Key Takeaways

  • Parallel Works and CoreWeave have deployed a fully managed AI and HPC environment for DARPA's NODES program, built on the ACTIVATE control plane and dedicated NVIDIA HGX H100 systems with NVIDIA Quantum InfiniBand networking.
  • Global AI-infrastructure spending reached $318 billion in full-year 2025 (IDC 2026), yet provisioning delays of weeks to months remain a persistent barrier for federally funded research programs, making operational orchestration as strategically important as raw GPU capacity.
  • HPC-AI cloud spending is projected to grow nearly 20% annually from 2025 through 2029, with biosciences explicitly identified as a high-priority vertical; the NODES deployment lands squarely in that growth corridor.
  • A 2025 cross-platform study found cloud HPC performance varies by workload, evidence that a managed orchestration layer, not just raw compute access, determines whether researchers can actually use the infrastructure they are allocated.
  • The arrangement illustrates a managed-service model gaining traction in federal science: a specialist operator absorbs identity, provisioning, scheduling, compliance, and support burdens so research teams can focus on the science.

The gap between available AI compute and usable AI compute has become one of the persistent friction points in government-sponsored scientific research. Agencies can, in principle, authorize access to large GPU clusters. But provisioning, identity management, allocation governance, and around-the-clock operational support have routinely added weeks or months before a researcher runs a first job. In multi-week simulation campaigns, where a single infrastructure failure can invalidate an entire run, that delay is a mission impact, not a scheduling inconvenience.

The numbers behind that friction exist against a backdrop of accelerating demand. Global AI-infrastructure spending reached $89.9 billion in Q4 2025 alone, up 62% year over year, with full-year 2025 spending totaling $318 billion (IDC 2026). The HPC, AI, and technical-computing market grew 23.5% in 2024, and Hyperion Research projects HPC-AI cloud spending to grow nearly 20% annually from 2025 through 2029, as reported by Scientific Computing World. More than 78% of HPC sites worldwide now use AI in some form (Hyperion Research, reported by Scientific Computing World 2026), confirming that machine-learning workloads have moved from experimental to operational status in technical computing.

For federal research programs, those growth rates make infrastructure readiness a strategic variable. A program that spends six months standing up compute for a two-year grant has already compromised its outcomes.

A Managed Layer for DARPA's NODES Program

On September 9, 2026, Parallel Works and CoreWeave announced the deployment of a fully managed AI and high-performance computing environment for DARPA's Network of Optimal Dynamic Energy Signatures (NODES) program, according to Parallel Works' press release. NODES addresses complex challenges in the dynamics of biological systems, research that requires large-scale, dependable computing to sustain simulation and modeling campaigns over extended periods.

The environment is built on CoreWeave's AI cloud platform and delivered through the ACTIVATE control plane, which handles user onboarding, identity and access management, provisioning, scheduling, allocation across research teams, and usage reporting, with around-the-clock direct-to-expert support for both the platform and researcher applications. The compute layer uses a dedicated reservation on NVIDIA HGX H100 systems, enterprise-scale storage, and high-speed NVIDIA Quantum InfiniBand networking.

"Breakthrough science shouldn't be delayed by infrastructure. DARPA needs researchers focused on biology, not on building infrastructure or managing compute environments. By combining CoreWeave's AI cloud platform with the ACTIVATE platform, we're giving NODES teams immediate access to the AI resources they need so they can spend their time advancing research." Matthew Shaxted, CEO, Parallel Works

The orchestration design addresses a common workflow split: ACTIVATE gives NODES researchers uniform access to both SUNK (CoreWeave's Slurm on Kubernetes offering) and CoreWeave Kubernetes Service (CKS) from a single platform. That means teams can run traditional HPC batch jobs and containerized AI workloads within the same underlying compute environment, a design choice that reflects where most research workflows now sit: somewhere between classical HPC schedulers and AI-native container pipelines.

Why Orchestration Outweighs Raw Compute Access

Raw GPU capacity has become easier to procure; the harder problem is governance. In a multi-team federal program like NODES, individual research groups require guaranteed allocations but should also be able to draw on shared capacity when it is available. ACTIVATE manages those sharing policies centrally while CoreWeave's observability stack monitors infrastructure health so problems are caught at the infrastructure layer before they interrupt a running job.

