On AI Economy

I am reminded of an example in Neil Degrasse's Tyson book, Stary Messenger, where he discusses the infrequent ability of us humans to understand exponential growth.

He cites the example of algae on a lake, where for a month you notice it doubling in size day by day, from a small section of a lake, until it covers half of the lake at the end of the month. At this point, when most people are asked to say how long will the algae take to cover the lake entirely, instinctively, they would guess it would need entire month for the other half to get filled, when in reality all it needed was one more day. Such is exponential thinking.

In such a lesson I am stuck, trying to even understand this exponential curve, yet I can only barely wrap around my mind around linear growth. When I first saw models, namely GPT-3.5 come out, I thought them cute, because of their improper logic, failure to handle math, dates, simple logical problems that had tangential context thrown in for surprise.

But it just kept getting better, and cheaper. In two years, from 2022 to 2024, the same level of intelligence output of GPT-3.5 became 285 times cheaper. As I write this post, I am struggling now to consume tokens as fast as they are given to me by my basic $20 a month Codex subscription.

At work I've noticed that GitHub keeps on having availability issues. But the reality is that they are seeing more code pushes than ever before. In just one year from Q1 2025 to Q1 2026, they are seeing code pushes increase by 80%. After another outage, GitHub said in this blog post that:

"Since April, monthly commits have grown from 1.4 billion to 2.9 billion. That growth explains the pressure on our systems, but it does not excuse these outages."

The chart below shows the number of git pushes per month, but stops at Q1 2026. I made some estimates on what Q2 and Q3 would look like, but it's clear that this is exploding.

One other development that surprised me has been the rise of India, as it now is approaching the number of git pushes on GitHub of the United States.

But then I thought, sure, these things can now produce intelligence faster and more cheaply, but then they will not be energy-efficient.

But here again, electrical efficiency is increasing at 40% year, meaning we can get greater intelligence output with less power.

A report comparing the top 500 supercomputers in the world shows that since 1995, these computer can make 312,500,000% faster computations per second and 78% more efficient between 2024 and 2025.

Despite the increases in efficiency, the increase in electricity demand has raised prices by 20%, and there is still so much more compute to come online still in the coming years as new datacenters get built.

As the quote mentions above, I am in a vortex that says that desk jobs will be completely replaced soon, that agents will replace me in a few year, but if we look at data from the US workforce, they estimate that another 185K jobs will be added in the next 10 years:

In 2025, all AI players mentioned in the chart below posted profits every year except for Oracle in 2022, when Oracle paid a lawsuit when CEO Mark Hurd left from Helwett Packard and then got sued due to the potential of him sharing trade secrets at a competitor.

When we look at free cash flow—the operating cash left after capital spending—the story is still pretty similar.