Written by: Techub News Compilation

Introduction

In the latest episode of investment podcast host Antonio Linares, he invited Thomas, an investor renowned in the community for his deep technological insights, to engage in a significant conversation about the future of AI infrastructure. This episode primarily focuses on a critical question: to what extent can AMD capture a significant share of the AI inference market? Thomas, noted for his in-depth research into the semiconductor industry, particularly AMD, systematically articulated his investment logic during the conversation and provided high praise for AMD's leader Lisa Su. In the current context of fierce competition in AI chips and an uncertain market landscape, this conversation offers valuable perspectives on understanding the trends in underlying hardware.

Summary

• AMD's Chiplet architecture grants it unparalleled flexibility, potentially allowing it to capture at least 50% of the future AI inference market.
• The value chain of AI is divided into five layers: Energy/Infrastructure, Data Center Construction, Chips, Base Models, and Applications. Currently, the value of the infrastructure layer has yet to be fully realized.
• The development of AI will lead to CPU shortages, and AMD, due to the close collaboration between its CPU and GPU architectures, is poised to be a major beneficiary.
• This is not an AI bubble; real value creation is occurring, especially for "creators" who can significantly enhance productivity through AI.
• AMD CEO Lisa Su is praised as the best non-founder CEO globally, with her strategic acuity and organizational culture shaping the company's success.

The Five-Layer Value Chain of AI Infrastructure and AMD's Positioning

At the beginning of the conversation, Thomas first mapped out his view of the AI value chain. He believes that since ChatGPT ignited the market, the opportunities in AI can be clearly divided into five levels: Energy and Infrastructure, Data Center Construction (excluding chips), Chip Layer, Base Model Layer, and Application Layer. Initially, his attention was focused on the model layer and application layer as the largest beneficiaries. However, through deeper research, he realized that the value of the infrastructure layer, particularly the chip layer, has yet to be adequately priced by the market and is still in a very early stage.

At that time, the market generally viewed Nvidia as the undisputed winner. However, Thomas pointed out that this perspective is not entirely accurate, and the market landscape is far from being defined. This prompted him to shift his research focus to the infrastructure domain and develop a distinct investment thesis around AMD.

AMD's Core Argument: Modularization to Win the Inference Market

Thomas's investment thesis for AMD can be summarized in one sentence: The flexibility of chiplet infrastructure will enable AMD to capture at least half of the AI inference market.

He explained that higher modularization allows chips to be optimized at the circuit level. When inference (running neural networks) is needed on edge devices, if the chip can adapt to the physical form of the network, it can optimize the flow path of electrons in the circuit. From a physical first principle perspective, if competitors cannot achieve similar optimizations at the circuit level, it will be challenging to compete.

The mainstream narrative often equates AI with training in large data center clusters. However, with the development of humanoid devices like autonomous driving and robotics, the argument for "Edge AI" is becoming increasingly credible. In edge scenarios, the requirements for latency, energy efficiency, and customization are extremely high, which is precisely where AMD's chiplet architecture shines.

Thomas further categorized the chip market into three types: the GPU market dominated by Nvidia and AMD; the dedicated processor market led by Broadcom and Google TPU; and emerging players like Groq and Cerebras that heavily use SRAM. He is more bullish on GPUs as they offer the highest flexibility. Among the GPU camp, AMD possesses stronger flexibility advantages due to its modular design.

He added that while Google's Gemini and Anthropic's models currently lead in specific areas, they primarily run on Google TPU or Trainium chips. One significant drawback of TPU is the difficulties in workload switching. Once configured for Transformer models, it becomes challenging to quickly switch to diffusion models or other new architectures. In contrast, GPUs, especially AMD's, provide flexibility in handling future algorithm changes.

CPU Shortages and the Future of Edge Computing

In addition to GPUs, Thomas presented a striking viewpoint: A major CPU shortage is imminent. He believes this is one of the reasons behind AMD's recent stock price increase.

He explained that many workloads in the currently popular agent architecture are running on CPUs. For example, operations like web searches, tool calls, and equipping AI models with skills all depend on CPU architecture. Therefore, GPUs and CPUs need to work closely together; the closer they are physically, the better. With the explosive growth of agent workloads, he anticipates a shortage in CPU demand.

Another advantage for AMD is its ability to network CPU and GPU cooperation, with outstanding architectural flexibility. In a shortage environment, having pricing power is crucial, and Thomas believes AMD is well-positioned to benefit from this.

Looking ahead, Thomas believes computation will exhibit a distributed trend, not solely concentrated in massive data centers or purely edge devices. Building medium-sized or even small data centers in specific areas for inference or training may prove more advantageous. AMD's technology is being used to construct this completely distributed computing infrastructure, giving it more opportunities in the future training and inference markets.

