About That AI Investment…

These companies represent both American and Chinese approaches to AI development. Their strategies differ, particularly in how they balance investment in computing infrastructure against improvements in software, model design, and efficiency.

AI vs AGI

AI, as virtually everyone knows, stands for Artificial Intelligence. AGI, short for Artificial General Intelligence, doesn’t yet have an agreed upon definition, but it generally means AI which has broad, flexible intellectual abilities, able to learn and perform a wide range of tasks at the same cognitive level as humans.

Today’s AI could indeed contribute to enormous productivity growth, without major further progress toward AGI. However, Artificial General Intelligence has been the “holy grail” of AI development in the US for some time. It’s the reason our hyperscalers will need to keep building more and larger data centers until either AGI is achieved or investors realize the hyperscalers may never get their money back. We suspect they’re overbuilding, as we do not believe AGI offers economic benefits that will ever be achieved, or will surpass the value proposition offered by “non frontier” Chinese AI models in which achieving AGI isn’t the focus. More about this below. 

There are two problems caused by focusing on AGI instead of using the AI we’ve already created to enhance productivity. The first is the ongoing, increasing power needs of almost unimaginable proportions. 

The second problem is much more dangerous and difficult, if not impossible, to solve. You can’t have AGI without autonomy. Anyone who follows the news these days is aware of the startlingly dishonest, in fact outright criminal acts autonomous AI agents have already committed. Some voices in the media feel the only way we will ever get control over the issue is if there is a disaster large enough to bring attention to the danger. At that point, with necessary regulations, this means autonomous AGI can never be permitted, at least not without large safety advancements many deem impossible to coexist with AGI.

Yet the hyperscalers are betting trillions of dollars on the idea they will achieve AGI, and that achievement will lead to untold riches. What could go wrong?

The economic opportunity is already substantial. The critical question, again, is whether we need to pursue AGI to realize it.

Why AI Still Has Difficulty Replacing Humans

There is an important distinction between using AI as an assistant and allowing it to operate independently.

An AI assistant can help a person work faster while leaving the human responsible for reviewing important decisions. This arrangement can capture many of AI’s benefits without requiring the system to be right every time.

The difficulty increases when companies attempt to replace human judgment entirely.

Today’s AI systems can produce plausible but incorrect information, a problem commonly known as “hallucination.” Their reliability varies by model, task, and circumstances, and there is no single error rate that applies to every system. Developers have made progress in reducing these errors, but they have not eliminated them.

This limitation is especially important when an AI system is expected to act autonomously in situations where mistakes could have serious consequences.

A system that occasionally produces an incorrect paragraph may be useful with relatively little supervision. A system that occasionally makes an incorrect financial transaction, disrupts critical infrastructure, or takes an inappropriate action in a military setting presents a very different problem.

This is where the economics of AI autonomy become questionable. If a system requires extensive monitoring, independent verification, restricted permissions, and human intervention to operate safely, some of the savings promised by replacing people may prove difficult to achieve.

Here is ChatGPT’s perspective on the issue:

Hallucinations and consequential decisions

“Language models can produce plausible but false statements, commonly called hallucinations. No general solution guarantees factual accuracy across unrestricted real-world questions. Training and evaluation can reward guessing rather than acknowledging uncertainty. 

Reliable documents, calculation tools, restricted permissions, and independent validation can improve performance; some checks can be automated. Nevertheless, better average accuracy does not establish acceptable reliability where occasional errors could cause enormous harm.

Authority over bank accounts, water systems, electrical grids, or military operations raises different questions from drafting reports. What happens when the system is wrong? What independent safeguards prevent harmful action? Verification costs and restricted authority could materially reduce the savings promised from autonomy.”

— ChatGPT

The distinction matters financially. If reliable human oversight remains necessary for many consequential tasks, the commercial market for AI could still become enormous, but the business case for unrestricted autonomy becomes considerably less certain. That is a very different outcome from the one investors may be anticipating when they finance the construction of enormous data centers.

The Chinese Challenge

Chinese AI developers present a serious challenge to the American industry’s investment strategy. Their progress suggests that increasingly useful AI does not necessarily require the same level of computing expenditure being contemplated by some American companies.

Hardware and software are both essential to AI, but they have very different economic characteristics. Hardware requires substantial capital investment, electricity, cooling, and ongoing maintenance. Software improvements, including more efficient models and better algorithms, can sometimes deliver major gains without requiring a proportionate increase in computing capacity.

