone to eat. 3. Geopolitical factors have officially become primary variables The frequency of words such as "local manufacturing," "supply chain security," and "energy independence" is increasing, and the underlying logic of technological competition has changed. If Su Zifeng's judgment is valid, then in the next 10 years, AI infrastructure will be a highly deterministic super long slope and thick snow track. Not only chips, but also the entire chain - electricity (nuclear/SRM), liquid cooling/submerged cooling, high bandwidth networks, software scheduling layers, and even data center real estate - will face sustained demand for real money. For investors, this means: The AI infralayer still has a huge opportunity to generate real cash flow rather than relying on stories to survive, far from reaching the ceiling.