What you need to understand about the AI Economy

I grew up in a household which embraced tech early. My dad was a software engineer during the internet’s rise. I think there has always been a computer in my household, and I learnt how to use prompts for MS-DOS (the system before Windows) as a kid so I could play Tetris. We had the Internet at the beginning, well before Google became a household name. Does Altavista ring a bell for anyone?

It was an exciting time. The message I had growing up, was that internet was going to change the world. In a good way. It was an optimistic message.

AI promises to change the world too, but the current AI boom definitely feels different.

It might be because I’m not in the same tech-embracing bubble as a small child, or it might be because I’m now an economist and a bit more cynical about life.

The biggest difference? Probably the anxiety of all the jobs it can replace. And how quickly experts think this could happen.

No doubt, it is astounding what AI tools can do now. For this piece, I can become a cartoonist, a tech expert and an economic historian – although at an amateur level. As an economist, I can read through research papers and analyse information quicker than I ever have before, although I can’t get AI to write anything close to what I want it to write. It also seems to fail when finding dates of economic data releases. It’s a few mistakes here and there, but enough to decide I can’t trust it, and I should just look up dates on my own.

AI is amazing, but it still needs human oversight.

And the more I’ve looked into AI more deeply, how it works and how it is interacting with the economy – there are some big challenges to resolve before we see true widespread adoption, and therefore the large displacement of jobs.

For this piece, I’ve focused on the general-purpose AI models like OpenAI’s ChatGPT, Anthropic’s Claude and Google’s Gemini. There are of course, other forms of AI, but these models are the ones creating the most hype and they are commonly being used as a foundation for more specialized models.

So, what does the industry have to overcome?

Firstly, there is:

1) The infrastructure needs

Analyst estimates point to OpenAI spending plans at close to $1 trillion over the next ten years to build the necessary infrastructure for widespread adoption. And there are big dollars from Amazon, Microsoft, Alphabet (Google) and Meta as well as Chinese companies, Alibaba, Tencent and Baidu.

But all this capital spending, mostly data centres and hardware – usually hasn’t accounted for the energy and water infrastructure necessary to support these data centres. The International Energy Agency (IEA) estimates that electricity demand would more than double by 2030 from the early 2020s if there is widespread AI adoption.

Some US big tech firms have reportedly signed deals to attach nuclear power to their facilities. But this takes time, and these aren’t likely to come online until 2030 at the absolute earliest. Which means any earlier builds of data centres before additional power infrastructure is built are going to place increasing strain on existing energy infrastructure. And this is on top of commitments to move away from fossil fuels for energy generation and electrification of motor vehicles.

This brings us to the next major hurdle for widespread AI adoption:

2) High costs

In the first industrial revolution, a string of innovations using machines were adopted in a variety of industries. Some of these new innovations especially in textiles were profitable very early on, not only because they could increase output per worker, but because these technologies cost less than the old technology, like the spinning jenny compared with weaving cotton by hand.

While AI can make existing work faster, whether it costs less than what we currently do already is still questionable.

Nvidia is getting a lot of attention with its double-digit earnings growth, but they aren’t building the AI, they are selling the chips to build the data infrastructure to power the AI models. Meanwhile, the model builders, such as Open AI, Anthropic and DeepSeek love to speak of the impressive revenue growth that they have, but information we do have suggests that they are not in profit. Google, Amazon, and Microsoft which have also ploughed big money into AI models are able to absorb the cost into their existing service offerings.

Of course, with startups, it is generally accepted for them lose money in the initial phase of their operation. As businesses gain more customers, revenue increases more than its costs.

But what is alarming about AI model builders’ cost structure, is that if they were to gain more customers, there are costs that scale up one-for-one with revenue.

Think of these costs like variable costs, which the tech world call “inference” costs.

AI companies incur this inference cost each time a user makes a query. This includes the computation cost of making the query, cloud services, energy consumption, cost of cooling, etc. AI model builders are mostly private companies, so we don’t get full financials to the public. but leaked information suggest that inference costs are exceeding revenues, and for OpenAI, they exceed revenues by a lot this year.

