Author: Techub News Compilation

Recently, the astonishing capital expenditure by tech giants and the crazy surge in the stock prices of semiconductor companies have sparked widespread discussions in the market about the AI bubble and sustainability. In this episode of Bankless, hosts Josh and Ejaaz take on the role of "capital detectives," attempting to trace where nearly one trillion dollars in AI investments from major companies like Google, Amazon, Meta, and Microsoft have flowed within the industry chain. The significance of this conversation lies in the fact that it is not only an interpretation of stock price phenomena but also a systematic sorting out of the underlying capital structure and technological paradigm shift in the AI industry, providing a clear roadmap for investors to understand the current AI investment landscape.

The "Waterfall Effect" Initiated by Trillion-Dollar Capital

The episode opens with a series of astonishing data points: a year ago, if one had invested in SanDisk, the return would be up to 40 times; investing in Micron yields 8 times; Intel and AMD also experienced nearly fourfold increases. Behind these numbers is the market's frenzied pricing for AI infrastructure demands. However, the hosts point out that people may have previously misjudged the situation—this is not simply a bubble but an industrial revolution driven by real capital expenditures.

The core driving force comes from the top: the four tech giants Amazon, Meta, Microsoft, and Google. Ejaaz notes that the Q1 2026 earnings reports are a critical turning point. These giants have not only proven that the promised capital expenditure of $630 billion has led to tangible profits and revenue growth but have also raised their spending plans for 2026 to $800 billion, expecting it to reach $1.1 trillion next year. "This is just four or five companies. I have to emphasize how crazy this is," says Ejaaz. Importantly, this expenditure is not based on leverage or borrowing but comes from the companies' own cash flow or cash reserves, providing a solid foundation for the current prosperity that differs from past bubbles.

Josh cites renowned investor Brad Gerstner's viewpoint, suggesting "this time it's different." The key is that the massive capital expenditure is flowing from these giants to the infrastructure layer needed to build more AI use cases, while upstream AI applications (like Anthropic and OpenAI) are already generating billions of dollars in annual recurring revenue. As long as AI continues to create value, this flywheel can keep turning. Capital is spreading from the core (mega-corporations) to the outer concentric circles.

Capital Flow Chart: From Retail Revenue to Infrastructure Stack

To clarify the capital pathways, the hosts constructed a clear "capital waterfall" model:

• Level 0 (Retail Entry): OpenAI and Anthropic. They attract 99% of retail users (such as ChatGPT and Claude users) and are the starting point of cash flow. Although still private companies, their paying customers and enterprise users form the source of the capital waterfall.
• Level 1 (Platforms and Mega-Corporations): Capital flows from OpenAI and Anthropic to platforms like Google, Amazon (AWS), Microsoft (Azure), and Meta. They not only serve as model laboratories but also provide critical cloud infrastructure and distribution layers, distributing AI models to businesses, governments, and end users. This is the center of capital aggregation and redistribution.
• Level 2 (GPU and Semiconductors): This is the segment traditionally perceived as “the shovelers” in the AI gold rush, represented by NVIDIA. NVIDIA provides the massive GPUs needed to build AI from scratch and has thus become one of the most valuable companies globally (although it has been surpassed by Google at this point). Its value lies in providing the initial computing power engine for the entire industry.

However, the story begins to show a critical "narrative breakthrough" here. Capital has not stayed within the GPU layer but continues to spread downstream and to surrounding areas.

Paradigm Shift: The Rise in CPU Demand Driven by the AI Agent Era

A major trend shift is occurring: AI is transitioning from large language models (LLM) and chatbots to the "agentic" (Agentic AI) era. Ejaaz explains that AI agents refer to AI instances capable of autonomously executing a series of tasks without continuous prompting from users. You can imagine the market space for autonomous AI workers is extremely vast.

The "narrative violation" here is that the core component driving AI agents to autonomously use tools, coordinate, and think is no longer the GPU but the CPU (Central Processing Unit). Josh points out that GPUs excel at solving complex mathematical problems for reasoning, but coordinating these models and providing them with "mind" and "intelligence" to execute tasks requires CPUs.

Ejaaz supports this transition with historical data: during the GPT-4 era, training or reasoning an entire model barely required CPU. But now, the demand ratio between CPUs and GPUs has approached 1:1 (i.e., one CPU core corresponds to one GPU core), and this trend is expected to reverse in the next six months, with CPU demand surpassing that of GPUs. This is fundamentally why stocks for Intel and AMD have soared—decades of accumulated CPU technology suddenly found a huge goldmine in the AI era.