Independent research reinforces why this matters. Usability Evaluation of Cloud for HPC Applications, a 2025 IEEE Computer Society study, evaluated 11 HPC proxy applications across AWS, Microsoft Azure, Google Cloud, and Lawrence Livermore National Laboratory's on-premises clusters at scales reaching 28,672 CPUs and 256 GPUs. It found that cloud HPC performance is not uniform across workload types, a finding that points toward why an intelligent orchestration layer is increasingly as important as the underlying GPU count.

"DARPA takes on many of the country's toughest technical challenges. Our job is to give researchers the computing power to go after them. We've spent years building an AI cloud that the world's leading AI labs and AI-native companies depend on. With Parallel Works, we're bringing those same capabilities to the NODES program." Michael Intrator, co-founder, chairman and CEO, CoreWeave

Infrastructure Risk and the Federal Context

Government research deployments carry compliance, data-governance, and reproducibility requirements that make self-managed cloud operationally burdensome. The Hidden Infrastructure Behind AI Growth | GBAF underscores that power, cooling, networking, and water resources are non-trivial operational concerns for AI compute at scale, challenges that a managed provider absorbs rather than passing to the research team.

Context on the broader deployment landscape is also worth noting: on-premises HPC infrastructure still held 46.89% of the scientific-computing GPU market share in 2025 (Mordor Intelligence 2026), which means cloud is not categorically the right answer for every federal workload. What the NODES deployment argues instead is that a managed cloud model, dedicated capacity with a governance and orchestration layer operated by a specialist, can combine the provisioning speed of cloud with the operational predictability that sustained scientific work requires. Parallel Works spun out of Argonne National Laboratory in 2015 and operates production environments for federal defense and science programs, according to the company, a background that shapes how ACTIVATE is designed for these constraints.

A Note on Market Data Consistency

Independent IDC data for Q2 2025 appears in multiple research summaries and warrants clarification. Two figures are frequently cited together: cloud and shared environments representing 84.1% of global AI-infrastructure spending, and accelerated servers accounting for 91.8% of AI-server infrastructure spending (IDC 2025). These are not conflicting statistics, they measure different dimensions of the same dataset. The 84.1% describes the share of total AI-infrastructure spending flowing through cloud and shared environments; the 91.8% describes the share of AI-server spending specifically attributable to accelerator-equipped systems. Research sources that appear to present these figures in different orders are drawing on the same IDC Q2 2025 data from different angles, not reporting different findings.

Common Questions

How does ACTIVATE differ from a standard cloud HPC portal?

Standard cloud portals provide access to compute but leave provisioning, scheduling, allocation governance, identity management, and expert support to the customer. ACTIVATE operates as a full control plane that handles all of those functions centrally, including around-the-clock direct-to-expert support for both the platform and researcher applications. That distinction matters most in multi-team federal programs where individual research groups cannot absorb the operational burden of managing their own compute environments.

Why does this deployment use both Slurm and Kubernetes rather than one or the other?

Scientific computing workflows frequently split between traditional batch-scheduled HPC jobs (which Slurm handles efficiently) and containerized AI training or inference pipelines, which fit Kubernetes better. CoreWeave's SUNK offering runs Slurm on top of Kubernetes, enabling both job types to share the same underlying GPU pool without requiring researchers to maintain separate environments. ACTIVATE presents both through a single interface and manages the allocation policies that govern how each team uses that shared pool.

Is managed cloud HPC cost-effective for programs that could build their own infrastructure?

The direct procurement cost of managed cloud typically runs higher than self-managed on-premises hardware at scale. The relevant comparison includes provisioning timelines, staff time diverted to infrastructure management, the cost of delays to active research campaigns, and the overhead of maintaining 24/7 operational coverage. For time-sensitive federal programs with multi-team structures and complex scheduling requirements, those indirect costs frequently outweigh the premium on managed services, though each program's economics will depend on duration, scale, and existing in-house capability.

Looking Ahead

With biosciences identified by Hyperion Research as a high-priority vertical for HPC-AI cloud growth and HPC-AI cloud spending forecast to grow nearly 20% annually through 2029, the NODES deployment represents an early, visible instance of a model likely to scale. As more federal agencies confront the same infrastructure-readiness problem, abundant compute authorization, scarce operational capacity to deploy it, managed platforms that absorb the complexity of provisioning, governance, and support will attract further evaluation. Whether that model generalizes depends on how reproducibly it performs across multi-year research timelines, how it navigates evolving federal data and compliance requirements, and ultimately whether it consistently delivers what the NODES arrangement promises: researchers who spend their time on science rather than on infrastructure.