The AI industry's underlying business model contains a structural flaw that worsens with every technological advance. OpenAI, Anthropic, and other AI companies charge users for "tokens" - the units of text processed by their models. But unlike traditional commodities such as oil or steel, the price of each token is rapidly falling as AI capabilities improve, "Hvylya" reports, citing The Atlantic.
Paul Kedrosky, an investor and financial consultant, called this dynamic "a death spiral to zero." As the value of a token drops, so does the value of what data centers can produce. Meanwhile, the advanced chips that make up the majority of a data center's cost are constantly superseded by the next generation, meaning the hardware depreciates almost as fast as it is installed. The supposed backstop for all the data center debt - selling the facility itself - is not actually a backstop.
The Iran war has added another layer of pressure. Energy prices are expected to stay elevated for years, because the targeted fossil fuel facilities in the Persian Gulf will take a long time to restore. Should energy costs continue to climb, so will the cost of manufacturing and operating computing equipment that already requires enormous amounts of power.
The revenues of Anthropic and OpenAI have been multiplying every year, which proponents argue means generative AI products are on track to become profitable. But on the current trajectory that would still take years, and there are reasons to think growth will slow. The main draw of AI tools is "efficiency" - executives hope AI will allow them to cut costs rather than grow output. The medium-term success of generative AI would likely put millions of people out of work, suppressing the very demand the industry counts on.
The hyperscalers - Microsoft, Google, Meta, Amazon - have spent far more than investors expected, yet revenue has not kept pace. Every time they announce additional data center investments, stock prices now fall rather than rise. The gap between expenditure and returns has left what The Atlantic described as a range of outcomes "somewhere from mildly bad to historically so."