Today’s data centre build-out dwarfs previous infrastructure investments

Kathy Gibson reports from Vertiv Week in Zagreb – Artificial intelligence (AI) is driving the biggest infrastructure build-out in human history – bigger than the railroads or the Industrial Revolution.

“People assume AI is a bubble, but it is showing no signs of slowing,” says Rod Evans, vice-president: AI, cloud and infrastructure at Nvidia.

He points to Nvidia’s financial results, which have broken corporate records for three quarters in a row, and has just announced that the coming quarter will show healthy growth again.

“This is no bubble,” he stresses. “The take-up of agentic AI has been extraordinary, and something fundamental is happening.”

With model parameters growing by 10x every year, their performance and complexity how grown by 1-million-times in the last decade. “Before this, the best we ever saw was 100-times,” Evans explains.

Although the number of reasoning workloads is growing at a rapid rate, the cost per token has come down by a factor of 10x every year.

Overall, Evans says global data centre demand is going to double by 2030.

Although most of the new capacity is planned in the US, many of these are being delayed so only about one-quarter of them will actually come online.

“So Europe is going to fill the gap,” he believes. “We are seeing massive build-outs all over Europe and North Africa that we wouldn’t have imagined two years ago.”

Driving the need for new data centre capacity – and a different type of data centre – is that fact that inferencing workloads are growing to where they will account for about 60% of all workloads by 2030.

And, because inferencing needs to happen close to where it is executed, Evans says this is going to drive pressure on metro-proximate colocation sies and increase demand for data centres near major centres.

AI token use is booming, and is on track to multiple by 24-times by 2030, for a projected consumption of 118-quadrilion tokens.

This is largely because agentic and reasoning models consumer 15-times more tokens per request the GenAI does.

“And this means AI factories are the new trend for data centres,” Evans says. “So AI is not only changing how we work, it is changing the face of compute, data centres and sustainability.”

The million-x improvement in compute over the last 10 years has seen data centres evolve to liquid cooling, and from AC to DC power.

“These AI factories are the industrial infrastructure of the AI era,” Evans stresses. “And they need to get to token production as quickly and efficiently as possible.”

He describes the AI infrastructure as having five layers: energy at the bottom, powering the chips, then the infrastructure, the models and the applications on the top.

“You cannot ignore any of the layers.”

It sounds simple, but there is a tremendous amount of design complexity that goes into each data centre, Evans says, while the deployment cost is huge. “So you have to ensure you get it right.”

The data centre rests on three pillars: the hardware infrastructure, the facilities infrastructure; and the software.

“This full-stack innovation drives performance and efficiency.”

Nvidia is possibly best known for providing the GPUs, but is also supplies reference architectures and modularised building blocks to help build a complete data centre.

Ongoing demand is driving the need to speed the building and deployment of data centres. – and these new data centres are quicker and less expensive if they are standardised and modular.

“There are no empty data centres,” Evans says.

Deliverable modules are key, since data centre operators are under pressure to maximise their capability per watt.

The post Today’s data centre build-out dwarfs previous infrastructure investments appeared first on IT-Online.