AI has a cost problem

{{ OpenAI has an unusual problem—too much demand. Users love its ChatGPT Pro too much, and that’s according to Sam Altman. The CEO says his artificial-intelligence start-up is losing money on its $200-a-month subscription deals. That’s because users making bigger demands on its AI computing resources are racking up costs that are not covered by the fee.

OpenAI’s most powerful model recently matched the average human on a set of problems designed to measure general intelligence but used more than $1,000 of computing power per task. The problem setters noted they paid a [third world] human roughly $5 per task to complete the same set. }}

https://www.barrons.com/articles/ai-stock-market-what-to-know-today-c1faf2b1

intercst

4 Likes

Mr Softie plans to drop $80B on AI data centers, this year, alone.

Steve

1 Like

Do we know what percentage of their capex goes to Nvidia?

DB2

Lots of people think that intelligence is some kind of magic when in reality it is brute force. The human brain has billions of neurons and trillions of synapses. The largest current data centers are dwarves by comparison. Even 100K Nvidia chip clusters are tiny by comparison.

With such huge capital expenditures only money making applications are economically feasible. Humanoid robots and RoboTaxis are two ventures that can generate revenues to cover the cost of AI. AI looks to bankrupt lots of startups. Follow the money!

For investing purposes, it would be interesting to speculate which applications could afford AI at current cost levels. For example, Tesla’s AI5 will be used as the inference chip in both EVs and Optimus robots. Such volumes help to bring down the unit cost.

Recently Elon Musk broke the 30K AI cluster coherence limit extending it to 100K (I think) but even that is tiny by comparison to human brains. The next great investment leap will be by those who can monetize AI.

The Captain

2 Likes

I have only seen an estimate of US/foreign investment split.

Steve

What I’ve heard is about 40-60% goes to the actual racks, server boards, CPUs/GPUs, hard drives, etc. The other ~half goes to the building construction, cooling, networking equipment, etc.

Mike

That did not age well.

Azure = Deepseek

Three Mile Island might as well stay mothballed.