At one point in the past, the phrase “When GM sneezes, the US catches a cold” was bantered around to signify how much of the economy was dominated by car manufacturers. The phrase was a modification of an older version, “When America sneezes, the world catches a cold.” Setting aside the idiom “catch a cold” for now, the phrase recognized the disproportionate economic influence the United States has on the rest of the world.
At this point in the current AI cycle (there have been others in the past), Generative AI (GenAI) has certainly taken over a significant portion of the technology cycle. The remarkable valuation of AI companies, e.g., Nvidia, hitting $4 trillion) does give one pause and invites one to ask what the real upside of GenAI is. The news cycle is filled with numerous AI successes and failures, and yet, as we approach the third anniversary of ChatGPT’s release (November 30, 2022), much of the “AI promise” still seems unattainable. There have been successes.
In contrast to the now infamous MIT report, which cited a 95% GenAI Failure, another study from the MIT Sloan Management Review. Practical AI Implementation: Success Stories highlights areas where AI is achieving success. From the study:
Tasks common to employees in many roles: Large language models are popular for tasks such as synthesizing and summarizing information and documenting meetings.
Specialized uses for specific roles and tasks: Enterprises with a higher risk tolerance are willing to utilize generative AI for business processes. Popular use cases include coding, supporting customer service, guiding the creative process, and creating content at scale. For example, CarMax utilizes generative AI to summarize customer reviews, which are then posted on research pages for customers to use.
Products and consumer-facing applications: E-commerce companies are introducing chatbots and enhancing personalized shopping experiences. Companies such as Adobe and Canva, both of which make graphic design software, are embedding generative AI tools into their products.
In addition, the recent TCP25 conference highlighted the use of AI for Science. Given some of the ongoing challenges for Gen-AI (e.g., hallucinations, bias and discrimination, intellectual property), there is clearly progress with Gen-AI solutions.
And it costs how much?
The pace of Generative AI has been staggering. The market over the last three years has been pretty much buying all the GPUs Nvidia can make (and anyone else, for that matter). The growth of models, however, seems to be reaching a point of diminishing returns.
(Prathmesh T/Shutterstock)
The recent hyped release of ChatGPT5 has been underwhelming in many cases (it should be noted that some people have seen improvements in certain problem domains). Industry analyst and critic/realist Gary Marcus has been tracking AI research for years. His latest analysis of ChatGPT-5 paints a disappointing picture. According to Marcus, who is an AI supporter,
“GPT-5 may be a moderate quantitative improvement (and it may be cheaper), but it still fails in all the same qualitative ways as its predecessors, on chess, on reasoning, in vision; even sometimes on counting and basic math.. Hallucinations linger. Dozens of shots on goal (Grok, Claude, Gemini), etc, have invariably faced the same problems.”
In addition, there now appears to be a computational wall in front of LLM scaling. In the preprint paper The Wall Confronting Large Language Models, authors P.V. Coveney and S. Succi state:
“We show that the scaling laws which determine the performance of large language models (LLMs) severely limit their ability to improve the uncertainty of their predictions. As a result, raising their reliability to meet the standards of scientific inquiry is intractable by any reasonable measure.”
There have been some suggestions by Marcus and others that the poor performance of GPT-5 may be a cost-saving measure. Indeed, the financial picture of many GenAI companies does bode well for profitability over the next several years. And the whole LLM-AGP discussion (Artificial General Intelligence) seems to be off the table.
In a recent interview in Slate, AI critic Ed Zitron talked about the finances of OpenAI and Anthropic:
“OpenAI, principally funded by Microsoft, runs all of its infrastructure. Microsoft owns all the GPUs required to run ChatGPT. While OpenAI is building more infrastructure, they’re doing so dependent on Venture Capital. OpenAI has approximately 700 million weekly active users, although it fails to define what that term means. And I question that number at its source. They’re expecting to lose anywhere from $8 to $ 2 billion this year and will. I think they’re expected to burn $44 billion between now and 2029. They have no path to profitability. Anthropic, principally funded by Amazon and Google, with Amazon running most of its infrastructure. In fact, I think that between Google and Amazon, they own around 30% of this company, which is remarkable. In both cases, these companies lose billions of dollars. Anthropic has leaked that they will lose $3 billion this year. I don’t think that’s true. I think it’s going to be more like five to $10 billion.”