This raises an important question: Are American companies investing too heavily in infrastructure when improvements in software and efficiency could deliver much of the same practical value at substantially lower cost?

Consider the enormous amount of data-center capacity already in existence in the United States, along with the additional capacity planned for construction. The United States has committed substantial resources to this buildout, reflecting the industry’s expectations for future AI demand.

Why, then, do Chinese AI developers appear able to compete despite operating under different infrastructure and investment conditions?

One possible explanation is that their approaches place greater emphasis on software efficiency, model design, and the practical performance of AI systems relative to their cost. The details differ among developers, and the economics of individual models are not always directly comparable.

Nevertheless, the competitive implications are significant.

If Chinese developers can deliver comparable capabilities with substantially less computing power, American companies may struggle to earn adequate returns on their much larger infrastructure investments.

The Chinese approach may not prevail, nor the cost differences remain constant. Yet the investment implications could be enormous if the market eventually discovers that comparable AI capabilities can be delivered at a fraction of the anticipated cost.

This is not simply a question of which country has the most powerful AI. It is a question of which companies can provide useful AI at a price customers are willing to pay.

There is also a strategic difference in objectives. Much of the investment rhetoric from major American AI companies emphasizes the pursuit of AGI and increasingly capable autonomous systems. However, the immediate economic opportunity may be in making existing AI systems more useful, reliable, and affordable.

If the latter approach produces most of the practical benefits, the companies that have invested most aggressively in infrastructure may not necessarily be the ones that earn the highest returns.

AGI Investment Risk

We do not need to reach AGI before serious AI risks emerge.

One concern is the use of increasingly capable AI by malicious actors. AI can lower the barriers to certain forms of cybercrime, assist with the discovery of software vulnerabilities, and potentially help attackers develop more sophisticated methods.

The precise scale of future harm is uncertain, but the direction of the risk deserves attention. AI could make some forms of malicious activity  faster, cheaper, or more accessible to people who previously lacked the necessary expertise.

This is particularly concerning where cyberattacks could affect financial institutions, critical infrastructure, or other essential services.

A second concern is recursive self-improvement. This refers to a system improving its own capabilities and then using those improvements to develop further improvements. In theory, such a process could accelerate development and make the behavior of advanced systems more difficult to predict.

Claims that a particular AI system has already achieved meaningful recursive self-improvement, or that it is improving exponentially, should be evaluated carefully. Demonstrating that an AI can assist with its own development is not the same as establishing that it can independently and continuously accelerate its capabilities.

Nevertheless, the possibility of rapid advances makes effective monitoring and safeguards increasingly important. If systems become substantially more capable between evaluations, existing safety procedures could prove inadequate.

In September 2026, Jacob Coxon, a researcher who had worked at Anthropic and OpenAI, publicly raised concerns about the pace of AI development and called for leading AI laboratories to coordinate on limiting recursive self-improvement. His comments reflected concerns about the direction of future development; they should not be interpreted as confirmation that an AI system has already achieved unrestricted, autonomous self-improvement.

Other AI industry leaders have also publicly discussed the need for greater caution. Bloomberg reported in September that leading executives were calling for a slower pace of development of advanced AI models and stronger safety measures.

These developments do not establish that an AI catastrophe is imminent, or that recursive self-improvement is already occurring at an exponential rate. They do, however, illustrate why the possibility of rapid capability advances deserves serious attention.

Where Does This Leave Us?

For investors, the stakes are substantial. If the profits supporting today’s prices fail to materialize, losses could be severe even while AI adoption continues to grow. We do not need to know the date of a market peak to recognize that retirement security should not depend on the most optimistic version of this story coming true. Our responsibility is to preserve the ability to participate in AI’s economic benefits while protecting clients from a speculative boom that could destroy years of accumulated wealth.

Important Information: This report expresses the author’s views and opinions about AI development, the economics of AI infrastructure, and the associated risks. Statements concerning future technological developments, productivity, corporate profitability, and investment returns involve substantial uncertainty. Actual outcomes may differ materially from the scenarios discussed. References to potential investment losses are not predictions that such losses will necessarily occur. This report is provided for general informational purposes and should not be interpreted as a guarantee of future economic or market outcomes.