Power and Cooling are the Main Battlefield, Not GPUs: The Essence of AI Investment and Misconceptions About the Bubble
Is the AI boom a "new dot-com"? From an interview with Martin Casado, a General Partner at Andreessen Horowitz, we organize the reality of infrastructure investment, how to view the bubble theory, ROI outlooks, and a map of investment opportunities. He begins by emphasizing that while "the late 90s were chaotic, with limousines, parties, and even janitors wanting stock options," the current situation is different from the 'true bubble' of that time. This article breaks down and explains the professional content while citing his remarks.
1. Where is the money going? "Infrastructure" is the main act
Casado asserts that the bulk of investment is going into "GPUs, data center real estate, power, and cooling (HVAC)." While there are labor and software costs, the majority of spending is on foundational infrastructure. He defines his own focus as investing in what could be called "tools for building apps," targeting computer science infrastructure (compute, networking, storage, DB, DevTools, security, and AI models).
2. What distinguishes a "bubble"—the decisive difference from the dot-com era
2-1. Differences in funding sources and debt structure
Around 2000, there was a fragile financial structure symbolized by WorldCom's massive debt (approximately $40 billion). Furthermore, an exogenous shock like 9/11 occurred. In contrast, current AI investment is "driven by giant tech companies with hundreds of billions of dollars in cash using their own funds." He states that the "fundamentals of who is providing the capital are completely different."
2-2. Maturity of demand and business models
Back then, "there were many users, but payment methods and models were not established." Now, revenue engines like advertising, SaaS, and cloud are established. Citing Meta's investment example (a mention of "up to $600 billion by 2028" was introduced in the interview), he describes growth that does not rely solely on new demand by "reallocating internal budgets from existing businesses to the AI side."
3. Thinking separately about "speculative valuation" and "systemic collapse"
Casado distinguishes between "bubble = speculative overvaluation" and "chain reaction collapse of the financial system." Valuations "go up and down," but that does not immediately mean systemic risk. Even if "red flags" are discussed in the short term, such as "AI revenue needs to be 40 times higher by 2030" or "$1 trillion for data centers," his position is that the foundation is different from the late 90s when viewed through the lens of "is there over-investment relative to long-term demand?" and "is there the capacity to endure?"
Quote: "A speculative bubble is possible. However, that does not immediately indicate a systemic problem".
4. When will ROI be visible?—"Scaling while private" has become the norm
"The best companies are no longer going public." Private capital has become abundant, and the path to "securing liquidity by selling to subsequent late-stage investors" has expanded. As a result, the classic model of IPO = exit has faded, and a new challenge of long-term private market exposure has emerged for both LPs and VCs. The gist is that "liquidity hasn't disappeared, but the way of thinking has changed."
5. Where are the opportunities?—"Core LLMs" and the "Long Tail"
The market map has a two-layer structure.
Core: SOTA LLM groups (OpenAI, etc.). Massive capital is essential.
Long tail: Many areas such as image, video, audio, and music generation. Casado states, "There are many promising companies here too," and says he is investing broadly. While past AI was limited to "20% improvements" and lacked economic viability, the structural shift is that "generative AI changes 'action' itself and brings an 'order of magnitude' better experience value compared to the past."
6. Culture drives technology—"A coffee pot becomes Netflix"
The "Hamster Dance" of the 90s or the live stream of the coffee pot at the University of Cambridge seemed like "play" at first. However, YouTube and Netflix were born as an extension of that. Casado says, "New behaviors that seem trivial are the prototypes for future giant industries." Generative AI use cases that look like "anime or play" now could also become the mainstream of the future.
Even if there are speculative waves, the quality of capital, the demand base, and the business models are decisively different from the dot-com era. That is why, rather than being swayed by short-term "frenzy," steadily building up "implementation capabilities starting with power and cooling," "long-tail apps," and "defense design" is the shortcut to obtaining sustainable returns in the AI cycle.