Designing Data Centers for 400kW GPU Racks | Researcher Conversations at GTC

Summary

Radiant presents itself as a newly launched neo-cloud with a global infrastructure roadmap spanning multiple sites and several hundred megawatts of capacity. Patrick O'Shaughnessy Legal says he has 12–15 years of HPC experience and joined roughly two years before launch to lead infrastructure planning and procurement. The company’s core strategy is to combine access to land, power, and capital—areas it says Brookfield helps unlock through the merger, including greenfield sites, planning permission, and behind-the-meter generation. On the technical side, Radiant is preparing for the latest Nvidia architectures and says it is already discussing 400 kW GPU racks for as soon as January next year. Its software platform, Radiant Cloud OS, is designed to turn bare metal into bare-metal services, VMs, or Kubernetes, while also providing observability, alerting, and workload reshaping for enterprise AI use cases.

Notes

• Radiant is positioned as a newly launched neo-cloud venture.
• Patrick says Radiant was announced only a couple of weeks earlier.
• He says he has 12–15 years of HPC experience.
• He joined Radiant about two years before launch to own infrastructure.
• His role is head of infrastructure, covering roadmap and procurement.
• Radiant says its strategy is organized around four pillars.
• Land and capital are explicitly part of the platform strategy.
• Radiant says its roadmap is global, not tied to one region.
• It says multiple sites are available and represent several hundred megawatts.
• Those sites are intended for the latest Nvidia architecture.
• The main constraints were land, power, and deployable capital.
• Brookfield is presented as solving those constraints.
• Brookfield gives access to greenfield sites with planning permission.
• It also brings power plants to supply the sites.
• Radiant says it can go from field to productive AI environment in about 12 months.
• The company frames this as accelerated infrastructure execution.
• Radiant says it is also partnering with Brookfield-linked data centers.
• The company wants to pre-design for future capacity coming online in 24–36 months.
• Radiant wants to align today’s designs with future Nvidia-driven infrastructure.
• The team says 400 kW GPU racks are already being discussed.
• That timeline is as soon as January next year.
• Data centers must be ready to accept those higher-density racks.
• Grid connection is one option, but he says it can be slower.
• Brookfield is said to enable behind-the-meter power generation.
• Radiant says energy sourcing is a central part of site planning.
• The software stack is described as comprehensive.
• Radiant Cloud OS is the lower-level platform.
• The lower-level platform turns bare metal into services.
• It supports bare-metal services, virtual machines, and Kubernetes environments.
• The platform is meant to re-provision infrastructure based on customer needs.
• Radiant says enterprise users may shift from training to inference.
• It specifically references a monolithic Slurm bare-metal training setup.
• The software stack includes AI application services above the lower platform.
• Radiant also includes data science, model registry, and inference services.
• Observability is a separate vertical in the stack.
• Telemetry is collected from servers, switches, equipment, racks, and the data center.
• That telemetry is turned into actionable alerts for operations.
• The goal is to maintain very high uptime and availability.
• Radiant frames uptime and availability as core enterprise requirements.
• The company argues that always-available service keeps customers happy.
• The interview emphasizes that capacity planning must anticipate future GPU generations.
• The roadmap appears to be built around rapid scaling rather than incremental expansion.
• Radiant’s pitch combines physical infrastructure control with software abstraction.
• The company’s differentiation is the ability to reshape infrastructure quickly for enterprise AI demand.
• Planning begins long before hardware arrival, using Brookfield relationships to line up capacity.
• The discussion suggests near-term pressure on datacenter power density.
• The operating model depends on high automation and telemetry-driven control.