The AI factory took center stage at Vertiv Week 2026 in Zagreb, Croatia. Team TECHx Media followed the event from the venue. One of the standout sessions came from Rod Evans, EMEA VP for Supercomputing and AI Cloud Infrastructure at NVIDIA. He gave a wide view of converged infrastructure for the AI era.

AI Is Driving a Huge Infrastructure Build-Out

Evans opened with a bold point. He said AI is behind the biggest infrastructure build-out in human history, larger than the railroads or the industrial revolution.

Many people still call it a bubble. Evans disagreed. He pointed to growth of 17% in Q1 results, another 17% in Q2, and a forward-looking 18% growth for Q3.

He also shared a telling example from RAISE, a conference in Paris. He asked 26 executives from IT companies and major enterprises across Europe if they used generative AI every day. All 26 raised their hands. A year earlier, he doubts he would have seen even one.

Model Growth and Falling Token Costs

Demand is rising fast. Model parameters are growing about 10x per year. That adds up to a million-fold gain in complexity and performance in ten years.

By comparison, the CPU era of Moore’s law delivered at best a 100x gain in a decade. Reasoning workloads are also climbing steeply. At the same time, token costs have dropped about 10x per year.

Data Center Demand and the Delivery Gap

Global data center demand is set to double by 2030. AI workloads account for about 71% of that change.

However, projects often struggle with permits and timelines. In the US, plans call for 12 gigawatts of new data centers in 2026 and 36 gigawatts in 2027. The expected delivery is closer to 5 gigawatts in 2026 and 7 gigawatts in 2027. That is roughly a quarter of planned demand.

As a result, Europe is expected to fill much of the gap. Evans listed build-out programs in Spain, Portugal, and the Nordics. He also named projects in Albania, North Macedonia, Croatia, Poland, and Romania. In addition, he noted interest in North Africa, including Morocco and Tunisia, where power and land are available and the region is closely connected to Europe.

Inference Moves Closer to the Edge

Inference is also changing the picture. Its share is expected to rise from 48% to 60% over the next five years. Global AI inference demand could reach about 93 gigawatts by 2030.

Evans expects inference to appear more at the edge. One attendee said edge computing did not really take off for general compute, but it will for AI. He compared it to web caching in the 2.5G and 3G mobile days, when operators placed caching in telephone exchanges to speed things up. Likewise, a phone request cannot travel all the way to California for inference. It will happen locally.

Tokens and Agents

Token use is set to surge as well. AI agent token use is forecast to grow 24 times by 2030, reaching 118 quadrillion tokens. Evans cited one large search provider’s volume, which rose from 9.7 trillion a month in May 2024 to 3.2 quadrillion in May 2026. He described that as a 7x jump in one year. Reasoning models consume most of the tokens generated today.

He shared a personal example too. At NVIDIA, engineers get unlimited tokens. Evans uses agents to sort through about 400 emails a day and even draft replies.

What Is an AI Factory?

So what exactly is an AI factory? Evans described it as the industrial infrastructure of the AI era. You put power in one end and get tokens out the other.

The journey has moved through three stages:

• An 8-way server system, the state of the art a couple of years ago.
• A rack with 72 tightly integrated GPUs.
• The AI factory itself, seen in containerized data centers and the new DSX design.

Cloud partners ask one question first: what is my time to first token? That is how they generate revenue. Therefore, every AI factory must reach production quickly and efficiently.

The Five-Layer Cake

Jensen Huang describes AI as a five-layer cake. The layers are energy, chips, infrastructure (the AI factory), models, and applications. Evans stressed that no layer can be ignored. Focusing only on applications while neglecting infrastructure creates problems.

Design complexity is high. When 72 GPUs are interconnected, the network becomes as fundamental as compute. In Evans’s words, the network is the computer. Costs are also large. Buying 10,000 GPUs of the Vera Rubin generation costs a little under $2 billion. Consequently, deployment quality and time to value matter greatly.

Three Pillars and Business Models

Evans outlined three pillars of the AI factory: hardware (racks and network), facilities infrastructure, and software.

He spends much of his time on business plans rather than hardware. His goal is to help AI cloud partners build sustainable businesses. He connects them with off-take partners, models, and software stacks to fill gaps in their plans.

Regional Autonomy and Open Models

Evans prefers the term regional autonomy over sovereign AI. He recalled a question Jensen Huang asks national leaders: will you be an AI taker or an AI maker? Countries will not have a choice about AI arriving. Those that build AI in their own economies stand to gain most.

Open models support this idea. NVIDIA has released many open models, and Evans said open models are advancing every year. They let companies keep critical data in their own country and use other regions for other work.

Still, Evans does not see a purely European company. He gave the example of BMW, which has plants in China, South Africa, South America, North America, and Europe. Global thinking is needed, along with careful choices on where data sits.

Reference Architectures and Exemplar Cloud

NVIDIA offers reference architectures for repeatable results. Partners who follow them can expect predictable performance. Through the Exemplar Cloud program, partners run benchmarks to show their infrastructure lands within 5% of NVIDIA’s own results.

