The National Research Platform (NRP) is a distributed, NSF-supported research cyberinfrastructure that pools compute, storage, networking, and AI resources contributed by universities and research institutions into a single Kubernetes-based environment known as Nautilus. Rather than being a traditional centralized supercomputer, NRP operates as a federated platform spanning dozens of institutions.

Compute and GPU Access

NRP's compute infrastructure is built around the Nautilus HyperCluster, which currently comprises approximately:

• 500+ compute nodes
• 1,400+ GPUs
• 32,000+ CPU cores
• Petabyte-scale Ceph and S3 storage
• Resources distributed across 70+ institutions/sites on multiple continents.

Available accelerator hardware includes a wide range of GPU generations, such as:

• NVIDIA A100
• NVIDIA A10
• NVIDIA L40/L40S
• RTX 3090 / 4090 class GPUs
• Tesla-series and Quadro GPUs
• FPGA resources at selected sites

Access Model

NRP is generally:

• Free to U.S. nonprofit researchers and educators
• Accessible through institutional identities via CILogon
• Designed around Kubernetes namespaces rather than traditional HPC job schedulers.

Users can access resources through:

1. JupyterHub for notebooks and interactive computing.
2. Coder (browser-based VS Code).
3. Direct Kubernetes deployment using kubectl.
4. Containerized services, batch jobs, and AI workloads.

An important distinction from many NSF HPC systems is that users launch containers on Kubernetes rather than submitting jobs through Slurm-type queues. This makes NRP especially attractive for AI services, web applications, model serving, and long-running workflows.

NRP's Hosted LLM Service

One of the newer offerings is a managed LLM-as-a-Service platform. NRP hosts frontier open-weight models on its own GPU infrastructure and exposes them through a single OpenAI-compatible API endpoint. Researchers can use the models without provisioning or managing GPUs themselves.

Access Methods

NRP provides:

• OpenAI-compatible REST API
• Open WebUI chat interface
• LibreChat/Open WebUI browser access
• Support for coding tools such as:
  • Claude Code
  • Copilot CLI
  • Kimi CLI
  • OpenCode
  • Chatbox and other OpenAI-compatible clients.

Models Hosted

The model catalog changes over time, but examples currently include:

• Qwen3
• Kimi
• Gemma 4
• GLM-5
• MiniMax-M2
• GPT-OSS
• Embedding models such as Qwen3-Embedding. 

These include:

• Multimodal models (text + images)
• Large-context models (up to hundreds of thousands of tokens)
• Coding-focused models
• Reasoning-focused models
• Embedding services for RAG workflows.

Hosted vs. Self-Hosted AI

NRP effectively offers two AI usage modes:

Managed inference

• Call hosted models via API.
• No GPU allocation needed.
• Ideal for teaching, agents, RAG, and software development.

• Request GPU resources in your namespace.
• Deploy your own inference stack (vLLM, TGI, Ollama, etc.).
• Fine-tune models or run custom weights. 

Why This Is Significant

For universities without major AI infrastructure, NRP provides something unusual: access to both research compute and state-of-the-art hosted LLMs at no direct cost to academic users. Instead of purchasing cloud GPU time, researchers can use shared national resources for:

• Training and fine-tuning models
• Inference and agent workflows
• Classroom AI instruction
• RAG systems
• Scientific AI applications
• Large-scale Kubernetes-native services. 

For someone at a research university, the hosted LLM offering is arguably the fastest way to get started: you can use frontier open-weight models through an OpenAI-compatible API immediately, and only move to dedicated GPU allocations when you need custom training or inference infrastructure.

Comparison with Traditional HPC

This combination of a large national GPU pool plus centrally hosted open-weight LLMs is what makes NRP increasingly attractive for AI-focused research and education. [nrp.ai], [nrp.ai], [training.n...autilus.io]

The ACCESS network (Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support) is the NSF-funded successor to XSEDE. Rather than being a single computing system, ACCESS federates HPC, cloud, GPU, storage, and AI resources from many institutions and allows researchers to request allocations through a common process. Texas A&M University (TAMU) is one of ACCESS's major resource providers, contributing several advanced computing systems through its High Performance Research Computing (HPRC) center. 

