Texas A&M Drives Drug Discovery Breakthroughs With NVIDIA DGX SuperPOD

Texas A&M University is one of the largest research universities in the United States, anchoring a system of seven institutions with more than $1 billion in annual research expenditures, operating at the frontier of discovery across basic science, engineering, agriculture, and medicine. But sustaining that position requires compute infrastructure that is difficult to scale.

Unlike industry, research universities must democratize access to world-class compute across dozens of independent labs and disciplines, without centralized IT budgets or procurement pipelines. For years, Texas A&M researchers worked within those constraints: writing grants to fund isolated clusters, waiting months for allocations on national facilities, or queuing jobs on aging hardware that couldn’t keep pace with modern AI and simulation workloads. Projects were scoped down or stretched across years. Researchers in computationally intensive fields—drug discovery, genomics, climate modeling—competed at a structural disadvantage against institutions that could provision GPU capacity on demand. 

The question wasn’t whether Texas A&M’s scientists needed more compute. It was how to deliver enterprise-grade infrastructure at an academic scale.

Dr. Reid T. Powell co-leads the Drug Discovery and Development Resource Center at the Texas A&M Institute of Bioscience and Technology TMC3-Collaborative Research Building Campus. This Cancer Prevention and Research Institute of Texas (CPRIT)-funded core facility supports between 20 and 40 comprehensive physical screening projects per year. His individual lab pursues drug candidates for cancer and Alzheimer’s disease, fields where computational screening can mean the difference between identifying a viable treatment candidate and missing it entirely.

Before VISION, Powell’s compute footprint consisted of seven workstations, the largest of which held three NVIDIA A6000 GPUs with 144 GB of VRAM combined. That ceiling forced a brutal tradeoff: Use fast, lower-precision models that generated more candidates, or use the high-precision co-folding with stronger enrichment of lead candidates but at a greatly more limited throughput. 

“If we could get 10% validation rates, we would be ecstatic,” Powell recalled. In practice, only 1%–10% of screened compounds would prove to be real binders from computation drug screens, meaning labs synthesized or ordered dozens of expensive molecules to salvage a handful of leads.

The compute constraint rippled outward. Powell could realistically pursue only one or two drug targets per year. Promising targets went unscreened. Grant proposals had to be scoped conservatively. And every run on the higher-precision models was time sensitive—supporting both the publication of research that advances science and the protection of promising innovations through patents. What the lab needed was not just more GPUs, but enough of them to run high-precision models at scale to meet the chemical diversity required to identify high-quality hits, with the potential to advance the science.

That answer came in the form of VISION—Texas A&M System’s NVIDIA DGX SuperPOD, ranked the number-one academic supercomputer on the June 2026 TOP500 list. Built on an NVIDIA DGX SuperPOD with nearly 760 NVIDIA Hopper GPUs delivering roughly a megawatt of compute power, VISION was purpose-built to serve the entire Texas A&M university system, from flagship research labs to historically underserved partner institutions. World Wide Technology (WWT), an NVIDIA Elite Partner, staged, configured, validated, and delivered the infrastructure and supported its deployment at Texas A&M.

“Early capacity modeling projects that Texas A&M System’s DGX SuperPOD could support 500–1,500 concurrent users per month across a diverse mix of workloads—from multimodal training runs using more than 250 GPUs to interactive debugging, notebook sessions using 5–10 GPUs, and large batches of small fractional-GPU jobs for activities like AI interference and user training sessions,” said Kim Andrews, executive director of IT research architecture at Texas A&M. DDN high-performance storage platform provides the data throughput needed to support these GPU-intensive workloads at scale. To support this diverse demand, Texas A&M uses NVIDIA Base Command™ Manager for cluster provisioning, lifecycle management, and infrastructure visibility, with NVIDIA Data Center GPU Manager (DCGM) telemetry providing real-time insight into GPU health and utilization. Integrated Slurm scheduling and a custom 11-tier quality of service framework help balance interactive workloads with extreme-scale jobs allocated up to 240 GPUs for as long as 72 hours. Dedicated Kubernetes nodes also support containerized inference workloads.

For Dr. Powell, onboarding was frictionless. He had pre-staged his Slurm job scripts knowing the system was coming online and used an agentic coding assistant to verify the configuration against the live cluster. Within 15 minutes of receiving access, he had passed a smoke test, launched a full-scale virtual screen, and immediately began benchmarking against his custom workflows, expecting to go even faster with additional machine- and driver-optimized containers now being provided by this NVIDIA partnership.

The contrast between before and after was significant. Powell’s first run on VISION completed 10.4 million molecular simulations in roughly one week—a workload he estimates would have cost over $1 million in rented cloud-based technology and taken substantially longer even if that cloud infrastructure could have matched the throughput.

More important than the speed was what VISION’s computing scale unlocked scientifically. Previous screens against one of Powell’s cancer targets had yielded roughly 120 candidate molecules, all predicted to bind to a single region. Running at VISION scale, Powell identified over 22,000 candidates spanning multiple binding regions and structural hypotheses. As a result, his triage yielded enough chemical diversity to optimize across multiple dimensions simultaneously: blood-brain barrier penetration, selectivity, potency, and manufacturability.

The effect downstream was transformative. 

Previously, Powell’s team might order more than 100 compounds to identify only a few viable leads. Now, computational screening helps narrow each order to 20 highly promising candidates, with the majority binding at or below the 10-micromolar threshold, which is the conventional target for hit discovery. That represents an 80%–90% hit rate, up from the lab’s previous 1%–10%.  

“Scale is allowing us to identify higher quality and diverse hits,” Powell explained. “This allows us to select molecules that are more drug-like, more manufacturable, and potent and selective out-of-the-box rather than just finding molecules that meet one parameter of what a potential therapeutic needs to be.”

At the system level, VISION’s impact extends well beyond a single lab. The cluster runs at 95%–98% GPU utilization, with 26 active research projects from the Texas A&M main campus and a beta cohort of 118 additional projects across six partner institutions waiting to be provisioned. The results made the case for something bigger.

As VISION transitions from early adopters to general admission, the team is preparing to onboard a combined 144 projects and nearly 500 accounts across the Texas A&M system—a process that will be managed carefully within US export control and licensing requirements. Prairie View A&M University is the largest single institution in the incoming cohort, representing a significant expansion of research capability at a Historically Black College or University (HBCU).

For Powell’s lab specifically, the compute bottleneck no longer sets the pace. He is writing grants proposing five to ten drug targets per year. He plans to extend VISION to phenotypic screening pipelines, generative peptide design using NVIDIA NIM™ microservices, and reinforcement-learning-based lead optimization. At the institutional level, VISION will serve as the backbone for AI training and education across Texas A&M’s medical and research programs, with plans underway for digital teaching assistants and large-scale coursework on the cluster. As the university’s computing needs grow, VISION is establishing a replicable model for how universities can offer enterprise-grade GPU infrastructure as a shared resource, making frontier AI research accessible to every faculty member, student, and institution in the network.