AI runs on borrowed money

AI is efficient, but almost nobody pays what it really costs. The difference is covered by investors and borrowed money, and how long that can last is anyone’s guess.

You pay for the image generator inside Photoshop. Several times a day it tells you there have been too many requests, the servers are overloaded, come back in a few hours.¹ In a few hours it is seven in the evening, and the working day is over. You paid for the tool to be there now.

The message could be down to a glitch, an outage or an attack. It could also be a sign that AI costs more to run than you pay for it.

Subsidised plans, expensive machines

You pay 40, 50, 100, maybe 500 dollars a month. For that you get work that used to cost many times more. The price covers only part of what it costs to build and run AI, which is why the companies building it are losing money. OpenAI’s operating loss for 2025 was $20.9 billion, on revenue of $13.1 billion.² Their data centres are being built on credit, and JPMorgan expects the debt behind it all to reach $4.1 trillion.³ They keep building, because more people use AI and more work moves onto it every month. The bottleneck for all that building is electricity. Data centres used 17% more of it last year, and shortages of grid connections, turbines and transformers now hold back new ones.⁴

At full cost, a plan could run to $5,000 a month instead of $500. At that price a company would rather hire a programmer.

The market splits

There is a second reading of the same numbers.

People who create huge value with AI will pay, because sometimes it replaces a whole company. Someone who needs one HTML footer a month will use a free model.

The cheaper models keep getting better

From inside the Claude and ChatGPT bubble, the expensive top end, that is easy to miss.

Chinese models cost several times less. DeepSeek charges up to $4 per million tokens of generated text.⁵ Anthropic charges $20 for the same amount from Opus 5.5.⁶ Their ability trails the top end. On the Artificial Analysis index, Opus 5.5 scores 58 and the leading Chinese models about 45.⁷ For many tasks, that could be good enough. Good, lower-cost Polish models have come out recently as well. Part of the saving comes from the language itself. Models are billed by the token, a small piece of a word, and a model built for Polish cuts Polish words into fewer pieces. Bielik now needs about half as many tokens per Polish word as its earlier version, so the same text costs less to process.⁸

With cheaper models this good, the move to AI looks irreversible.

Does it add up?

Investors and bankers ask a different question, about the money behind the infrastructure. How many people will pay?

Even 100 million users at $200 a month would bring in $240 billion a year. Next to the trillions going into data centres, that is small change. Corrections are starting.

Microsoft has walked away from a planned data centre in the UK.⁹ OpenAI has cut its projected compute spending to about $600 billion by 2030.¹⁰ Central bankers are comparing the boom with earlier investment bubbles.¹¹

If it is a bubble, it will burst the way the dot-com bubble did, and then the market will settle and real, organic growth will begin.

Anyone who lived through that one has seen this film before. The internet came through its crash. AI could come through its own.

Content gets a price

The rules on what AI may take from the web, and at what price, are still being written. The EU’s AI Act already sets duties for general-purpose AI models.¹² The UK competition authority has ordered Google to let publishers keep their content out of its AI features in search.¹³

The gates will close as the market settles, and some already have. For new customers, Cloudflare now blocks AI training and agent use by default on pages with ads.¹⁴

Google pays news publishers for using their content in its AI.¹⁵ Each deal like Google’s is a new cost for the AI provider, and sooner or later it shows up in the price you pay. File sharing went the same way. Free downloads gave way to Spotify, where rights holders get paid and listeners pay a subscription.

So what do you do while AI is priced below its cost? One answer is to make the most of it while it lasts. Push your productivity as far as AI tools allow, and build tools that automate as much as they can. The other answer is to remember that today’s price is a promotional one.