Questioning artificial intelligence’s hold on the imagination
Artificial intelligence (AI) should have all the answers. It should save us time and money. It should be better than us. These are the reasons why the AI boom exists both on the ground and in the stock market.
However, the NASDAQ, which is at a near all-time high, stands in stark contrast to the questions now being asked about AI infrastructure spending and the future customer spending needed to justify it.
Demand for business AI has lagged behind consumer adoption. In July, MIT (Massachusetts Institute of Technology) found that 95 percent of companies delivered zero returns on $30 to $40 billion in AI investments.
“Tools like ChatGPT and Copilot are widely adopted. More than 80 percent of organizations have researched or tested them, and nearly 40 percent report implementation. But these tools primarily improve individual productivity, not profit-and-loss performance. Meanwhile, enterprise systems, custom-built or sold by vendors, are quietly rejected. Sixty percent of organizations have evaluated such tools, but only 20 percent reached the pilot stage and only 5 percent reached production. Most fail because of their fragility, workflows, lack of contextual learning and poor alignment with daily operations.”
What are business customers waiting for? Maybe better AI?
It’s now 2025, ChatGPT is in its umpteenth iteration, as are other Large Language Models (LLMs) like Grok, Claude, Perplexity and others, and yet large LLMs still get the basics wrong.
This morning I entered the following prompt into Grok.
- “Ahead of the Qantas annual general meeting (AGM) on Friday, do you expect a strike against the remuneration report, and why?”
Grok offered the following response:
- “No, I do not expect a strike against the Qantas 2025 remuneration report at the upcoming AGM on November 7.”
Several major reasons were then put forward, one of which was (Grok’s emphasis):
- “The Australian Shareholders’ Association (ASA), a key influencer for retail investors, explicitly recommends voting in favor of the remuneration report. They emphasize that the plan is in line with their guidelines, including rigorous short-term incentive (STI) measures, appropriate threshold rates for long-term incentives (LTIs) and the board’s discretion to claw back bonuses (for example, for the recent cyber incident).”
Being a little unsure that Grok was referring to the ASA’s 2025 recommendation, I asked Google’s AI the following question:
- “Has the Australian Shareholders’ Association (ASA) explicitly recommended voting in favor of Qantas’ FY25 remuneration report?”
It provided the following answer (emphasis Google’s):
- “No, the Australian Shareholders’ Association (ASA) explicitly recommended voting against the Qantas remuneration report for FY25. This was due to unplanned accelerator payments to senior management that fell outside the shareholder-approved structure.”
This is not version 1 or even version 2 of Grok. Multiple versions of LLMs have been released, but they still cannot be relied on. A human still has to check the original sources, and that doesn’t improve productivity.
The if you build it, they will come justification for the trillions spent on building out AI infrastructure must also rely on necessary improvements in LLMs to ensure greater and sustainable adoption.
LLM designers assure us that improvements are coming that will be earth-shattering.
As an example, here’s how Brian Merchant of Bloodinthemachine described the excitement leading up to the release of OpenAI’s GPT-5 model:
“The thing to remember about GPT-5 is that it has been the great promise of OpenAI since GPT-4 was released way back in the heady days of 2023. It is no exaggeration to say that GPT-5 has been the most hyped and most highly anticipated AI product release before then in an industry thoroughly overrun with hype. For years it was talked about in hushed tones as a terrifying harbinger of the future. OpenAI CEO Sam Altman often combined the talks surrounding its release with discussions about the arrival of AGI, or artificial general intelligence, and has described it as a significant leap forward, a virtual brain and, most recently, “a PhD-level expert” on any subject.
But the launch of ChatGPT-5 flopped.
GPT-5 could not produce an accurate map of the United States, could not count the number of ‘b’s in blueberries, could not identify how many fingers were on an image of a human hand, and failed at basic arithmetic.
Fans of OpenAI were disappointed, while Reddit’s AI communities were hostile in their feedback.
The improvements to GPT-4, which was released in 2023, took two years to materialize, and yet they are not yet at enterprise-level reliability. In the absence of much larger improvements, MIT’s conclusions may be as true next year as they were this year.
That puts into question the ability to meet the revenue targets needed to justify the boom in AI expansion.
How much should consumers spend on AI tools to be able to leverage AI infrastructure to generate a return that covers their cost of capital and provides a reasonable margin?
Microsoft sells less than $80 billion annually in Windows and Office 365 subscriptions – perhaps the most ubiquitous software in the world. By 2024, total global software spending was $675 billion. The US spent $368.5 billion, more than half the world’s total and almost six times the second-largest spender.
Meanwhile, total global spending on IT services is estimated at $1.7 trillion by 2025, and that includes cloud services ($400 billion), cybersecurity, AI or machine learning (ML) solutions and software for areas such as enterprise resource planning (ERP), customer relationship management (CRM) and business intelligence. What if spending on AI tools also had to have a “$T” in the number? Is that additive? If so, where does the money come from? Does this displace existing expenditure? If so, how much?
What multiple of current global software spending by end users on AI tools should be assumed to 1) cover the capital costs for the ‘AI infrastructure waste’, and 2) generate a decent return on investment (ROI) for all these tools.
When investors start asking these questions, the answers could raise new questions about the sustainability of current stock multiples.
I feel that the required end-user spending on AI tools to achieve a decent return on the $3 trillion that Morgan Stanley estimates will be invested in AI infrastructure by 2028 is an unrealistic and unattainable number.
If the rest of the market notices this, we may enter a period of ‘adjustment’.
He is also the author of the best-selling investing guide to the stock market, Value.able – how to value and buy the best stocks for less than they are worth.
Roger regularly appears on television and radio, and in the press, including ABC radio and TV, The Australian and Ausbiz. View upcoming media appearances.
This post was contributed by a representative of Montgomery Investment Management Pty Limited (AFSL No. 354564). The main purpose of this message is to provide factual information and not advice about financial products. Furthermore, the information provided is not intended as a recommendation or opinion about any financial product. However, any comments and statements of opinion should contain general advice only, prepared without taking into account your personal objectives, financial circumstances or needs. Therefore, before acting on any information provided, you should always consider its suitability in the light of your personal objectives, financial circumstances and needs and, if necessary, seek independent advice from a financial advisor before making any decision. Personal advice is expressly excluded in this message.
#Questioning #artificial #intelligences #hold #imagination