Megawatt Racks and Campus-Scale Systems: Data Center World 2026 Shows AI Is Rewriting the Rules

The data center is no longer a room full of servers. It is a single, tightly integrated compute system purpose-built for AI. That was the message from engineering leaders at Oracle Cloud Infrastructure, Nvidia, and Google at Data Center World 2026, as reported by Data Center Knowledge’s Shane Snider.

What is new here is not that AI demands more compute. What is new is that the unit of design itself has changed. The industry has moved past incremental improvements to power and cooling. It is now redesigning the entire facility, and the campus, from the ground up.

Ram Nagappan, vice president of AI infrastructure at Oracle Cloud Infrastructure, described a fundamental split. AI workloads now come in two distinct patterns. Large-scale training connects tens of thousands of GPUs in tightly coupled clusters where latency and proximity are everything. Distributed inference prioritizes availability and responsiveness at a broader scale. A single facility must now support both.

That is a harder engineering problem than supporting one or the other. It forces tradeoffs in layout, resilience, and network design. “You have to take both into account when you build the data center,” Nagappan said.

The density numbers tell the story. Varun Sakalkar, distinguished engineer in Google’s datacenter technology and systems group, said racks that once pushed 30 to 40 kW are now measured in hundreds of kilowatts. Designs are approaching the megawatt range. That is not a gradual curve. It is a step change.

Sakalkar described the resulting environment as bimodal. Traditional compute and storage infrastructure continues on a gradual density curve. AI systems operate on a much steeper trajectory. The data center must support both simultaneously. “We’re not designing a rack anymore – we’re designing a system,” he said.

That system-level thinking is the key takeaway. It is not about better cooling or more power. It is about treating the entire facility, and the campus, as a product.

Power is the binding constraint. Sean James, distinguished engineer for energy systems at Nvidia, said operators are increasingly relying on on-site generation to accelerate deployment. But he called those approaches temporary. “Behind-the-meter power is a good stopgap,” James said. “It’s not the preferred long-term solution.”

The real challenge is grid capacity. Operators are working to secure grid-connected capacity while adding energy storage to manage increasingly volatile AI workloads. Training clusters introduce sharp, dynamic load patterns. “You can see that impact all the way back at the power plant,” James said, describing how generation must ramp to match workload behavior.

Energy storage is becoming essential to smooth those fluctuations. It maintains power quality and meets emerging grid requirements such as ride-through during voltage events. The data center is no longer a passive consumer of power. It is an active participant in grid dynamics.

Cooling has moved past debate. “Liquid cooling is here,” Sakalkar said. “At this point, the conversation is about standardization.” Operators must manage hybrid environments where liquid-cooled AI systems coexist with air-cooled infrastructure. That mix complicates both design and long-term planning.

James pointed to scaling challenges inside liquid systems themselves. Component supply chains and the number of connections required inside high-density racks are becoming constraints. Water use is emerging as both a sustainability and operational risk. “Data centers need to engineer out water where they can,” he said.

The timeline is compressing. Operators are responding by shifting work off-site and standardizing designs. James described an approach that relies on front-loaded design to ensure flexibility across GPU generations, increased use of prefabrication and factory integration, and modular architectures that can be assembled quickly. That model allows developers to deliver capacity faster while preserving optionality as hardware requirements evolve.

The largest shift is at the campus level. Instead of optimizing individual buildings, hyperscalers are treating entire campuses as integrated systems. Sakalkar described this as a shift toward viewing the campus as a product, one that must balance flexibility, scale, and rapid deployment. That includes designing for multiple workload types, maintaining flexibility across hardware generations, and coordinating deployment across the supply chain, installation, and commissioning.

Unlike traditional phased buildouts, many AI campuses are now deployed in large increments, with infrastructure and compute coming online in tighter synchronization. The campus is the unit of design.

For AI builders, the implications are direct. The compute you can access is no longer determined solely by GPU availability. It is determined by power availability, cooling architecture, and campus-level integration. The data center is becoming a first-class constraint on AI research and deployment.

The industry is running out of room for incremental change. The traditional data center model is under strain. AI is not just increasing demand. It is changing the shape of that demand, introducing new constraints in power, cooling, and time.

James offered a direct message to a new generation entering the industry. “Question assumptions,” he said. “If something doesn’t make sense, it probably doesn’t.”

The question now is whether the industry can redesign fast enough to keep pace with AI capability growth. The data center is no longer a support system. It is part of the AI itself.