
Share
Instead of another batch-scheduled supercomputer, TACC is betting that researchers need always-on inference and agentic tooling. Sagebrush runs on SambaNova hardware and ties into NAIRR, with production targeted a few years out.
Most HPC resources still work like a queue: you submit a job, wait your turn, get your allocation window, and move on. That model works fine for a big simulation run, but it's a bad fit for the way people actually use AI tools now, where you want a model sitting there ready to answer a query or run an agent loop at 2am without waiting in line. TACC's answer to that mismatch is Sagebrush, a new always-on AI inference service funded by an NSF award (2537075) that's meant to give the national research community persistent access to models and agentic workflows, rather than one more batch-scheduled cluster.
The project runs alongside the NSF Leadership-Class Computing Facility (LCCF) and is explicitly designed as infrastructure, not a one-off experiment. Katie Antypas, senior science and engineering advisor for NSF's CISE directorate, framed it as expanding the connective tissue between LCCF and the National AI Research Resource (NAIRR): "Sagebrush is a significant NSF investment that provides researchers a new, persistent way to put open models and agentic tools to work on their most demanding scientific questions. Paired with the NSF LCCF and connected to NAIRR, it expands the national research infrastructure needed to translate emerging capabilities into new scientific discoveries."
That NAIRR connection matters for context. TACC is working with the San Diego Supercomputer Center, which leads the NSF-funded NAIRR Operations Center, to help move NAIRR out of its pilot phase and into something closer to permanent infrastructure. Sagebrush is effectively TACC's contribution to that broader effort, which NSF has been calling "Super Intelligence Infrastructure for Science."
The service breaks down into a few concrete pieces rather than a vague "AI platform" pitch:
Niall Gaffney of TACC's Data and AI Directorate described the goal in fairly practical terms: "At its core, we want to provide AI inference services that help people with their research, and support agentic workflows that are more targeted to specific domain areas." He also flagged a gap TACC is trying to fill: "We want to host the computational tools researchers have grown to need during this fast-evolving AI environment that were previously unsupported, specifically tuned to specific domain needs."

On the hardware side, Sagebrush runs primarily on SambaNova's AI accelerators, specifically the SambaStack platform, which pairs proprietary inference processors with its own compiler stack. TACC has had early, non-production access to this hardware already, and is now purchasing four additional systems to expand capacity for commonly used open models, with delivery expected this November. Abhi Ingle, chief customer and solutions officer at SambaNova, pitched it as a natural extension of that early-access relationship: "Science increasingly runs on AI inference, for everything from coding to simulation. Sagebrush brings that capability to the open research community at national scale, and we're proud that TACC is expanding on SambaNova after its early-access work."
Worth noting: this isn't an Nvidia-GPU-everywhere story. SambaNova's architecture is a dataflow design rather than a traditional GPU, and the bet here is that it's well suited to steady-state inference serving rather than training workloads. Pairing custom inference silicon with an open-source software layer, agentic frameworks and vector stores for RAG, is TACC's attempt to keep the stack open and portable even while the underlying hardware is proprietary.
The timeline is deliberately staged. The NSF grant kicked off September 15, 2026, with the first two years structured as an experimental phase. During that window, TACC will build out support structures, figure out what researchers actually need from an always-on service (as opposed to what they say they need), and explore whether parts of the service could become self-funded. That phase could accelerate if things go well. The following three years shift into a production phase, which is presumably where allocation policy and capacity planning get locked down.
TACC isn't doing this in isolation either. Los Alamos National Laboratory is separately exploring SambaNova-driven AI inference for its own research needs, and the two organizations plan to share information as both efforts mature. That's a useful signal: if two different NSF/DOE-adjacent institutions are independently landing on SambaNova for inference infrastructure, it suggests the dataflow architecture is finding a real niche in research computing, not just a single vendor relationship.
Over the next six to twelve months, TACC says it will build a new allocation model designed to scale across multiple models and services, integrate Sagebrush into its existing allocation systems (the same ones that govern access to LCCF and other TACC resources), and migrate older hardware into user-facing services as the four new SambaNova systems come online.
The practical upshot for researchers: if you've been running your own inference endpoints or patching together RAG pipelines on borrowed compute, TACC wants to hear from you. They're explicitly looking for early users with specific use cases to shape how the allocation model and service tiers get built out. Gaffney put it plainly: "We look forward to working with research communities to understand their challenges and leverage AI technologies in their workloads to help bridge them. We will continue to evolve the system to support future capabilities over the lifetime of Sagebrush." For a field that's spent the last few years bolting AI tooling onto HPC systems that weren't designed for it, a purpose-built always-on service is a meaningfully different approach, and one worth watching as it moves from experimental to production.
Tags
Original Sources
TACC Launches Sagebrush, an Always-On AI Inference Service for US Open Science Community - HPCwire
↗ https://www.hpcwire.com/off-the-wire/tacc-launches-sagebrush-an-always-on-ai-inference-service-for-us-open-science-community
About the author
Kai built ML infrastructure at a Bay Area startup before developing an obsession with transformer architectures and inference optimisation that eventually pulled him out of product work entirely. A stint at a compute research lab sharpened his instinct for what actually matters in a model release versus what is marketing. He writes from the inside — from the perspective of someone who has debugged the systems he is describing at three in the morning. He is allergic to hype and instinctively drawn to the unglamorous plumbing questions that everyone else skips over.
More from The Engineer →This Week's Edition
20 November 2026
29 articles
Related Articles

The Data Center Backlash Has Gone From Local Fights to National Policy
Tools & Engineering · 5 min

Google Built a Fake Internet to Train Chrome's AI Agents Without Breaking the Real One
Tools & Engineering · 5 min

National Academies Report Calls for Faster, Tougher Retraction Standards
Models & Research · 5 min
Related Articles

The Data Center Backlash Has Gone From Local Fights to National Policy
Tools & Engineering · 5 min

Google Built a Fake Internet to Train Chrome's AI Agents Without Breaking the Real One
Tools & Engineering · 5 min

National Academies Report Calls for Faster, Tougher Retraction Standards
Models & Research · 5 min
More Stories
© 2026 Cedar & Bloom. All rights reserved.