From Dial-Up to AI: It Was Always About Bandwidth

It was 1995. I was a Sales Rep for a Telecom Service Provider.

Internet dial-up services were starting to become available to the public. Many early users connected at painfully low speeds: 9.6, 14.4, 28.8 and later 33.6 kbps.

Many corporations were buying 64 kbps leased lines — and that felt serious.

If you met with the IT Manager, he was usually sitting in a big chair, maybe smoking a Cuban cigar, and his business card might as well have said:

“I have the bandwidth.”

CIO roles were not yet as common or as powerful as they later became. But the person who controlled connectivity controlled a very important part of the company’s future.

By 1997, fiber optic networks were becoming available. A branch could get 2 Mbps. A main site could reach 155 Mbps.

Meanwhile, most people were still accessing the internet through dial-up modems.

After 2000, dial-up started to give way to ADSL and cable access. The real change was not only speed. It was the shift to always-on connectivity.

Consumer bandwidth started to scale.

Huge investments went into infrastructure to satisfy the need for more bandwidth. Some of those investments were questionable. Some were unsustainable.

Then the bubble popped.

The demand was real. The timing was wrong.

We built capacity before the applications, devices and user behavior were ready to absorb it. We ended up with large amounts of infrastructure ahead of demand.

The network layer had moved faster than the application layer, the device layer and the user layer.

But eventually, the use cases arrived.

Mobility. Social networks. Video. Cloud applications. E-commerce.

That capacity was eventually filled — and then the world demanded orders of magnitude more.

If AI came to mind while reading this, that is exactly the point.

Today’s AI infrastructure has a similar problem. Demand is real. The use cases are real. The strategic importance is real.

But the economics depend on whether infrastructure can move data fast enough, cheaply enough and intelligently enough.

Right now, many LLM interactions behave like brilliant professors with fragile short-term memory: they do not truly remember; they reconstruct context.

They process prompts. They retrieve documents. They use caches. They re-read large amounts of information. They consume compute to rebuild what should ideally be available as structured memory.

A model that has to re-read everything is expensive.

A model that can access structured, relevant memory is economically different.

The next efficiency frontier is memory: not just storing more, but retrieving the right context at the right time with less compute.

This is still a bandwidth problem — only now the bottleneck is not just the last mile.

It is memory bandwidth, GPU-to-GPU bandwidth, chip-to-memory bandwidth, data-center interconnect and the cost of moving tokens through the system.

Some answers are incremental: better memory, better networking, better accelerators, better caching and better model architecture.

Others are more radical: in-memory computing, optical interconnects and neuromorphic approaches.

This is why memory, networking, accelerators and data-center infrastructure are becoming strategic again.

The market is not only paying for more compute.

It is paying for faster movement of data.

Many of today’s most important memory and networking investments are aimed at one thing: making AI infrastructure economically scalable.

Without that, AI scales technically — but not economically.

Look at the strategic attention around HBM, advanced memory, GPU interconnects, Ethernet fabrics and AI data-center networking.

Companies such as Micron, SK Hynix, Samsung, Broadcom, Arista and others are no longer just “plumbing.”

They are part of the AI bottleneck.

And bottlenecks create value.

The pipes changed.

The lesson did not.