OEM partners adopt the architectures and add their own features. A design review board then checks each solution. Deployment can be a pod at a partner, on premise, or in the cloud. Evans said the company does not mind which route is used.

Physical AI Arrives

Evans called physical AI the next big area of growth. He pointed to the recent robot Olympics in China, where robots set records and showed striking skill.

He also cited a Daily Telegraph article from November. Its headline said Western executives came back from China terrified of what they saw. Factories there had few or no people, only robots. The article quoted Ford’s chief executive, who said companies must change to survive.

Evans linked this to restructuring at Jaguar Land Rover and job cuts at Volkswagen. Driverless taxis are another example. A friend’s wife in Phoenix uses Waymo everywhere and feels safe. Evans expects rollouts in major cities within 12 to 18 months, with London announced early next year.

Financing and the DSX Blueprint

At GTC in Santa Clara in March, Jensen Huang met 20 chief executives from major data center partners. He asked why lenders fund apartment blocks without pre-sales, yet data centers need 70% sold first. Evans noted that there are no empty data centers today, and GPUs are hard to find.

The second point was inconsistency. Builders design data centers in different ways. To help, NVIDIA created the DSX blueprint and released it to the industry as open source. Evans said the same approach was used for the rack-based design. He added that Apple built a new internal server range using it. The goal is faster financing and faster builds.

Power, Cooling, and Space

Every data center has finite resources: power, cooling, space, and connectivity. Connectivity usually takes care of itself. The other three can be controlled.

Evans warned against building a gigawatt site as one big building. It could take ten years to see the first output. Modular delivery, with compact cooling, power, and space, is a better path.

Efficiency is the next challenge. On a gigawatt site, only about 60% to 70% of grid power becomes useful output. Also, a 100 megawatt grid contract may deliver only 93 megawatts. For that reason, Evans advises on-site generation such as gas, hydro, solar, or wind to close the 7% gap.

Smarter Power Use With Max-Q

Energy savings can be large. Evans said saving 50 terawatt hours of annual energy equals about 35 metric tons of CO2, or the energy use of 6.8 million homes.

He also explained a power-tuning approach. A new rack can draw up to 224 kilowatts at maximum power. Running in Max-Q mode lowers that to 164 kilowatts per rack. Therefore, more racks fit within the same site. Operators can run at full power for training and switch to Max-Q for inference.

800 V DC and Megawatt Racks

The next rack generation, due in 2028, will approach one megawatt per rack. Traditional power would need about 32 cables into one rack. That will not work.

The answer is 800 V DC distribution. AC power enters the data center and converts to 800 V DC. At the rack, it steps down to 54 V DC, a standard telco level. Evans hopes efficiency can rise from 60% to 70% up to 80% to 85%. Vertiv is working with NVIDIA to make 800 V DC a reality.

The DSX framework covers power optimization, infrastructure software, and platform software. More than 250 industry partners are working on it. Evans said this gives builders confidence that designs will not be made obsolete.

Liquid Cooling and Heat Reuse

Cooling is moving to full liquid designs. The Grace Blackwell generation uses direct liquid cooling to the chips. The Vera Rubin generation goes further, with cable-less racks and liquid cooling for all switches and power supplies.

Water enters at 45°C and leaves at 65°C. This works in most places worldwide. The warm water can then be reused. Examples include district heating in the Nordics and warm water tanks for salmon spawning. Evans also described drying wood on site in forested regions, so water is not shipped across the country inside logs.

Closed-loop cooling reduces chillers, evaporative coolers, and water use. About 95% of the water is recycled.

Evans said NVIDIA has worked closely with Vertiv on data center designs with liquid cooling. He cited these gains from direct liquid cooling:

• 20% lower annual cooling energy consumption
• 40% less rack space
• 50% faster deployment
• 40% smaller footprint

Digital Twins Shape the AI Factory

Building a gigawatt AI factory is a huge task. A video shown at the session said it takes tens of thousands of workers, nearly 5 billion components, and over 200,000 miles of wire. That is nearly the distance from the Earth to the moon.

The NVIDIA Omniverse blueprint for digital twins helps plan these sites before construction. Engineers combine 3D and layout data with power and cooling systems from Vertiv and Schneider Electric. They also use NVIDIA Air to simulate network logic, layout, and protocols.

Cadence Reality simulates air and liquid cooling. ETAP from Schneider Electric simulates power block efficiency and reliability. Teams can run what-if scenarios in seconds instead of hours. This cuts errors and speeds bring-up. It also helps test the cost and downtime of retrofits.

Final Thoughts

Evans closed with a quote from Jensen Huang at Davos. Twenty years ago, all of this was science fiction. Ten years ago, it was a dream. Today, we are living it. AI has crossed from generating answers to performing work, and the next industrial revolution will be built on it.

The message from Vertiv Week 2026 is clear. The AI factory depends on power, cooling, networks, software, and partnerships working together. Data center builders, operators, and regions across Europe will all play a part.

Related post

View all