[felixfurtak]

People keep comparing the AI boom to the Dotcom bubble. They’re wrong. Others point to the Railway Mania of the 1840s — closer, but still not quite right.

The real parallel is Canal Mania — Britain’s late-18th-century frenzy to dig waterways everywhere. Investors thought canals were the future of transport. They were, but only briefly.

Today’s AI runs on GPUs — chips built for rendering video games, not thinking machines. Adapting them for AI is about as sensible as adapting a boat to travel across land. Sure, it moves — but not quickly, not cheaply, and certainly not far.

It works for now, but the economics are brutal. Each new model devours exponentially more power, silicon, and capital. It just doesn't scale.

The real revolution will come with new, hardware built for the job (that hasn't been invented yet) — thousands of times faster and more efficient. When that happens, today’s GPU farms will look like quaint relics of an awkward, transitional age: grand, expensive, and obsolete almost overnight.

[l9o]

I think specialized hardware will emerge for specific proven workloads (transformer inference, for example), but GPUs won't become obsolete. They'll remain the experimentation platform for new architectures. You need flexibility to discover what's worth building custom silicon for.

Think 3D printers versus injection molds: you prototype with flexibility, then mass-produce with purpose-built tooling. We've seen this pattern before too. CPUs didn't vanish when GPUs arrived for graphics. The canal analogy assumes wholesale replacement. Reality is likely more boring: specialization emerges and flexibility survives.

[roommin]

Sure, but your R&D infrastructure isn't going to be 1.5 trillion dollars.

[fhennig]

> Today’s AI runs on GPUs — chips built for rendering video games, not thinking machines. Adapting them for AI is about as sensible as adapting a boat to travel across land.

A GPU is fundamentally just a chip for matrix operations, and that's good for graphics but also for "thinking machines" as we currently have them. I don't think it's like a boat traveling on land at all.

[ozgung]

Another definition: A modern GPU a general purpose computer that can make parallelized and efficient computations. It's optimized to run limited number of operations but on large number of data points.

This happens to be useful both for graphics (same "program" running on on huge number of pixels/vertices) and neural networks (same neural operations on huge number of inputs/activations)

[aurareturn]

Nvidia’s enterprise GPUs having nothing to do with graphics anymore except for the name.

[felixfurtak]

GPUs are massively parallel, sure, but they still have a terrible memory architecture and are difficult to program (and are still massively memory constrained). It's only NVidia's development in cuda that made it even feasible to create decent ML models on GPUs.

[realaaa]

I think it'll be a combination of hardware of course, but also better software - surely there is a better way of doing this (like our brains do) which will eventually require less power