Say, you were starting a business, a restaurant.
You would have a category expenditure described as Capital Expenditure. You would have to buy cooking ranges, mixers, ovens, utensils, cutlery and the like. This is a starting expense, and you expect longevity. You do not expect to buy another cooking range for a decade, if not more.
Then you have operating expenses. Under this, you would have rent, salaries, ingredients, marketing and the like. In this, rent and salary will go out each month. The property owner does not care if 1 customer came or 10,000; the entire rent gets paid. These are called fixed costs.
By comparison, ingredients are truly a variable cost; you don’t use them up unless you make a dish. Hopefully, you do not make a dish unless a customer shows up.
Now, let us say your fixed cost is 50000. And your variable cost per customer is 50. You are charging the customer 100. You are making 50 over, and above your variable cost on every sale, hence you need to sell 1000 units to arrive at what is known as operational breakeven. At 1000 customers, you can pay off your fixed cost of 50000.
The moment you exceed 1000, you start recovering Capital Expenditure. At some point, you reach full breakeven.
Now, let us consider AI businesses, especially those claiming to build foundational models and raising billions as if they were millions.
Their Capital Expenditure is the data centres that they require. Data Centre capacity is measured in terms of Megawatts. The frontline companies, such as OpenAI and Anthropic, are building 100 MW data centres. This requires 50+ acres of land to begin with, and then they require an array of Nvidia GPUs. Each GPU costs about $5000. A data centre could have close to 100,000 to 200,000 of these. Then you have the building and all the other paraphernalia needed. Typically, a data centre costs about $3-5 billion to set up.
A 100 MW data centre is capable of catering to 10 - 50 million users based on the load it is receiving. For instance, if I send Claude a query asking it to convert a report into an interactive website, this can take 20 minutes of operation on several GPUs. Hence, the ability to cater to a large number of users can come under stress if the requests require a high degree of sophistication.
Then come the operating expenses. Depending on the cost of power, operating a 100 MW data centre can cost between $50 million and $150 million per year. Most of the foundational AI companies require Gigawatt scale to cater to the demand that they are receiving. Hence, the 100s of billions being spent.
But here is the fun part. The life of a GPU is between 5 and 10 years, depending on the amount of use it is subjected to.
Have you ever bought a laptop and, after a decade of use, it starts to sprout all kinds of issues? It starts to freeze up, crash, and be generally very slow. That is the result of, for the lack of a better expression, “wear and tear”.
The AI companies are throttling use because of the demand that they are seeing. Safe to say, very high wear and tear. So the 100s of billions being spent will have to be spent all over again in 5 years!!
Imagine you are building a hotel and you have to demolish and build all over again after 5 years! No hotel would ever break even.
Have you heard of an AI company that has managed to break even?
In no world is a $4 billion monthly revenue run rate small. If, at that scale, you are unable to breakeven, there is something fundamentally wrong financially. OpenAI boasts 50 million paying subscribers, and Anthropic claims 30 million paying subscribers.
At what scale should operating income overtake operating expenses? This is not an ad subsidised business such as Google or Facebook.
Also, it is one thing to get revenue, and it’s another how you get it. Across companies there are leaderboards set up for engineers who use the highest numbers of tokens used by the engineers on AI. Bad incentives lead to bad results.
Axios reports that an unnamed corporation managed to “accidentally” torch roughly $500 million in a single month. The cause? The company introduced Anthropic’s Claude to its workforce but forgot to put caps on usage. Employees ran long, unchecked queries and complex agentic workflows, racking up an enormous tab before the finance department even realized the meter was running.
Source: Inc
This is lunatic.
If you paid $25,000 a month to an engineer, you could have hired 20,000 engineers for a month with $500 million dollars. That would be the number of software engineers that a company like Apple or Facebook has on their rolls. I hope they get some output from all that AI use, although I highly doubt it.
This kind of wanton spending is propping up revenues. This is sure to disappear sooner or later.
Ultimately, the problem is that all of these companies are starting with the stated goal of putting a lot of humans out of work. They hope to achieve this using regurgitation machines.
There is a simple problem they are all struggling with. The human brain runs on 20W of energy; to replicate that capability, these companies require several megawatts of energy. That asymmetry is proving hard to overcome and manifesting itself in the form of economics.