AI’s fatal flaw

AI can improve efficiency, boost productivity, uncover hidden connections and perhaps even help cure disease. At a more mundane level, it will remind you to slow down while driving and tell you when it’s time to replace the water filter in your refrigerator.

[What follows] is not an argument against AI itself. It is an argument against paying irrational prices for AI stocks. AI’s usefulness does not justify paying bubble valuations or chasing momentum in a speculative frenzy.

In other words, AI will survive. Many of today’s stock prices will not.

At Strategic Intelligence Australia, our goal is not only to help you build wealth, but also to protect the wealth you’ve already earned. When it comes to preserving wealth, it’s better to miss a few points on the upside than watch your portfolio melt away on the downside.

I believe now is the time to reduce your exposure to the AI sector. The analysis that follows explains why.

Perhaps the greatest flaw in the AI business model is that it is not scalable.

Scalability means the cost of acquiring each new customer falls as the customer base grows. Fixed costs are spread across more users, while the cost of serving each additional customer steadily declines.

Every business begins with substantial upfront expenses: legal and accounting fees, product development, hardware, management, marketing and administration. Once those costs are absorbed, growth should become increasingly profitable.

When only a handful of customers share those costs, the expense per customer is high. As millions of customers arrive, however, the cost per customer falls dramatically. At the extreme, the cost of serving each additional customer approaches zero, while the revenue from that customer flows almost entirely to the bottom line.

That is why scalable businesses command such rich valuations. Venture capitalists and Wall Street place enormous value on recurring revenue with declining marginal costs.

But OpenAI, Meta, DeepSeek, Mistral AI and other frontier-model developers are all facing the same problem. As demand increases, they must spend exponentially more on computing power, chips, data centres, electricity and infrastructure simply to keep pace.

The economics remind me of an old Wall Street joke about a company that loses money on every sale.

The CEO says, ‘We’re going broke.’

His banker replies, ‘Make it up in volume.’

That, in a nutshell, is the logic behind much of today’s AI spending. If every additional customer increases losses instead of profits, scaling the business doesn’t solve the problem. It simply makes the losses larger.

More compute, less profit

The culprit is what the industry calls compute — the total cost of processors, hardware, data centres, electricity, software development and every other resource required to produce AI output.

Training today’s frontier models requires enormous datasets — essentially the public internet, plus as much personal information as companies can gather. Running those models demands trillions of calculations across increasingly complex neural networks.

AI can certainly improve efficiency. The question is whether customers are willing to pay enough to cover the industry’s exploding costs. So far, the answer seems to be no.

There are practical alternatives. One is to build much smaller, highly curated models designed for specific tasks instead of trying to answer every question imaginable. Another is to abandon the race toward mythical superintelligence and focus instead on systems that are simply ‘good enough’ to solve real-world problems.

The AI industry is unlikely to embrace either approach. Too much money has been invested in the promise that AI represents the greatest technological breakthrough since the steam engine — IPO valuations depend on that narrative.

But investors are beginning to recognise the difference between technological promise and economic reality. Once investors understand that today’s frontier AI models become more expensive — not less — as they scale, lofty valuations and trillion-dollar IPO dreams could evaporate remarkably fast.