055 - Hardware Dependency

While most major tech companies and startups are busy throwing everything they can into buying the largest quantity of high-end GPUs they can get their hands on, going so far as to treat them like currency, we're laughing at them.

The most advanced AI isn't built on the technology stack of Transformers, RL, or similar systems, and it doesn't rely on GPUs. Google, Microsoft, and others may as well be competing for who can accumulate the largest Beanie Baby collection since they're massively overinvesting in the wrong kind of hardware.

Last year we went so far as to warn Nvidia of the dead end they were headed for, but they appear quite content to keep milking the current cash cow until they fall off the cliff with it.

Most people are already aware that when the "Generative AI" bubble bursts it will reduce the current frenzied demand for GPUs, but the bottom will fall out of that market when every major tech company finds themselves with 100 times more GPUs than actually serves a purpose.

The most advanced systems primarily bottleneck on normal RAM, but even 64 GB of RAM proved sufficient to run a "Global Workspace" on par with that of humans. Running similar systems in real-time where the entire knowledge graph is held in active memory, never requiring time to load and unload, can be accomplished with a single high-memory server.

Normal RAM is extremely energy efficient compared to GPUs, often clocking in at about 1.1v per stick. That difference in energy efficiency directly translates to systems producing far less "waste heat", and much lower cooling costs. Normal RAM also has a far longer hardware lifecycle, with an average of 6.5 years between one generation and the next.

The tech giants backed themselves into a corner with the wrong technology stack, and their stockpiles of very expensive paperweights will serve as a lesson for future generations.