The entire foundation of computing is coming apart.
But there’s no need to panic. Because it’s happened before.
In the early days of the internet, one server did everything. It handled traffic, stored data, delivered content and kept websites running.
That worked… until it didn’t.
As more people came online, those machines started to struggle. So a new kind of infrastructure emerged.
Instead of one machine doing everything, each task got its own solution. Routers directed traffic, while storage systems handled data. Some systems moved data closer to users. Others spread out demand.
That specialization is why companies like Cisco (Nasdaq: CSCO), Amazon (Nasdaq: AMZN) and Google (Nasdaq: GOOG) became so important during the internet buildout.
They were each attempting to make a part of the internet work better.
The same thing is happening again today.
Only this time, it’s happening with the chips that power artificial intelligence.
The End of General-Purpose Compute
For decades, the central processing unit, or CPU, has been the center of gravity in computing.
Image: Wikimedia Commons
It’s flexible and reliable enough to handle most workloads, which makes it incredibly valuable in a world where computing needs are relatively simple.
But AI’s needs are far from simple.
Training AI models takes a lot of computing power. Running them at scale requires speed and efficiency. And both depend on moving huge amounts of data without slowing things down.
So the old model of relying on a single, general-purpose CPU doesn’t work anymore.
That’s why the AI industry is now assigning each task to a chip designed specifically for it.
Graphics chips, or GPUs, have long been the go-to for training AI because they can handle a lot of calculations at the same time.
Image: Wikimedia Commons
From there, customization has spread.
Google has its TPUs, which are custom-designed AI chips for training and running models.
Amazon has its Trainium chips for training and Inferentia chips for running A models.
And Microsoft is building its own Maia chips to improve how its systems run.
Even memory isn’t just a supporting component anymore. In many cases, it’s just as important as compute itself.
High-bandwidth memory, or HBM, has become a critical piece of the system because AI needs to feed data into chips fast enough that they don’t sit idle.
Some analysts estimate the HBM market will reach $54.6 billion in 2026, up 58% from the prior year.
Image: globalxetfs.com
Demand for AI memory is now so strong that supply is being locked up years in advance.
And it’s becoming a real bottleneck.
SK Hynix, one of the world’s largest memory chipmakers, says much of its high-end memory for 2026 is already sold out.
That’s why I pounded the table about Micron Technologies (Nasdaq: MU) in Strategic Fortunes when DRAM prices started skyrocketing in late 2024. I could see where this was going.
But memory isn’t AI’s only constraint.
Power is starting to limit how fast new AI infrastructure can be built too. Training and running AI models also require enormous amounts of electricity, and in some cases, access to power determines where new data centers can even go.
In other words, AI has been growing so fast that bottlenecks are popping up everywhere.
Because of this, companies are being forced to redesign how everything works together.
That’s why the biggest AI infrastructure players are now designing their own chips. Because even small efficiency gains at the chip level can translate into massive advantages across their entire AI systems.
Amazon, Google, Meta (Nasdaq: META) and Microsoft (Nasdaq: MSFT) alone are on track to spend around $665 billion on AI infrastructure in 2026.
One reason behind this enormous amount of spending today is that the industry is breaking computing into pieces and rebuilding it in a more specialized way.
Data centers are no longer built around interchangeable machines. They’re being redesigned as tightly integrated environments where different types of chips handle different parts of the workload.
So compute, memory and networking are all being optimized together.
This also happened in the Internet era, when computing evolved from standalone servers into layered systems. Each layer handled a specific function, and together they created a faster, more scalable network.
That’s what’s happening inside AI infrastructure today.
It’s a leading reason why the semiconductor market is growing so quickly right now.
Because demand isn’t just increasing in volume, it’s also increasing in complexity. And that’s pulling the entire semiconductor industry in a new direction.
From general-purpose chips…
To purpose-built systems.
Here’s My Take
The real story here is that AI isn’t just changing what compute looks like. It’s changing who controls it.
We’re moving away from a world where general-purpose chips could be bought by anyone and used for almost anything. That made computing widely accessible.
But specialized systems don’t work that way.
They require custom chips, tightly integrated hardware and massive amounts of capital to build and operate. And that naturally concentrates power in the hands of the companies that can afford to build and run them.
This isn’t new.
During the internet buildout, profits didn’t stay evenly distributed. It concentrated in the companies that controlled key layers of its infrastructure.
The same thing is starting to happen again.
Only this time, it’s happening at the foundation of computing itself.
And it means the gap between the companies building AI infrastructure and everyone else is likely to widen.
Regards,
Ian KingChief Strategist, Banyan Hill Publishing
Editor’s Note: We’d love to hear from you!
If you want to share your thoughts or suggestions about the Daily Disruptor, or if there are any specific topics you’d like us to cover, just send an email to [email protected].
Don’t worry, we won’t reveal your full name in the event we publish a response. So feel free to comment away!
