United for sustainability Climate Crisis Club 

Main Menu

The AI bubble

Started by K-Dog, May 27, 2026, 02:29 PM

Previous topic - Next topic

K-Dog

#15
Quote from: RE on Nov 21, 2026, 11:30 AM
CEPR.NET 2026-11-20

The Economics Behind the Collapse of the AI Bubble

The decline in AI stock values stems from unrealistic expectations clashing with economic realities as affordable, privacy focused Chinese alternatives outperform costly U.S. models in practical applications. Most users prioritize efficiency and customization over high priced frontier systems that fail to generate sufficient revenue. The market is correcting toward sustainable utility rather than speculative hype.

Tic Toc.

RE
Quoteprivacy focused Chinese alternatives outperform costly U.S. models in practical applications.

Private because they can be hosted locally and that also makes them practical.  This is actually unfair because apples are being compared to oranges.  Local models vs Data centers, data centers that run on a Chinese budget are localish.  The author gets the right idea through a different perspective.

His comparison muddles together the Mixture Of Experts technology which China implements since they do more with less hardware.  Excess firepower always being the American philosophy is good for the weapons industry, that in this case being NVIDIA. 

MOE tech allows large parts of models to only be loaded into GPU memory as needed as apposed to having to load all of a model at once.  American data centers will keep models in memory and switch users between them.  Local users and countries with fewer resources. (no foolish investors for them) need the ability to manage GPU memory more efficiently.

The article summary above was generated using an 80 GB model which loads this way.  It lets me use the GPUs for other things as it loads what it needs from memory and it does not need that much GPU space.  Since I built a workstation I get excellent results.  My memory swaps are very fast, and I can keep the whole model in RAM.  It takes about 20 seconds to generate a summary.

This author sort of knew what he was writing about.

He sniffs the foul odor from the financial side, and I sniff out the same thing from the technical side.  It is a VERY dangerous thing to build hardware because software needs it.  It is actually a huge no-no because software can change much faster than hardware can.  And that is exactly what has happened.

Countless investors were sold on the need for data farms.  How could you host a frontier model without one.  Be a data center king and rule!  How could you know that technical advancements would lower the hardware requirements leaving data centers tokenless.

Forgetting to consider that your data center might be obsolete before the concrete gets hard.  At least they did not use vacuum tubes.  The power requirements for that would melt Antarctica.  Which some people will want for real estate.  American politicians would approve such a plan for sure.  It is a good thing chips were invented before LLMs or we would already be at 2 degrees C.  Training an LLM with vacuum tubes would require the same land as covering the moon with one giant data center.  As Elon would be happy to have you pay for a moon sized data center, it is a good thing tech moved on to the transistor*.

I have a 260 GB model on disk and I do not have enough memory to keep it all in RAM so when I use it layers load from disk.  It is slow, but usable.  I may have to have a cup of coffee to get results, and I have only used the model once so far to test it, but the results were damn good. Colibri is the project.  The quality is so good waiting for the results is worth it sometimes.

* If small vacuum tubes were used to store 16 bits for a single parameter of an 80GB model using a square foot of space is a reasonable guess.  Each floor in a building can store several square foot assemblies in a rack of some kind in a square foot of space.   20 assemblies on top of each other, on each floor of a 10 story building.  That is 200 / square foot of building.  The building is square so each side is ( 80000000000/200)1/2 = 3.79 * 5280 ) or it is almost four miles on a side.  That much room without walkways, power and cooling distribution, or parking lots. 
Currently eight sticks of memory on the computer's motherboard that I am using to write this has more memory that that one building using vacuum tubes which would be at least four miles on a side would have.
That would be the bubble of all bubbles.  The analogy taken to the absurd makes the argument's current thesis very reasonable.