Across federal agencies, defense organizations, and national research institutions, artificial intelligence initiatives are accelerating rapidly. Agencies are investing in GPUs, modern analytics platforms, cloud infrastructure, and advanced AI models to improve decision-making, operational awareness, cybersecurity, logistics, and scientific research. But many organizations are running into the same problem: the infrastructure underneath their data was never designed for AI.
Federal agencies already possess enormous volumes of valuable information, most of it unstructured. Satellite imagery, ISR feeds, geospatial data, simulation outputs, cyber telemetry, research datasets, video, sensor data, and operational records are being generated at an unprecedented scale. In fact, most agencies are overwhelmed by the amount of unstructured data they already manage.
So, it isn’t about a lack of data; the real issue is data fragmentation.
Federal AI Won’t Scale Until Agencies Modernize the Data Layer
This fragmented, unstructured data is stored across NAS systems, regional data centers, object stores, cloud environments, edge locations, and classified and unclassified domains that evolved independently over decades. The result is data copy sprawl. Teams end up copying datasets between storage systems, cloud environments, and compute clusters simply to make data accessible to applications. Data is extracted, reformatted, staged, duplicated, and migrated repeatedly across environments, introducing operational complexity and risk, governance challenges, and impacting time-to-value.
Modern AI and HPC
Modern AI and HPC workloads are fundamentally different from the enterprise workloads that traditional storage infrastructure was built to support. AI pipelines demand high-throughput, low-latency parallel access to massive datasets. Training environments continuously ingest, process, checkpoint, retrain, and validate data at scale. Inference systems increasingly require real-time access to datasets that don’t reside “next to” the GPUs.
Agencies often discover that the bottleneck in AI infrastructure isn’t the GPUs, but the inability to efficiently access and operationalize data across fragmented environments.
This becomes especially apparent when AI projects move from pilot programs into production-scale deployments. Small proof-of-concept environments can often work around infrastructure limitations using isolated datasets and temporary workflows, but production environments cannot.
For federal agencies, these challenges are compounded by security and operational requirements that commercial enterprises don’t often face at the same scale.
Data sovereignty, classification boundaries, zero-trust architectures, and strict access control requirements all influence how data can be stored, accessed, and moved. Many agencies operate in multi-domain environments where data must be consistently governed across both classified and unclassified systems. Edge environments may have intermittent connectivity and limited local infrastructure. Mission teams require rapid access to data while simultaneously maintaining strict governance and auditing requirements.
In many cases, security policies themselves become operationally difficult to enforce because governance is fragmented across dozens of independent storage systems and administrative silos. This is why many agencies are beginning to rethink the data layer itself.
Rethinking the Data Layer
A key attribute of this data layer is a global namespace that provides a single logical view of data across distributed environments. Instead of forcing organizations to manually move data into a newer, faster storage system before it can be used, data can remain in place while remaining accessible to authorized users and applications across environments. This changes the operational model significantly.
Rather than constantly moving data to applications, organizations can intelligently position data where it is needed while maintaining centralized visibility and governance. Data generated at the edge can be accessed by AI systems in regional data centers or cloud environments without creating unnecessary copies. Research organizations can collaborate across geographically distributed environments without introducing data copy sprawl. And agencies can modernize infrastructure incrementally while continuing to leverage existing storage investments.
Once all this fragmented data is unified, the next barrier is getting it to the compute resources and applications that need it. This is difficult because the datasets are massive and hard to move. This has sometimes been referred to as “data gravity.” Historically, data gravity referred to the idea that applications and compute needed to move closer to where data resided. But AI has flipped that model: GPUs have more “gravity” than the data, because they have massive energy demands, and energy has become a major constraint. It is no longer about bringing compute to a massive storage system; instead, it is about bringing your data to the compute. And those compute resources likely reside in another location, another cloud region, and perhaps even another cloud provider.
And that is where data orchestration comes in. Data orchestration automates data placement and management throughout its lifecycle, so you can intelligently tier data between flash and disk and bring data to the compute without creating data copy sprawl.
Federal agencies are entering a period where AI success will increasingly depend on data infrastructure agility. The organizations that succeed will be those that can operationalize data efficiently across complex, distributed, and highly secure environments. Modernizing the data layer will increasingly determine how effectively agencies can execute their missions.
The author, Eric Bassier, is Senior Director of Product Marketing at Hammerspace.