Could the AI bubble pop?

Key Concepts

• Artificial General Intelligence (AGI) / Superintelligence: AI that possesses human-level intelligence and can self-improve, potentially leading to a "winner-takes-all" scenario.
• Bubble vs. Race: The current AI investment landscape is debated as either a speculative bubble or a critical race to develop AGI.
• Winner-Takes-All Market: The belief that the first company to achieve AGI will gain a monopoly and control the future.
• Infrastructure Left Behind: The concept of tangible assets (like railway tracks or power grids) that remain useful even if the originating companies fail, contrasting with the obsolescence of AI hardware.
• Open Source vs. Commercial AI Models: The availability of free, open-source AI models potentially undermining the value proposition of expensive commercial models.
• Value Capture: The uncertainty surrounding which entities will profit from the AI revolution, especially with the rise of open-source alternatives.

The AI Investment Landscape: Bubble or Race?

The discussion centers on the unprecedented pace of investment in Artificial Intelligence (AI), with some, like Sam Altman, suggesting it could be a bubble. Despite widespread acknowledgment of this possibility, investment continues unabated. This phenomenon is attributed to the belief that the current AI market might not follow traditional economic patterns, potentially being both a bubble and a race.

The AGI Prize and Valuation Challenges

A key driver of this intense investment is the belief held by leaders in AI labs that the first company to develop powerful AI, specifically Artificial General Intelligence (AGI) or "human-level intelligence," will "own the future." This potential outcome is so transformative that traditional valuation methods become difficult. One cited figure for the potential market value of building AGI or superintelligence is a staggering $1.46 quadrillion.

The "Winner-Takes-All" Argument and Counterarguments

The "winner-takes-all" hypothesis suggests a single company will achieve a monopoly in AI. However, the recent history of the industry shows a constant shift in leadership among AI developers, with many labs appearing to be "neck and neck." This dynamic suggests that dropping out of the race is undesirable, as winning could grant a monopoly.

Conversely, the argument is made that a monopoly outcome is less likely. An analogy is drawn to the telecom boom of the early 2000s, where companies like Global Crossing and 360 Networks, once vying for supremacy, are now largely forgotten. This historical precedent questions the assumption of a single winner.

Defining Superintelligence and its Implications

The concept of superintelligence is defined as an artificial intelligence that surpasses the intelligence of any human ever and is capable of self-improvement, leading to an exponential "take-off." If such an AI is developed, the theory suggests that the first company to achieve it would immediately develop a second, even more advanced superintelligent AI.

Probabilistic Valuation and Lottery Tickets

The immense potential value of AGI, even with a low probability of success, can justify significant investment. For instance, a 1 in 1,000 chance of a company becoming a $1.46 quadrillion entity still translates to a potential $146 billion valuation. This probabilistic thinking, coupled with the allure of controlling the future of technology, can lead investors to view their investments as "lottery tickets" rather than traditional financial instruments, potentially fueling a bubble.

The "AI Dust" Phenomenon

Similar to the crypto boom, there's a concern that simply adding "AI dust" to an unimpressive company can attract significant investment, indicating a speculative phase of the AI boom.

The Legacy of Technological Booms: Infrastructure vs. Obsolescence

A crucial distinction is drawn between past technological booms and the current AI surge, focusing on what tangible or lasting infrastructure is left behind.

Historical Examples of Lasting Infrastructure

• Railways: The construction of early railways, like the Greenwich to London Bridge line built in the 1830s, left behind durable brick arches. While the companies that built them may have failed, the physical infrastructure remained useful.
• Electricity: The electricity boom resulted in the construction of grids and power stations, which continued to be utilized even if the initial investors or companies disappeared.
• Telecommunications: The fiber optic cables laid during the telecoms boom, despite companies going bankrupt, eventually became the backbone of modern communication networks.

The Obsolescence of AI Hardware

The question arises whether AI leaves behind similarly durable infrastructure. The primary physical components of AI are data centers filled with chips. However, these chips, particularly GPUs (Graphics Processing Units) like those from Nvidia, have a rapid obsolescence rate.

The Cost Breakdown of Data Centers

A rough estimate suggests that approximately 50% of the cost of building an AI data center is attributed to computing power (chips), with concrete and other infrastructure being less expensive.

Limited Reusability of AI Hardware

While GPUs have roots in computer gaming, AI data centers are not easily repurposed. They are optimized for AI tasks, often lacking the necessary CPUs and interconnectedness for functions like global streaming game consoles.

Potential Niche Repurposing

There are some AI-esque use cases outside of general AI development that could potentially utilize this hardware, such as weather modeling and complex financial trading. In the event of an AI bubble collapse, these industries might benefit from years of "really, really cheap" computing power.

The Ethereal Nature of AI Models and Open Source Competition

Beyond physical infrastructure, the value left behind in AI is also considered to be the models themselves and the expertise embedded within them. However, this intangible aspect presents unique challenges.

The Rise of Open Source AI

A significant concern for investors in commercial AI models is the proliferation of high-quality, open-source alternatives. Companies like Meta (with its Llama series), Alibaba (with its Quen series), and Deepseek (with its DeepSeeks models) are releasing powerful AI models freely.

Undermining Commercial Value

The existence of these capable open-source models raises the question of why consumers and companies would pay substantial amounts for expensive commercial AI systems when comparable or nearly comparable open-source options are available. This is a "trillion-dollar question" for the industry.

Uncertainty in Value Capture

While it's widely accepted that AI will transform the world, it remains unclear who will capture the value generated by this transformation. It's even uncertain if the companies actively developing AI will be the ones to profit.

A Transformed World Without Direct Payment

The scenario is envisioned where almost every industry is transformed by AI, but the users of these AI systems do not pay the developers for them, due to the widespread availability of effective open-source models. This contrasts sharply with previous technological booms where tangible infrastructure or proprietary technology created clear avenues for value capture.