The (future) cost of AI

ChatGPT and other LLMs are seeing wider use in society. We also see Stable Diffusion and other generative AI being used for creating illustrations. While we are able to run smaller models on our own computers, as long as we have the hardware (usually a modest GPU), the bigger and more powerful models are usually provided as cloud services, hosted on servers in data centers. There may be free plans, but access to better models usually require some form of subscription, or are charged based on usage.

Right now, the price of using such AI services may still be low, with companies keeping prices low to attract a wider user base. However, such a pricing model is not sustainable. Think about the hardware needed to configure such servers. Think about the money needed for building and running data centers, including construction, staffing, and electricity. Data center GPUs like the H100 can easily cost the price of a car. The cost of setting up data centers for AI services and operating them can easily run up to billions of dollars. Do you think the low prices now will be able to actually cover the money invested? Of course not.

This is where the sinister part of capitalism comes in. Companies providing AI services are not trying to cover costs now in the short term. The short-term goal is to expand user base and develop user reliance, and eventually drive out the human alternative.

Wait, what is driving out the human alternative? Well, it is simply the forcing out of the competition from humans providing that same service, whether it is artists who draw illustrations, translators, writers, musicians, singers, narrators, maybe even video creators. Low pricing of AI services providing such services now makes it hard for humans providing such services to compete, eventually forcing humans to seek other jobs. In a few years, we will see entire sectors of jobs disappear, replaced by cheaper AI services. In another few years, such skills would be lost due to lack of use and practice.

And that will be when companies collect on the actual cost of AI. When there is no human alternative left, companies can easily hike up prices, justifying price hikes by showing users how much it actually costs to run their data center. And users will have to pay because they have no where else to turn to. Such a day is not an "if"; it is a "when". When that day comes, it will be too late; our reliance on AI will be so entrenched in our lives, we have no choice but to pay. And pay we will, because AI is not cheap. All these companies didn't spend billions on training AI models and building data centers to earn 10 dollars from you each month. We will end up paying for years of R&D and infrastructure investments, plus the operating costs of that infrastructure.

If anything, just think about self-driving cars. Today's self-driving cars have not reached Level 5 autonomous driving. Such cars will probably be driven by cloud services; the core model driving the vehicle will be located in the car, but required data (such as maps and traffic information) will be made available to the AI model via the cloud. With time, all cars will be replaced by self-driving cars; to prevent accidents, countries may even require everyone to switch to self-driving cars. A few years later, no one will actually be able to drive a car. Other forms of mobility (like trains) will probably end up being replaced too. Once getting around becomes a cloud service (except walking), companies can then charge you anything they want.

Convenience and speed comes at a price. We may not be paying that price today, but companies are not charities, and the day will come when they come to our doors, ready to collect. And we will have to pay, with no other alternative.