Texas Center Triples AI Inference System
The Texas Advanced Computing Center at the University of Texas at Austin is tripling the size of its SambaStack AI inference system with SambaNova hardware, the center announced on Monday, November 16. A $5 million National Science Foundation award funds the expansion, with $3.5 million going to AI accelerator hardware.
The money isn't buying training capacity, the expensive headline side of AI. It's buying inference, the part researchers actually wait on when they run models. Tripling that queue capacity is a plain, measurable upgrade for the scientists who share the system.
The Texas Advanced Computing Center (TACC) at the University of Texas at Austin is tripling the size of its SambaStack AI inference system, the center and SambaNova announced on Monday, November 16. The expansion adds SambaNova rack systems to Sagebrush, the system TACC uses to provide AI inference services to researchers. A National Science Foundation award of $5 million funds the project, with $3.5 million of it going to AI accelerator hardware.
The distinction that matters here is inference versus training. Training builds a model, once, at enormous cost. Inference runs the finished model, every time a researcher asks it a question, tests a variant or processes a dataset. A shared research center lives and dies on inference throughput, because that's the queue its users stand in. Tripling the inference system means three times the capacity for the work researchers submit daily, not a bigger trophy machine.
TACC first chose SambaNova technology in 2024 and is keeping its existing supercomputing systems in place while the new hardware is added. The center operates some of the most powerful academic supercomputers in the United States and provides computing services to researchers well beyond its own campus. The expansion connects to the National Artificial Intelligence Research Resource, the federal program meant to give academic researchers access to AI computing they can't buy on their own.
The funding route matters as much as the hardware. The $5 million comes from the National Science Foundation, the federal agency that bankrolls much of the country's academic computing, and $3.5 million of it is earmarked for the accelerator hardware itself rather than operations. TACC is also involved in the National Artificial Intelligence Research Resource's Super Intelligence Infrastructure for Science initiative, the federal effort to build shared AI infrastructure for researchers who don't sit inside a large technology company. An inference expansion at a center with that role doesn't serve one campus. It serves the national queue of academic projects that apply for time on shared systems, which is why a tripling, rather than a marginal upgrade, is the detail that counts.
SambaNova designs its own AI chips alongside the SambaStack platform that runs models on them, and the expansion is built on the company's SN50 accelerator, introduced earlier in 2026. SambaNova claims the SN50 runs five times faster than rival chips. That's the company's own figure: no independent verification of the speed claim or of the deployment timeline has been reported, and chip-vendor comparisons depend heavily on which model and workload get measured. The verifiable part of this announcement is the hardware purchase and the capacity increase, not the multiplier.
What researchers get is concrete. The added racks go into service for the inference workloads TACC already runs for its users, on a system those users already access. For a field where academic labs routinely wait for commercial cloud credits or shared cluster time, a federally funded tripling of inference capacity at a national center is the kind of infrastructure spending that shows up later as published results, not press releases.
The conversation.
Loading comments…
A little more perspective.
Reading is free. Verified members can comment and reply for $2 a month.