Based on these and other public estimates, many AI companies need to figure out a way to break even sooner rather than later.
Recently, as reported in The Wall Street Journal, OpenAI announced a deal with Oracle to provide $300 billion in cloud infrastructure starting in 2027. It is unclear where OpenAI will obtain these funds or how Oracle will secure the necessary GPUs and power for such an agreement.
Can you make a buck on inference?
At this point, AI companies need to ask, “Can we make money selling inference?” An additional analysis by Ed Zitron, based on data from The Information, indicates that in 2024, OpenAI’s revenue was likely in the region of $4 billion, and the operational loss was $5 billion after accounting for revenue. Thus, the 2024 burn rate was approximately $9 billion. Furthermore, according to Zitron, the information also reports that OpenAI had 15.5 million paying subscribers in 2024, although it’s unclear what level of OpenAI’s premium products they use. That equates to spending $580 per customer with a loss of $258 per customer.
OpenAI will need to either charge more, reduce costs, or find other large markets to become profitable by selling ChatGPT inference. Basically, there has not been a “killer app” for AI.
The computational cost of inference varies. There is speculation that GPT5 was designed to be more power-efficient than previous versions, thereby helping to decrease costs. Even a simple query lights up quite a few GPUs in the data center, however. New and more powerful GPUs will help decrease costs, but the amount of inference revenue may not be enough to close the cost gap. There is also research into smaller domain-specific models that can perform just as well as the large Foundational models, but use much less power. Getting a handle on inference costs will be critical if AI is to become profitable. At this point, there does not seem to be a clear path to profitability. One would assume that AI investors are acutely aware of this situation.
Of bubbles, black swans, and risk
Nassim Nicholas Taleb has popularized Black Swan theory (or dealing with unknown risk) to explain high-profile, hard-to-predict, and rare events that are beyond the realm of normal expectations in history, science, finance, and technology. By definition, “Black Swan Events” (BSEs) are unpredictable and can be particularly hazardous in certain situations.
In later treatments, Taleb also presents ideas that can help mitigate BSEs by implementing robust structures that can sustain negative impact and even grow stronger. These ideas are not new; diversification in investing is sage advice, and the age-old proverb “don’t put all your eggs in one basket” provides a robust strategy. The difficulty lies in trying to predict things that cannot be predicted (BSEs). Bubbles are not robust and will eventually deflate or collapse due to external circumstances and events.
With these practices in mind, consider the figure below from Eye on the Market by Michael Cembalest, Chairman of Market and Investment Strategy for J.P. Morgan Asset & Wealth Management (scroll down to page 6).
US real GDP growth contribution from tech capex spending, Source: Eye on the Market by Michael Cembales, J.P. Morgan, September 2, 2025 (page 6)
According to the figure and source report, the tech sector’s capital spending has contributed to roughly 35%-45% of the overall US GDP growth over the last three quarters. Corporate AI investment reached $252.3 billion in 2024. A large portion (or almost all) of the tech sector capex has been fueled by AI expenditures. Everything from GPUs to data centers are selling at a premium rate.
There have been concerns that the AI market is a bubble. Like those in the past, the dot-com and real estate bubbles can have deep economic effects across industries. As reported in The Verge, OpenAI CEO Sam Altman has indicated he believes there is an AI bubble:
“When bubbles happen, smart people get overexcited about a kernel of truth,” Altman told a group of reporters last week.
“Are we in a phase where investors as a whole are overexcited about AI? My opinion is yes. Is AI the most important thing to happen in a very long time? My opinion is also yes.”
Based on the figure above, even a partial decline in AI capex spending will have a significant impact on the US economy. In essence, there are seven eggs in the current US economic AI basket: Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, and Tesla. Given the state of the adolescent and expensive GenAI market, the current situation presents a high level of risk when viewed in the context of the US economy. If the AI market sneezes and drops the basket, we may all catch a cold.
This article first appeared on our sister publication, HPCwire.