A powerful AI data center can contain thousands of expensive chips.

Each one can perform an enormous number of calculations.

So the formula seems simple:

More chips = more AI power.

But there is a problem.

Those chips need to talk to each other.

And if they cannot do that fast enough, buying more chips does not necessarily make the whole system equally faster.

This is becoming one of the less visible bottlenecks in the AI boom.

The chips are getting faster.

Now the roads between them have to keep up.

Imagine 1,000 People Building One LEGO Model

Imagine 1,000 people trying to build one enormous LEGO model.

Every worker is incredibly fast.

But each person is building a different piece.

One worker finishes a section and needs to hand it to someone else.

Another needs instructions from across the room.

Another is waiting for a piece being held by someone else.

Soon, the problem is not how fast each person can build.

It is how fast everything can move between them.

AI data centers face a similar problem.

A GPU can perform enormous numbers of calculations.

But large AI systems do not run on just one GPU.

They can use thousands of them.

And those GPUs need to exchange enormous amounts of information while they work.

That means some of the most important technology in an AI data center is not inside the chip.

It is between the chips.

One GPU Isn't Enough

Modern AI models can be too large or computationally demanding for a single GPU.

So the work is divided across many processors.

One GPU calculates one part.

Another calculates another.

Then they exchange information.

More calculations happen.

More information moves.

This happens again and again.

The faster the processors become, the more important those connections become.

Because an expensive GPU waiting for another GPU is still waiting.

And a waiting GPU is not producing useful work.

Think About a Highway

Imagine building a city with 1,000 factories.

Every factory is incredibly productive.

But there is only one narrow road connecting them.

Soon, trucks pile up.

Factories wait for parts.

Warehouses fill.

Production slows.

So you build another factory.

But that does not fix the problem.

Eventually, you need:

a bigger highway.

AI infrastructure faces the same problem.

The GPU is the factory.

The network is the highway.

And as companies install more AI chips, the amount of data traveling on those highways becomes enormous.

This Is Why Nvidia Sells More Than GPUs

Most people know Nvidia for GPUs.

But Nvidia has also built a major business around connecting those GPUs.

Its AI networking technologies include:

NVLink

InfiniBand

and

Ethernet.

You do not need to remember the names.

The important part is what they do.

NVLink helps processors communicate at extremely high speeds over short distances.

InfiniBand and high-performance Ethernet can connect larger groups of machines across an AI data center.

Nvidia describes its networking technology as the connective fabric that allows tens of thousands of GPUs — and eventually potentially far more — to work as one coordinated system.

The goal is simple:

Make thousands of separate processors behave more like:

one enormous computer.

That requires more than fast GPUs.

It requires fast roads between them.

Now Other Companies Want to Build the Roads

On September 15, 2026, Reuters reported that a startup called Delos Data had raised:

$100 million.

The company was founded by former Intel engineers.

But it is not trying to build another general-purpose AI GPU.

Instead, Delos is developing chips and software designed to move data faster inside increasingly complicated AI data centers.

That distinction matters.

The company is not betting that the world simply needs another processor.

It is betting that the processors already being installed need better ways to communicate.

Why would investors put $100 million into something most consumers will never see?

Because AI data centers are becoming more complicated.

They may contain Nvidia GPUs.

AMD accelerators.

Specialized AI processors.

Custom chips.

Different machines may perform different jobs.

But they still need to exchange data.

The more complicated the system becomes, the more important the roads become.

Bad Communication Wastes Expensive Hardware

This is where a networking problem becomes a money problem.

AI accelerators are expensive.

The servers surrounding them are expensive.

The buildings are expensive.

The electricity is expensive.

And the cooling is expensive.

Now imagine paying for all of that infrastructure only to have expensive processors spend part of their time waiting for data.

Former Intel CEO Pat Gelsinger, whose venture firm participated in the Delos funding round, pointed to this problem.

Poor communication between chips, he argued, can leave expensive hardware consuming electricity while waiting for information.

That means inefficient networking wastes more than time.

It can waste money and energy.

So the important question for an AI data center is not simply:

How many GPUs did we buy?

It is:

How much useful computing can we get from the GPUs we already bought?

A company can own an enormous amount of computing power on paper.

But if the processors cannot work together efficiently, some of that expensive capacity is wasted.

This Is the HBM Problem — One Level Bigger

Inside an AI accelerator, there is already another data-moving problem.

