On Mon, 3 Aug 2026 at 06:25, Simon Richter <[email protected]> wrote:
> > Everyone points at the dark fiber post Internet-bubble which took ten
> > years to eventually light up as the analog to the AI bubble. But what
> > most people forget was that after the dot.COM bubble, there were
> > $100k+ USD Sunfire servers that were selling for pennies on the
> > dollar, and Sun servers ten years later don't hold value the same way
> > dark fiber did. So that might be the closer analog for the portion of
> > the AI bubble which comes from people buying Nvidia GPU's.
>
> Exactly my point: there is very little being built that will last, the
> GPUs that were used in a datacenter are losing value even faster than a
> ten years old Sun server, and most of the other hardware is
> special-purpose.
Part of the assumption of the AI bubble bursting is that the foundation of
AI development sticks around and more of it becomes free as eitehr open
weighs or trully free AI models. Say 80% of AI companies go bust or get
consolidated and scale back. There are now thousands of very depricated
big-AI-capable server racks on the secondary market or on the
rent-by-the-hour market, along with free models available for download that
are similar in capability to current fronteir models and are large enought
that running them at home is not really possible. That opens up the
possiblity of all the small and medium companies (and also a good chunk of
enthusiasts) picking up that hardware on pennies on a dollar and running
their own models for their own uses in a completely independent way. We get
wide-AI instead of the early tall-AI situation. Similarly how everyone had
to use Geocities or Blogger.com to host things early on, but after a while
could self-host with their own hardware or choose one of thousands of
smaller hosting providers locally.
And this is assuming there is *absolutely* nothing that can use that
hardware. There are big chunks of science where we can use big floating
point compute, if it is cheap enough. There is a ton of data coming out of
all kinds of ground and space-based observatoreis that can be analysed in
millions of different way to gather new science, for example. Protein
folding research can eat up all the compute in the world and still come up
wanting more.
Sure, it is annoying that DDR5 prices are insane right now. But worrying
about resources being wasted in the AI bubble bursting? I don't worry about
that at all.
--
Best regards,
Aigars Mahinovs
