It's common to hear tech executives extol the boundless possibilities of AI. However, in a shocking turn of events, one of AI's leading proponents is raising serious concerns. The CEO of OpenAI, Sam Altman, is cautioning that the industry's rapid growth may be about to come to a painful halt due to killing costs rather than killer machines.
While companies are locked in a fierce race to build the next groundbreaking AI, Altman suggests the real enemy might be a simple spreadsheet. The current path of creating ever-larger and more powerful models is, in his view, financially unsustainable.
Where the Money Really Goes
We often hear about the eye-watering, multi-million dollar price tag to train a system like GPT-4. But Altman points to a even bigger money pit: the cost of actually using it.
Every time you ask an AI a question, it consumes computational power. This is known as an "inference cost." For a simple query with today's models, it's cheap—just a few cents per chat.
The problem arises when we look to the future. Imagine asking a super-powered AI to solve a massively complex problem, like developing a new drug or planning a city's logistics. It would need to "think" for a long time, using enormous resources.
You could instruct the AI to "use the whole stack, spend $100,000 of compute on this one prompt," as Altman stated. All of a sudden, the price is a fortune rather than pennies. This leads to a basic commercial issue: how can you develop a product for the general public when the cost of a single sophisticated task could be thousands of times more than the price you could possibly charge for it?
The Race for a Fix and the Imminent "Implosion"
Altman refers to this discrepancy between price and value as a possible "implosion." The entire business case for many AI applications may crumble under its own weight if the economics don't become better.
So, what's the way out? The industry is betting everything on a technological moonshot: a revolutionary leap in "compute." Compute is just a fancy word for raw processing power, the kind provided by advanced chips (like the GPUs made by NVIDIA that are currently in short supply).
According to Altman, a double win—a significant expansion in the amount of compute that is available and a sharp decline in its cost—is necessary for the future of affordable AI. His bold intentions to raise trillions of dollars to boost global chip production capacity are motivated by this.
Altman's caution serves as an important reminder of reality. It serves as a reminder that the path to advanced AI is a significant economic problem as well as a scientific conundrum. The next major development in plumbing is necessary to make the technology not just smarter but also affordable enough to be used if the AI revolution is to be genuinely successful.