The GPU needs information from memory.

That is why HBM — High Bandwidth Memory — has become so important.

HBM helps deliver enormous amounts of data to the processor quickly enough to keep it working.

Now zoom out.

The GPU needs data from memory.

Then it needs to exchange data with other GPUs.

Zoom out again.

Servers need to communicate with other servers.

Then racks need to communicate with other racks.

The AI system starts to look less like a collection of chips and more like:

a giant machine for moving data.

And every step can become a bottleneck.

The Highway Itself Is Becoming Expensive Technology

A normal office network does not face the same problem.

If an email takes a fraction of a second longer to arrive, almost nobody notices.

AI computing is different.

Thousands of processors may be working together on the same job.

Small delays can accumulate.

Congestion matters.

Bandwidth matters.

Latency matters.

That is why companies are building specialized networking equipment specifically for AI.

Nvidia's current ConnectX-9 SuperNIC, for example, can deliver up to:

1.6 terabits per second of throughput.

That number is difficult to visualize.

The important point is not the exact speed.

It is that networking equipment itself has become extremely high-performance computing technology.

Nvidia says ConnectX-9 is designed to support what it calls "gigascale AI factories."

And the network is becoming more specialized.

Nvidia's Spectrum-X Ethernet platform is designed specifically for AI workloads rather than ordinary office networking.

The company says its latest platform supports networking across large AI clusters and even between multiple AI data centers.

The road itself has become sophisticated technology.

And that creates another market inside the AI boom.

Even the Cables Are Changing

Keep zooming in and another bottleneck appears.

How do you physically move all this information?

Copper cables have limits.

Traditional optical connections consume power.

As AI clusters become larger, the physical connections between machines also have to evolve.

Nvidia is now developing networking systems that integrate optical technology more closely with switches.

Its co-packaged optics products are designed to reduce power consumption and increase bandwidth density as AI clusters grow.

Some of these systems are being designed for connections of:

1.6 terabits per second and beyond.

Think about what has happened.

We started with an AI model.

Then we needed a GPU.

Then faster memory.

Then faster networking.

Then better switches.

Then faster optical connections.

One bottleneck reveals the next.

Follow the Bottleneck

The AI story is often reduced to one company:

Nvidia.

But zoom out and the system becomes much larger.

↓

GPU

↓

HBM

↓

Networking

↓

Servers

↓

Power

↓

Cooling

Every layer exists because the layer above it needs something.

A faster GPU needs faster memory.

Thousands of GPUs need faster networking.

More servers need more electricity.

More electricity creates more heat.

More heat requires more cooling.

Solve one bottleneck and another one becomes visible.

And when a bottleneck becomes important enough, money starts flowing toward companies trying to remove it.

AI Is Becoming an Infrastructure Business

This helps explain why some of the businesses benefiting from AI do not look like AI companies at first.

Memory manufacturers.

Networking companies.

Data-center developers.

Transformer makers.

Cooling companies.

Power producers.

None of them writes the answer you see on your screen.

But all of them help make that answer possible.

And the more expensive the processors become, the more valuable it becomes to keep those processors working efficiently.

The famous part of AI is the chip.

The less visible part is everything required to stop that chip from waiting.

First, the GPU needs data from memory.

Then it needs data from other GPUs.

Then thousands of processors need to behave like one enormous computer.

The world's fastest factories are not very useful if the roads between them are jammed.

Sometimes the valuable part isn't another factory.

It's the highway connecting them.

BEYOND THE OBVIOUS.

Related on OVIQQ

Before AI chips can communicate with each other, each chip has another data problem of its own: getting information from memory fast enough.

Why Does AI Need So Much Memory?

Sources

Reuters — Delos Data, a Chip Startup Founded by Intel Veterans, Raises $100 Million for AI Networks
Reuters source
NVIDIA — ConnectX InfiniBand Adapters
NVIDIA ConnectX InfiniBand Adapters
NVIDIA — Spectrum-X Ethernet Networking Platform
NVIDIA Spectrum-X
NVIDIA — Optical Transceivers and Cables for AI Networking
NVIDIA AI Networking Interconnects
NVIDIA — ConnectX-9 SuperNIC Firmware Release Notes
NVIDIA ConnectX-9 documentation

Search This Blog

OVIQQ

Why Do AI Chips Need to Talk to Each Other So Fast?