That means they don’t have a scalable business model - with their current cost structure, they can’t just scale up and be profitable. And there is also the fixed cost of implementing, training and maintenance of the models.

The only way that these general AI models have a long-term future is if these variable costs come down, or they find a way to charge a lot more for their products.

There is a lot of attention in working out how to bring down these inference costs and make models run more efficiently. But they face an uphill battle when it comes to energy costs which can only rise if demand for AI increases.

3) Social acceptance and politics

Finally, it’s hard to see universal acceptance of what AI will bring.

This is a time when voters are unhappy about the rising cost of living in many developed economies, so a growing number of data centres pushing up demand for electricity and driving prices higher, isn’t likely to be a popular development.

Most consumers would not see the benefit of AI beyond ChatGPT assisting with writing emails, unless they hold shares in a hot tech company.

They care more about their jobs. While AI is not at a point where it can fully replace whole occupations, like many technological advancements, the roles that are more likely to disappear are those that are lower-paid, lower-skilled or junior roles.

On top of the cost-of-living concerns, in the US, social inequity is high, and income inequality is at its highest since at least the 60s when official statistics of gini inequality measurements begin.

Donald Trump’s presidency came on the back of a movement that have been lamenting the loss of jobs in manufacturing, and the working class that has been left behind.

If nothing is done to distribute incomes, wealth and opportunity more evenly across America, these grievances are likely to get worse, especially with a future where AI is growing in use at the expense of jobs.

It could lead to a push towards blocking AI advancement.

In economic historian Carl Benedict Frey’s book “The Technology Trap”, Frey provides example after example throughout history of when labour-saving technology has been blocked or opposed. From Emperor Vespasian in Ancient Rome banning the use of a construction device, to the Luddites smashing textile machinery in 19^th century Britain, banning or attempting to block technological advancement at the expense of existing jobs or the status quo has been the norm in our history. Frey also emphasized the importance of the social environment and politics in whether technology advanced or not.

There is also the question of intellectual property. A number of lawsuits in Europe and in the US against OpenAI, Anthropic, Google, etc. have generally leaned in favour of plaintiffs against the AI companies that use copyrighted content without permission for AI training. Two major ongoing cases to watch out for are The New York Times against OpenAI, and a group of US authors and publishers, also suing OpenAI for using pirated book datasets to train their models and there are a number of lawsuits ongoing in different parts of world. Settling these lawsuits and the need for licensing deals are going to add to the costs for these AI companies.

What this all means and the investment implications

This wasn’t meant to be a piece to bash AI, it was about understanding how the economics of AI stacks up. And it does seem like there is a lot for the industry to overcome if there is to be widespread adoption.

Does that mean that there is an AI bubble?

There are bubbly aspects to AI-related assets, and there is a big risk that the over trillion dollars of investment planned would not pay off for many years to come if at all. Investors are putting the horse before the cart and like the idea of AI more rather than seeing that these companies can viably make profit in the future.

That doesn’t mean that AI won’t. It all comes down to whether they can feasibly bring down the costs of running their models to less than what people are willing to pay for. There are industry experts who know more about the tech than I do who seem to think they can, but there is an uphill battle when considering the impact of AI on global energy demand and its impact on energy costs. Licensing content to train models also looks like it will be more costly, not less. And it’s not yet clear that the social and political environment is favourable in supporting AI.

Economic historians say that the industrial revolution didn’t properly take off until coal and steam power came into play. Before this, early factories relied on water and wood for power, which limited the productive capacity of the new inventions that mechanised labour.

It might be the case that we can’t have a proper AI revolution until we have another power revolution. Even if AI model builders can remake their business models to be profitable, AI use could be constrained until we see some proper breakthroughs in generating more, cheap and clean power.

And if the industry can’t bring down its costs enough or shift its business model, it may cease to exist in the way we know it today.

I am reminded of the story of the Concorde – we loved the idea that we could fly at faster than the speed of sound. But its high cost meant that it remained a niche business, people didn’t like how noisy it was, and ultimately public trust eroded after one fatal accident.

It’s a cautionary tale for AI. In the long-run, the idea of AI just isn’t enough.