Scalable AI infrastructure: Lessons from Los Alamos National Laboratory

CIOs across industries face a common bottleneck: data pipelines and compute architectures designed for traditional analytics cannot scale to handle large-scale artificial intelligence. As organizations accelerate their deployment of large-scale models across core corporate divisions, many data centers are straining under massive power requirements and complex multi-node orchestration.

To overcome these constraints, leaders can look at the advanced computing initiatives at Los Alamos National Laboratory (LANL). Task-driven environments like LANL handle massive, high-consequence data matrices. By co-designing next-generation infrastructure architectures to run sophisticated AI workloads, the laboratory offers an example for building scalable, resilient systems capable of accelerating complex domain-specific workflows.

The core challenge: Architecture and power bottlenecks

AI projects frequently stall during the scaling phase. The primary infrastructure hurdles include:

• Data silos and throughput bottlenecks: Training large language models (LLMs) or multi-modal systems requires immense parallel processing. Traditional network topologies create localized data chokepoints, leaving high-end graphics processing units (GPUs) underutilized while waiting for data ingestion.
• The power density gap: Standard enterprise data centers typically support 5 to 15 kilowatts (kW) per rack. High-density AI infrastructure requires up to 40 to 100 kW per rack, necessitating extensive upgrades to liquid cooling systems and electrical distribution.
• Workflow orchestration complexity: Managing distributed training workloads across thousands of computing nodes demands tight integration between hardware, scheduling software, and datasets to prevent cascading failures.

For LANL, these challenges are magnified. The laboratory requires advanced computing platforms capable of running highly secure, automated simulations without relying on commercial cloud infrastructure. IT leaders face a parallel demand: the need to execute proprietary models on-premises to protect corporate intellectual property and adhere to strict regulatory frameworks.

The solution: Co-designed AI-enabled supercomputing

To address these infrastructure bottlenecks, LANL leveraged a co-design methodology, collaborating with HPE and NVIDIA to develop specialized, high-density computing environments. Rather than assembling disparate hardware components, the laboratory treated the entire data center ecosystem as a single integrated platform.

This integrated approach is demonstrated through three main strategic pillars aligned with the U.S. Department of Energy’s overarching Genesis Mission.

1. Advanced computing platforms and unique partnerships

The foundational layer relies on tightly integrated CPU and GPU architectures, such as the Venado Supercomputer. Venado utilizes NVIDIA Grace Hopper Superchips, which fuse an ARM-based CPU and a Hopper GPU onto a single module. This design eliminates traditional PCIe bus bottlenecks, expanding coherent memory bandwidth to allow faster data movement during large model training.

2. Advancing scientific AI and agentic workflows

Looking toward future scalability, LANL is developing next-generation AI-optimized systems called Mission and Vision. Slated for deployment through 2027 and 2028, these platforms are co-designed using HPE Cray Supercomputing alongside advanced NVIDIA Vera CPUs and Rubin GPUs. These systems are engineered specifically to run agentic AI workflows—leveraging custom automated software setups, such as LANL’s ArtIMis ecosystem, where intelligent agents execute multiple tasks in parallel to accelerate discovery science.

3. Transforming scientific computing through decentralized infrastructure

To counter regional grid limitations and meet high-density power demands, LANL expanded its compute footprint outside its primary facility. This includes a new $1.25 billion research complex built in collaboration with the University of Michigan. The facility will sit on a 144-acre site and draw approximately 100 to 110 megawatts of power to house centers dedicated to critical national security AI challenges and academic collaboration. For grid-independent reliability, LANL collaborated with NVIDIA and advanced fission developers to study deploying localized sodium-fast nuclear reactors to directly power dedicated computing clusters.

Quantifiable benefits

The architectural principles validated by LANL translate directly into tangible operational metrics for enterprise IT landscapes:

• Maximum compute efficiency: By adopting unified memory architectures, organizations can maximize GPU utilization rates. Eliminating data transfer delays between the system processor and the accelerator ensures that high-value hardware investments spend less time idle, accelerating time-to-solution for complex production models.
• Data protection and control: By deploying high-density infrastructure on-premises, organizations can execute frontier reasoning models entirely within a controlled network perimeter. This architecture allows organizations to extract deep insights from proprietary datasets without exposing sensitive financial, medical, or corporate data to external public networks—mirroring how LANL deploys OpenAI’s reasoning models on a classified network to conduct national security research.
• Predictable operational scaling: Co-designed architectures simplify the infrastructure lifecycle. Standardizing onintegrated infrastructure like HPE Cray supercomputers with NVIDIA acceleration allows IT departments to scale compute capacity predictably as data pools expand.

For CIOs, the takeaway from Los Alamos National Laboratory is clear: successful artificial intelligence requires a shift from general-purpose computing to an integrated advanced computing model. Organizations must treat compute, networking, storage, software, and power as an interdependent stack.

Looking forward, this architectural approach serves as the literal foundation for executing complex, automated operations. Under the U.S. Department of Energy’s initiative, LANL was recently awarded funding for seven dedicated projects designed to pioneer transformative scientific workflows. These initiatives—ranging from closed-loop autonomous frameworks for nuclear fuel qualification to evolutionary agentic systems for hardware co-design—illustrate the practical end-state of high-density AI infrastructure: transitioning from raw processing power into self-sustaining, intelligent software workflows that accelerate the speed of discovery.

By collaborating with established technology partners like HPE and NVIDIA, teams can deploy infrastructure tailored to their workload demands. This strategic approach mitigates data chokepoints, manages power density constraints, and secures core data assets—positioning the organization to successfully scale AI capabilities into core operational drivers. For more information, visit hpe.com/cray and hpe.com/ai. 

***********

As AI becomes increasingly central to economic competitiveness, scientific advancement, and national priorities, organizations require infrastructure that balances performance with security and sovereign control. Together, HPE and NVIDIA co-engineer rack-scale AI systems that integrate AI computing, high-performance networking, and supercomputing expertise to support large-scale AI workloads. This provides enterprises, governments, and research institutions with a trusted foundation for sovereign AI initiatives while maintaining control over critical data, models, and operations.