AI is more like a shoe factory than a software firm

The rise of AI has awakened a demand from everyday users worldwide who can personally attest to its usefulness. The supply of AI is a different matter. It is perhaps less viscerally understood by most of us, and its dynamics are easily misunderstood. AI relies on data centers using computer chips, memory, and electricity. The output is tokens – the building blocks of an AI answer – billions of them a day across chatbots, assistants and generative AI tools.

All well and good – but how should we, as investors, think about that output? In my view, AI more nearly resembles physical products coming off a production line – shoes, say – than the infinitely scalable output of an algorithm.

To grow, a shoe company needs to buy raw materials and invest in new factories. This puts pressure on current cash flow in exchange for higher earnings in the future. AI companies are no different. Every year, chip companies and leading AI developers release new technology that produces cheaper tokens, or “shoes.” These falling costs often cause investors to worry that too many factories have been built or that greater efficiency will reduce the need for additional capacity.

More is more

However, Jevons’ paradox shows that when technology makes a resource cheaper, people often use more of it – not less. In the early industrial revolution, when James Watt’s steam engine unlocked the potential of coal power, economist William Stanley Jevons saw that the result was a massive increase in coal consumption. Why? Because before Watt, coal had fewer use cases with less efficacy than afterwards.

Now we’re already seeing something similar with AI. As the cost of producing artificial intelligence falls, new uses become affordable, more people adopt AI, and existing users employ it more frequently.

We’re still in the early stages of AI adoption – equivalent, in shoe terms, to the introduction of sandals. Boots, loafers, and sneakers are still off in the future. But eventually, nearly everyone will use AI, and most people and businesses will use several forms of it. Token demand is benefiting from three trends happening at the same time: more users, more usage per person, and AI agents that can operate independently of people.

I can see your HALO

Unlike traditional software, which can be written once, copied, and distributed to millions of people at very little additional cost, every AI question requires another “production run.” Each answer uses chips, memory, electricity, and other data-center equipment. In this sense, the AI application may look like a traditional asset-light software product, but every response still requires real-world computational resources to produce. A token is not just another copy of a shoe design; it’s another shoe coming off the line.

Historically, chip demand was tied closely to the number of people using technology. There were natural limits to growth, and this contributed to repeated boom-and-bust cycles. Demand for personal computers and smartphones has depended on how many people needed these devices and how often they replaced them.

Not quite unlimited demand

AI agents could change this relationship. Agents can operate around the clock, and many can work at the same time. Demand is therefore no longer limited by the number of people or the hours in a day. It’s limited by whether the work produced is worth more than the cost of producing it.

AI won’t eliminate the normal ups and downs of the chip industry. Companies can still build too much capacity, and periods of weaker demand will occur. But agents could make long-term demand less dependent on computers, smartphones, and human usage. The world isn’t building the AI factory once and then operating a low-cost software business on top of it. As intelligence becomes cheaper, we will find more reasons to consume it—and the shoe factory will need to keep expanding.