AI Capacity Hype – or Predictable Demand?
I had a great discussion with my friend 5.6 to find out whether the energy and capacity hype will collapse and just create another financial bubble burst.
The billion-dollar race for data centers. Oracle’s Jupiter, Microsoft. Google, AWS… invest trillions of dollars combined. This smells like a huge risk. We wanted to know.
AI Capacity Hype – or Predictable Demand? The data look like crazy hype, the circular businesses too. But the math taught as something else.
More than enough power and compute for Desktop AI for 2,5 billion people.
In a weekend session with GPT 5.x, we started calculating today’s consumption and a continuous growth rate through 2030. We realized that there is enough capacity to serve four of the eight billion population. And the power may also be enough, as we have three major “Time Zones,” so one is always less active, etc. So far so good.
Then we looked at Enterprise AI consumption under full load.
Let me explain the difference: DESKTOP AI is all the chatbots, copilots, and individual agents that float around worldwide.
ENTERPRISE AI is end-to-end AI applications that can run, even under full load, 24 x 7 x 365. They are not isolated functions that run at night while others work during the day. They are an integral part of a Workflow Intelligence Network. Sales applications split the load so that lead generation, research, and company assessments run at night, market analysis, changes, and shifts run during the day, and all interactions with customers run in their respective time zones. This happens with humans in the loop and also autonomous AI Agent units.
Now, if half of the 4 billion internet and AI users work in offices and those offices get fully AI-powered, we would need to provide capacity for 2 billion people. But we were shocked by our own calculation: The combined capacity of all providers, after they have built their cloud centers, can support only 169 million employees. That’s less than 10% of the required availability.
What did we take into consideration:
1) TODAY’S AI UTILIZATION !!!
Today we use chatbots, copilots, and agents maybe 10 times a day and get a response 10 – 20 seconds later. That is an AI utilization per person of 100-200 seconds of the available 86,000 seconds each day. In other words, the AI utilization is negligible. It’s almost nothing. NOTE: This is why productivity is as low as the utilization.
2) AI PRODUCTIVITY GAIN
The AI Productivity gain grows proportionally with its AI utilization. An important discovery is that the productivity gains we pay for with AI are a fraction of their monetary value.
3) AI CONSUMPTION GROWTH
As AI utilization grows in industrial / enterprise settings, AI compute consumption also grows significantly. In the end, this is no news. This was already recognized in the early days of the industrial revolution, when machine utilization eventually became a critical factor in a company’s economic health. When you get to over 50% AI utilization, you will consume 43,000 seconds per day of AI compute. That is close to 300 times the consumption.
4) REALITY CHECK
Even under best conditions, there is no way that a demand for 2 billion employees with fully loaded AI capacity will arise by 2030.
How much or how little will become incremental demand is currently hard to tell. But it has become obvious that the race for capacity on a global scale is no hype – sooner or later it will become natural market demand.
5) INTERESTING SIDE EFFECT
At 50% AI Utilization, an AI system produces approximately 1,000 – 3,000 results. 80% can be autonomously checked by other AI units. But at least 20% may need to be reviewed by humans in the loop. That’s ~400 decisions, suggestions, proposals, concepts, plans, legal offers… that may require 5-10 additional employees. After all scenarios we calculated, companies will go for higher AI Utilization, keep their entire workforce, and may hire more people
