I was listening to a few of Nikhil Kamath’s podcasts over the past few days, and a few things struck me. AI is supposed to result in greater productivity. That productivity should show up in the form of economic growth. Whether any of that accrues to the working class or not is a question for another time. If the last 30 years have shown us something, it will not.
But even so, the pertinent question is: Between OpenAI and Anthropic, $100 billion in revenue is being generated; then where is the productivity? Where is it showing up in the economy? Over the last year, any economic growth can be simply attributed to price rises. So what is happening?
The current AI narrative is that AI will take over the jobs of many people and reduce costs for corporations, and this will reflect as stronger bottom lines.
In other words, the purpose of AI is to allow wealth to access skill while denying the ability of the skilled to access wealth.
It is not framed as a tool that is meant to increase productivity. It is framed as a tool that will reduce cost. The jury is out on whether that would indeed happen or not.
But since it is mainly meant to replace people, there will not be any rise in GDP through the rise in productivity. It is only meant to move the source of productivity from people to machines. It is supposed to move expenditure going out as salary to an invoice to OpenAI or Anthropic. Therefore, if at all any value is being created by AI model companies, it is not going to deliver an increase in GDP.
With the current narrative, AI will not deliver any GDP growth.
On the other hand, Harshil Mathur from Razorpay argued that the cost of a token is approaching zero. His evidence was the $5000 machine at home that runs an open-weight model, and the cost of electricity for running this for an entire month is Rs 1000 ($10). If the cost of AI is negligible, then why are frontier AI companies haemorrhaging money? Most of the data centres are being built on other people’s dime. It is large tech companies like Meta, Google, Oracle, and Amazon that are taking on debt on their books to build these data centres. The frontier model companies are merely renting the capacity.
If the token pricing is that low, they should be making money hand over fist because they have managed to push the capital expense into other people’s balance sheets. They essentially have only variable costs. Outside of servicing customer queries, the only major cost is training newer models. Training models is a pursuit that has no finish line.
When are they going to stop training?
Just like Apple did not produce iPhone X and then announce “this is it”. In the last 10 years, there has been zero change in form factor or broad functionality. But they need to keep producing newer models.
Similarly, most startups using AI agree that frontier models are unnecessary to do the job. Open-weight models do the job just well enough. That has not stopped any of the AI labs from training newer and newer models every 6 months.
AI is already as cheap as it could be, but the vain pursuit of an unknown goal keeps burning money.
75% of the income that frontier AI companies such as OpenAI and Anthropic make comes from selling tokens to other AI startups that have not even started making money. So startups that are hoping to find a use case and efficiency are the ones mostly contributing to the revenues of these companies.
VCs are keeping the AI party going. When the VC AI craze dries up, so will the revenues.
Most companies that have discovered a well-defined use case for AI are using open-weight models since they are much cheaper to run.