Josh outlines three phases of AI evolution: starting from LLM chatbots that only needed GPUs, to the phase requiring more tokens for "chain of thought" reasoning, and now to the agent era needing complex coordination. Each paradigm shift has led to a massive increase in the demand for computing resources (especially tokens), and the core feature of the current phase is a sharply rising dependency on CPUs.

Memory: The "Achilles' Heel" of GPUs and Covert Monopoly

If CPUs are the "brains" of AI agents, then memory is the "lifeline" for the entire AI computing power. Josh points out that memory accounts for 50% of the BOM (bill of materials) for GPUs, making it the most critical bottleneck in the current supply chain. "Without memory, everything doesn’t work," he says. This also explains why memory stocks have performed so crazily over the past year: SanDisk increased by 40 times, Micron increased by 7.25 times.

Ejaaz reveals the "covert monopoly" pattern in the AI memory sector. High-end AI memory (High-Bandwidth Memory, HBM) is primarily dominated by three companies: Micron (USA), SK Hynix (South Korea, the largest supplier), and Samsung. This type of memory is crucial for processing large models with trillions of parameters like GPT-5.5, as it enables lightning-fast data access. However, its manufacturing process is exceedingly complex, and supply is extremely tight, with orders backed up until the end of 2028.

Another type of memory is NAND flash, dominated by SanDisk. Its function is to provide "temporary storage," allowing AI models to maintain longer context windows and remember earlier content in conversations. Although not as "sexy" as HBM, it is still an essential foundational component.

An interesting "paradox" has arisen: when companies like DeepSeek release more efficient models that use less memory, it was expected that memory demand would decline. But in fact, the opposite is true. Ejaaz uses the "Jevons Paradox" to explain: efficiency improvements lead to lower costs, which in turn stimulate larger demand. The architectural changes in DeepSeek v4 (using more agents for thinking) actually increased the overall consumption of NAND flash. Therefore, enhancements in memory efficiency haven’t stifled demand; on the contrary, they have pushed up total demand by unlocking more application possibilities.

Next Stop: The Challenges of Power and Infrastructure

The capital waterfall continues to flow down to Level 5: Power and Infrastructure. Once massive GPUs and CPUs are deployed in data centers, a fundamental issue arises: how to effectively power and manage them?

Ejaaz references a report pointing out that the XAI cluster, which has over a million high-end NVIDIA GPUs, has a power utilization rate of only 11%. The bottleneck lies not with the GPUs themselves, but with the lack of sufficient generating capacity or chip architecture to effectively deliver power to the GPUs. This indicates that the next flow of capital will likely be towards power supply and infrastructure providers, such as General Electric (GE) and Constellation Energy, which not only provide power but also ensure that it is delivered to CPUs and GPUs at the right time and with the correct capacity, and address cooling issues.

Josh mentions an even more fundamental Level 6: Raw Materials. Extracting "thought" from sand (silicon) requires large amounts of basic materials. For instance, the price of lithium carbonate soared from $75,000 to $187,000 in a short time, a staggering increase. This layer represents the most fundamental physical cornerstone of the AI stack.

Prosperity and Warnings: Is This Time Really Different?

After depicting an exciting capital landscape, the episode returns to rationality, exploring potential risks. Ejaaz notes that the memory industry has historically gone through numerous cycles of boom and bust, with the chart illustrating dramatic fluctuations. Every historical downturn has eliminated a group of players, reducing the number of HBM suppliers from 14 to the current 3.

So, will this time be different? Ejaaz reveals that he has listened to podcasts featuring executives from several memory companies, all of whom unanimously stated, "This time it's different." The reasons are: first, AI has brought an unprecedented surge in demand; second, the demand side has real revenue support (AI products are profitable); third, the current investments are not based on leverage. The $800 billion or even over a trillion dollars in capital expenditure comes from the existing cash flow of companies.

Josh concludes that this could be a significant paradigm shift in the past one or two decades. The capital that has historically accumulated at the top of giants like Google and Amazon is now flowing downstream in the industry chain. For investors, the key question is: where will the capital ultimately flow? Who will be the winners in this capital expenditure feast? Semiconductors and memory have already performed excellently; will they continue? Will power, infrastructure, and raw materials become the next hot spots?

Ejaaz's final thoughts are quite representative: the current market indeed shows signs of a bubble, but this is a "different type of bubble" due to the lack of leverage support. Every quarter's financial report will serve as a litmus test for its health. And the data from the first quarter indicates that money is indeed being made tangibly. The narrative may change in the field of AI in an instant, but at least currently, the capital waterfall driven by real cash flow is reshaping the entire landscape of technology investments.

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。