In the 1830’s after the success of the Erie Canal everybody knew that canals were the way of future commerce, and all the states began furiously building canals so they could interconnect and reap the benefits of a national network.
In 1837 there was a recession, and by the time it ended the tax base for the canal build-out had collapsed. When the economy recovered five years later, railroads had become a thing and that’s why there’s a half-finished canal in almost every state in the Midwest somewhere.
There are two scenarios for AI general intelligence success.
One company creates a much better product and can charge a large premium for its service. This corporation will quickly become a large percentage of the S&P 500’s market cap (like 20% to 30%).
The 4 or 5 players in the AI space deliver roughly the same product and price competition keeps user fees low. The AI industry gets a much smaller market cap under this scenario, likely smaller than it’s current value, but we won’t know for a few more years.
Under either scenario, there is likely to be less need for AI data center throughput. Once a successful model is created, they typically run a distillation process on the code to make it run faster and require lower compute resources. There will be a lot of idle data centers under either scenario. Fortunately, there’s likely to be a big demand for mass surveillance of a large, restive unemployed population. The idle data centers can be repurposed to that task.
The main winners will be Apple and Window/Intel. The desktop hardware is necessary for privacy and jurisdiction requirements for professionals, ie accountants, engineers, doctors, and lawyers. When you own the local machine you use AI for free after the software is installed. The DC cloud computing will never make sense on earth. In space is a different econmics.
Meanwhile Moore’s Law times 7 is what has been happening with AI chips. This is quickly turning the chips into commodities.
We are still 80 years plus off from sentient life in the machine. AI does not know death, sex, or why. None of us know why, but AI’s mixed up ideas are not pondering why.
That coder I worked with last week, his disdain for AI coding helped me. I quickly realized to completely discuss the game I am programming with Claude in greater detail. Claude saw some patterns over the last month it had not understood was giving up major code in the main asset to suit a different model of play. Now I can simply run all the components of code through Claude and it will be debugged into the specs and model of play action I want. Claude will stop getting lost in the cross currents of a game being programmed as a work in progress.
The coder with the disdain probably is not moving with the times. Not everybody does.
The Erie Canal was completed in 1825 and was thought to be a costly boondoggle. But it opened easy access to the Great Lakes from NYC. It allowed grain from the midwest to reach eastern markets cheaply. Many other canals followed. The anthracite canals across New Jersey did quite well.
The many midwestern canals failed. Most say its railroads that did them in. The first steam locomotives were not very powerful. By 1840 the technology was much improved. Canals were seasonal and could not compete.
I’m familiar with the Wabash and Erie Canal that followed the Maumee River west from Toledo then connected with the Wabash River to Evansville, IN. Connecting Lake Erie with the Ohio River. Poorly constructed. Sabotaged by residents. The Wabash RR laid rails on the canal banks along the Maumee and took the Wabash name. Terre Haute still has the canal.
Most of the discussions are not about AI in general but about LLMs in particular. Tesla data centers are not going to go broke. Some LLM data centers more than likely will.