We are deep into takeoff of the technological singularity. Things might still feel normal for now, but not for much longer.
This is the Eve of the birth of a new world. For better or worse.
The stakes have never been higher. Navigating this is still quite challenging for all involved. But to understand this is happening is the first step in doing so.
I spent the past week looking at purchasing a little box called a DGX Spark. It’s a little computer you can experiment with running local AI models on.
It isn’t very powerful. People in-the-know called this thing E-Waste. Junk. A brick. They were saying this as recently as a month ago.
This thing was selling for $4k at launch. Then the prices climbed to $4,700. Then this past week it climbed to $5k. Now, it’s gone.
The same thing has been happening to every single piece of hardware that can run AI inference. There are GPUs out there that are 6 years old that are now selling for nearly twice their original price.
In fact, the price is starting to even come up on GPUs that are nearly a decade old.
And for new hardware like the Mac Studio with 256GB of RAM, deliveries are now pushed out to half a year from now.
For those familiar with computers, something like this has effectively never happened.
We are used to computers doubling in speed or processing power every couple of years. This has been the exponential case we’ve been enjoying for over half a century. It has been an enormous consumer surplus; we can get the best processors, with the best memory, with extremely cheap and abundant file storage, right in our phones and home computers.
Not so, anymore.
As I watch any consumer hardware that can run AI disappear from shelves, I feel as if I’m watching the tide roll out from an incoming tsunami.
The reasons for this are several-fold.
The first is that open-source models have gotten good enough. We have AGI at home now. We have a guy that you can put inside your computer and literally lives there, in your house. Not in some data center in the cloud. Not in Sam Altman’s office.
Right there. In your house. And the guy in your computer can fully use the computer. He can use any app in your computer. Any website. Write any code he needs. Do anything you tell him to do.
The open source models are about that good now.
And the next factor driving this: subscription plans are getting more expensive.
People have been complaining a lot lately that their $200 per month AI plans used to feel limitless. But they feel much less so lately. The frontier models are quite large. They can use a lot. And nobody knows how much they’re even supposed to get (this isn’t disclosed, believe it or not).
This making it a lot harder for people to work. They’re on X begging the codex and claude teams for usage limit resets throughout the week. It’s not great, and definitely not a sustainable place to be for your livelihood. So people are turning to open models more and more.
The other couple factors here: the frontier models are super geniuses. They can now solve millennium prizes. They can cure blindness. They can build room-temperature superconductors.
Whatever it is they’re tasked with doing, they appear capable of marshaling the resources to actually accomplish it.
This is terrible news for the worker-consumer. While we might become flush with an abundance of new technologies and medical discoveries, this also means that letting us use these models while they could be busy making these discoveries is, essentially, a waste of compute.
One MORE factor: people are beginning to content with the concept of Recursive Self Improvement (RSI). New and better models are going to come out faster and faster. Some of these models will also get smaller and smaller. And be open source!
What this means is the marginal value that any AI inference hardware can provide is suddenly much more valuable. What used to be graphics card you used to play dumb video games and you were about to get rid of is now a guy inside your computer that does infinite stuff for you. And he’s getting smarter every 2 weeks with every additional model release and optimization for faster inference.
And simultaneously, because demand for these guys in computers are so high, the GPU makers have basically stopped making almost any smaller enterprise or consumer hardware at all. It’s all become extremely scarce, which is forcing the entry point for a lot of this hardware to be closer to $20k minimum, and the bulk of the hardware still being produced is selling for several hundred thousand dollars.
And the FINAL cherry on top of this, is that all those guys in computers in the cloud need their own computers to run a bunch of code on and do tasks on. So that’s causing these AI-agent companies to go buy any last thing they can so those guys can work.
Anyway, I don’t see any of these trends reversing any time soon. I think the coming months, and especially around the holidays, will see a large consumer upset as they begin to realize what the AI boom is costing them in their own electronics consumption availability.
And remember, this is just the tide going out on personal compute availability. These are only the initial signs of the magnitude of changes upon us. The wave hasn’t even begun to hit yet. But we are now seeing its first very real, very tangible signs.