The AI Bubble: Real Utility, Excessive Expectations
Nearly half of the S&P 500 has turned into a bet on AI and this presents a systemic risk to investor portfolios. In this series we’ll dive into why AI, despite its utility in certain domains, can’t live up to the sky-high earnings expectations.
As an investment manager, programmer, and overall tech enthusiast, I find my opinion about new technologies can starkly depart from those held by others in my field. Artificial Intelligence is a broad field, but when most are talking about AI they are actually specifically talking about Large Language Models (LLMs) and other forms of generative AI like diffusion models for images. So I am mostly going to focus on LLMs throughout this series.
To put it simply, Large Language Models are a technology that is both groundbreaking and fundamentally limited. LLMs provide real utility within specific domains, but the scale of the productivity gains, the breadth of utility, the adoption timelines, and the profitability of providing models do not match the hype.
The markets are finally catching on that the LLM craze might have some trouble capturing the profits that many expected, but understanding the why of that is far more important than simply forecasting what might happen next.
This will be a series of posts on the current state of AI, its effectiveness, its drawbacks, how the market for AI may play out in the future, and a reflection on the history of financial bubbles.
We will cover these topics:
- Market Expectations
- With billions of users already using their services and few areas of natural expansion left, tech giants have gone all-in on AI as their next big driver of growth.
- Around 45% of the S&P 500 has direct exposure to bets on AI. The projected revenue growth required to sustain those valuations is monumental.
- Much of the US’s current GDP growth is related to the buildout of data centers.
- What happens if the reality of these data centers does not match the projected profitability? And just how much revenue do they really need to succeed?
- Effectiveness of Models
- One of the primary domains that LLMs are touted for is coding. To non-programmers, LLMs seem to have cracked the problem, but for those who care about software quality the results can be… lacking. So how effective are LLMs in the programming space, where do they shine, and where do they fall flat?
- Model Scaling
- We have seen huge improvements in models as they have scaled them up, but they already have the sum total of human-created data in their datasets. What are the limits of model scaling?
- Model Improvements
- LLMs have already consumed the sum total of human-produced information. What is the future look like for LLMs?
- Limitations of LLMs and Promising Alternatives
- There are fundamental limitations to what LLMs can accomplish. Simply scaling LLMs up has been unreasonably effective in increasing their capabilities, but over the past 18 months we have really reached a ceiling on that scaling.
- Other subfields of AI like symbolic AI and world models can be used to augment or replace LLMs as more generalized models with internal models of the state of the world.
- Hardware Scaling
- While hardware architecture plays a part, in the past, most performance between GPU generations has mostly come from process node shrinks. We are nearing the end of scaling from process node shrinks alone. What does the future hold for hardware improvements?
- Moats
- While there remain many open questions about how effective LLMs are and the ROI of their integration into businesses, a lot of the value of leading AI companies like OpenAI and Anthropic comes from their status of offering frontier models. Open weight models like Qwen, Deepseek, Kimi, MiniMax, and GLM have shown how precarious that status is, with frontier models only 6 months ahead of their open weight analogs.
- How strong are the moats of the frontier AI companies and will that actually generate the profits that they require to sustain their valuations?
- Security
- Data sovereignty is important to enterprise customers. Where do the frontier AI companies fit in when more companies move to increase security around their data?
- Financial History
- Railroads, telecom, Japanese stocks, the Nifty Fifty, real estate, crypto, AI. We have seen market boom and bust cycles across many industries with various levels of fallout. The hype around AI is simply the latest in a long line of financial bubbles related to new technologies and chasing trends. This will likely turn into its own series of articles, but I will start by putting the current hype cycle of AI into context. We will also take a look at past boom and bust cycles within the AI industry.
- Private Credit
- In the search for higher yielding debt, banks, pensions funds and insurance companies have invested in private credit.
- There is near-zero transparency in the private credit market, allowing unseen risks slowly build.
- Datacenter buildouts are increasingly tapping private credit markets as venture capital and private equity exhaust their cash reserves.
As you can see, there is a lot to unpack. This series will walk through each of these topics in detail, examining both where AI delivers genuine value and where market expectations have completely outstripped reality. The next post will explore these expectations for AI companies and the growth those expectations demand.