Compute and GPU Access at Texas A&M Through ACCESS

Texas A&M contributes multiple systems, but the two most important ACCESS resources for AI and accelerated computing are:

• ACES (Accelerating Computing for Emerging Sciences)
• FASTER (Fostering Accelerated Scientific Transformations, Education, and Research)

Researchers obtain access through the ACCESS allocation process rather than a local TAMU account. Allocations are awarded in Service Units (SUs) and are available to U.S.-based academic researchers. 

ACES

ACES is a next-generation accelerator testbed specifically designed for emerging AI and scientific computing workloads. It combines traditional CPUs with a diverse collection of cutting-edge accelerators. 

Key hardware includes:

• NVIDIA H100 GPUs
• NVIDIA A30 GPUs
• Intel Data Center Max (Ponte Vecchio) GPUs
• Intel FPGAs
• Graphcore IPUs
• NEC Vector Engines
• NextSilicon accelerators
• Intel Sapphire Rapids and AMD EPYC processors 

Notable characteristics:

• ~110 nodes
• NVIDIA NDR200 interconnect
• SLURM scheduler
• Composable hardware infrastructure via Liqid fabric
• Designed specifically for GPU and accelerator experimentation
• Classified by ACCESS as a Category II advanced architecture resource 

For AI researchers, ACES is one of the most interesting systems in ACCESS because it provides access to accelerators that are often unavailable on traditional university clusters. 

FASTER

FASTER is an NSF MRI-funded composable supercomputer designed to support GPU-intensive data analysis and AI workflows. Unlike conventional HPC clusters where GPUs are physically attached to specific nodes, FASTER uses Liqid composable infrastructure to dynamically connect GPUs to computational workloads. 

Key features include:

• Approximately 184 Intel Ice Lake compute nodes
• NVIDIA A100 GPUs
• NVIDIA A10 GPUs
• NVIDIA A30 GPUs
• NVIDIA A40 GPUs
• NVIDIA T4 GPUs
• HDR InfiniBand networking
• NVMe-based storage architecture 

A major advantage is the ability to request large numbers of GPUs for a single workflow without being restricted by a node's fixed local GPU count. Researchers can configure CPU and GPU ratios more flexibly than on many traditional clusters. 

Additional TAMU Resources

Texas A&M also operates:

Grace

• TAMU's flagship supercomputer.
• Roughly 6 PFLOPS peak performance.
• Hundreds of CPU nodes plus approximately 100 NVIDIA A100 GPU nodes.
• More than 5 PB of Lustre storage.

Launch

• Another ACCESS-connected cluster used for production HPC workloads and workforce development activities.

AI and LLM Capability

ACCESS differs from NRP in an important way.

ACCESS Focus: Compute Allocation

The ACCESS program's primary model is:

• Allocate compute resources.
• Researchers run their own training and inference jobs.
• Users manage software stacks and model deployments.
• Systems are primarily accessed through Linux login nodes and SLURM batch scheduling. 

This resembles traditional HPC more than a cloud-hosted AI service.

AI-Oriented Services on ACES

The newest ACCESS documentation for ACES explicitly identifies:

• AI tools and support
• Model inference services
• Specialized AI hardware
• NAIRR Pilot participation
• GPU-centric research workloads 

This indicates that ACES is evolving beyond traditional HPC toward support for AI workflows and inference services. However, ACCESS does not currently provide the broad, centrally managed, OpenAI-compatible hosted-LLM catalog that NRP advertises. Researchers generally deploy or run models themselves on allocated resources. 

Why Researchers Use ACCESS at TAMU

For AI and computational science, TAMU's ACCESS systems are particularly attractive when researchers need:

• Large-scale GPU training
• Access to H100s and cutting-edge accelerators
• Experimental hardware architectures
• High-performance interconnects
• Large-memory scientific computing
• Custom model development and fine-tuning
• Benchmarking across different accelerator technologies 

In contrast, NRP's biggest advantage is that it offers a ready-to-use hosted LLM platform. ACCESS, especially through ACES and FASTER at TAMU, is stronger when researchers need direct control of the hardware and software stack for training, inference, benchmarking, or systems research.