The AI Bubble Debate and Real Value Creation

When asked whether we are currently in an AI bubble, Thomas answered negatively. He believes the loudest voices declaring a "bubble" often come from ordinary users who only use ChatGPT or Copilot for simple Q&A. They feel disappointed that GPT-5 did not include features like emojis and have not run any real AI workloads. From their perspective, it indeed seems like a bubble.

However, for those engaged in truly creative "white-collar labor," the situation is entirely different. Whether engineers coding with the latest cloud models or designers using creative tools, they see that the law of scaling has not stopped; the work completed by AI is becoming increasingly autonomous, accurate, and powerful. Thomas, as a deep user himself, subscribes to all mainstream AI services and compares them daily, experiencing firsthand the rapid advancement of AI capabilities.

He cites his software engineers as an example; they now use AI coding tools to run multiple terminals and agents simultaneously, likely increasing productivity by 100 times or more. AI primarily benefits "creators," and jobs like software engineering, which can create infinitely, benefit significantly. He quotes Anthropic co-founder Dario Amodei: "Every task that software engineers do will be automated within 12 months." If this argument holds, its impact on the socio-economy would be disruptive.

Lisa Su: Vision and Execution of the Best Non-Founder CEO

When discussing AMD's management, both Thomas and Antonio spoke highly of CEO Lisa Su. Thomas even stated that Lisa Su is the best non-founder CEO in the world today, surpassing well-known managers like Tim Cook.

Antonio added details about Lisa Su's leadership: when she took office, AMD was in a difficult position, and Intel was a giant opponent. She defied the odds and fully committed to the chiplet architecture that her predecessor had developed for two years, establishing it as the company's core strategy for the future. At the same time, the only message she repeatedly emphasized to the outside world was, "We will work closely with our partners." At that time, analysts deemed this statement hollow.

However, it is this technical decision to fully commit to chiplet architecture, combined with the organizational culture of “close collaboration with partners,” that produced a tremendous multiplier effect. The chiplet platform provides unparalleled flexibility, while deep iterations with partners ensure that products precisely meet market needs. This combination, seemingly simple but extremely difficult to maintain, has created immense value for AMD. Antonio recalled buying AMD at a stock price of 4.2 USD (with a market cap of about 30-40 billion USD); today, the company's market cap has surpassed 400 billion USD. Lisa Su consistently makes low-profile commitments and speaks cautiously, even in the face of skepticism during the most challenging times; this trait is particularly valuable among CEOs.

Thomas summarized that making such a significant strategic bet requires a founder-like mindset and determination, and Lisa Su has accomplished this. He predicted that in the next five years, this point will become clear to everyone.

Beyond Chips: The Long-Term Impact of AI on Socio-Economics

The latter half of the conversation transcended specific stock analyses, exploring the more grand and profound socio-economic transformations that AI may bring. Both parties agreed that the process of value creation is accelerating at an unimaginable pace. Elon Musk previously predicted that companies with valuations exceeding 100 trillion USD will emerge in the future, which sounds incredible but is not impossible given the ongoing effectiveness of the AI scaling law.

Thomas presented an intriguing viewpoint: the value ultimately created by AI may be "less than" some people's expectations of singularity or superintelligence (i.e., there will not be "God-like" AI), but the actual scale of value creation it brings will far exceed everyone’s most optimistic estimates, possibly being ten times or more than during the internet bubble period.

Regarding wealth distribution, Thomas maintained a relatively optimistic attitude. He believes that the immense value created by companies like Meta and Google over the past 20 years has benefited ordinary people very little. However, AI is commodifying "knowledge" itself, and everyone will benefit from it. Future companies with valuations in the trillions may not be because they monopolize short-term attention, but because they provide universal healthcare, universal legal services, or allow everyone to create their homes on demand.

Antonio posited that the future economy may revolve around "self-actualization." Once machines take over most repetitive production and service work, humans can dedicate more time to areas machines cannot replace: empathy, social interaction, art, exploration, etc. He even mentioned the combination of biology and AI at the edge (e.g., health monitoring and intervention), believing this will be a larger market than humanoid robots because it directly addresses humanity's most fundamental needs—health and longevity.

Finally, both acknowledged that the future is filled with uncertainty, with changes potentially being destabilizing. However, as investors, in the absence of certainty regarding the ultimate application layer winners, investing in the "shovels"—the chips that everyone needs—is a logically sound choice. And AMD, with its dual advantages at the physical layer (circuit optimization) and organizational top layer, is becoming the sharpest and most flexible of these "shovels."

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。