TACC's director puts the NSF supercomputer's allocation split on the record, and says rising Blackwell prices held UT's state-funded add-on to 432 GPUs.
Ninety percent of the available time on Horizon, the National Science Foundation's new leadership-class supercomputer at the Texas Advanced Computing Center, goes to NSF users, TACC executive director Dan Stanzione told Supercomputing News (SCN) in a September 14 email. The University of Texas at Austin's own block of new Blackwell GPUs is a separate, state-funded purchase of 432 that sits outside Horizon's base configuration of 4,000, and Stanzione tied that count to cost. "As pricing has continued to increase, it's also not 1,000," he wrote.
Both purchases are NVIDIA Blackwells, and on the two accounts now on the record, they were priced years apart. TACC agreed with NVIDIA on a fixed price for the NSF system's 4,000 GPUs five years ago, Stanzione told The Next Platform in February; the publication, citing him, put the Horizon deal with NVIDIA and Dell in 2021. NVIDIA has held to that price, he said, "despite what has happened to the retail price of the GPUs," while costs for Dell and the rest of the system, which had no such terms, rose sharply. The 432 are funded from a fiscal 2026 state appropriation, a count Stanzione links to rising prices.
The record supports that contrast and stops short of a per-GPU comparison. TACC has not disclosed the fixed price; the state money also paid for networking and data-center work, and neither of UT's announcements ties a Blackwell count to the appropriation. Read together, the two accounts put the NSF partition's price before generative-AI demand moved the GPU market and the state-funded purchase after it.
The email also covered where the rollout stands. About 15 research teams are on Horizon now, with roughly 40 expected in October, and TACC plans to publish a Linpack result using emulated 64-bit arithmetic that, by his account, the TOP500 list is not accepting. Horizon is the first system of NSF's $457 million Leadership-Class Computing Facility award to UT Austin.
NSF users reach their share through two routes, Stanzione wrote. One is LRAC, the Leadership Resource Allocation, the peer-reviewed track on the LCCF allocations page. The other is the National AI Research Resource pilot, NAIRR, which offers Horizon time through a Deep Partnership call of 200,000 to one million GPU-hours. The remaining 10% is for UT and discretionary use.
As Stanzione defines it, Horizon's service unit, or SU, is one hour on a single GB200 board, which carries one Grace CPU and two Blackwell GPUs, so one SU equals two GPU-hours. With 4,000 GPUs on the system, he put the available pool at about 15 million SU a year.
By SCN's arithmetic, 2,000 boards running every hour of a year would produce 17.5 million SU, so the available pool is about 86% of the wall-clock ceiling; Stanzione did not itemize the difference. Ninety percent of 15 million is 13.5 million SU, or roughly 27 million Blackwell GPU-hours a year for NSF users. The UT and discretionary share comes to about 3 million GPU-hours.
His numbers also explain a sentence on the LCCF page, which caps early-operations LRAC requests at 500,000 SU for six months and describes that cap as nearly 5% of the full system's annual capacity. Against a 15 million SU year, it is 3.3%. Stanzione wrote that the 5% refers to the early-user period, which runs for less than a year.
A maximum award of 500,000 SU is a million GPU-hours. For scale, that is less than a sixth of the more than 6 million GPU-hours the Apertus team reported for the production run of its largest open model, at 70 billion parameters, on the older GH200 superchips of Switzerland's Alps system.
For Frontera, the machine Horizon replaces, TACC's system page puts the share of available hours allocated through NSF's process at up to 80%. Stanzione gave no track-by-track breakdown for Horizon's 90%, and his May 18 note to users says a full call across all tracks for 2027 is still to come.
UT described its Center for Generative AI expansion twice last November. UT News said on November 10, 2025, that the center "is doubling its computing capacity to more than 1,000 advanced graphic processing units," and that the Texas Legislature had appropriated $20 million toward a portion of the center's additional GPUs.
A week later, a second release announced more than 4,000 NVIDIA Blackwell GPUs that would go into Horizon and said, "More than 1,000 of the advanced GPUs will be dedicated to UT's Center for Generative AI," with that block reserved for UT faculty and student researchers.
The November 10 wording can describe the center's total capacity. Its 2024 cluster was specified at 600 NVIDIA H100s, and in January 2026 the Institute for Foundations of Machine Learning announced a GB200 NVL72 system of 72 Blackwell GPUs housed at TACC. Add 432, and the total is 1,104. The November 17 sentence sits in a release about Blackwells and reads as a count of new ones, which is how the Austin American-Statesman read it the next day, writing that more than 1,000 of the newly acquired GPUs would be set aside for UT faculty and student researchers. Neither UT release says whether the center's GPUs would be drawn from the NSF partition or added beside it.
Stanzione's email answers that. The center's GPUs are additional hardware outside Horizon's base configuration and are not among the 4,000 for NSF, he wrote, and the addition brings the total to 432 Blackwells, most of which are not yet available.
UT Austin's University Media Relations said on September 18 that it was working on a response to questions SCN sent on September 14 about the GPU count and the state appropriation. SCN gave UT until the end of the day on September 22 and followed up that morning; no answer had arrived by publication.
The 2026-27 General Appropriations Act appropriates $40 million in state general revenue for fiscal 2026 to support TACC, with a legislative finding that the center needs upgrades. The rider does not mention GPUs or the Center for Generative AI. The rider's text directs the full $40 million to TACC. By Stanzione's account, that $40 million was the entire UT rider and half of it went to TACC, where most of the money is paying for the 432 GPUs and the networking and data-center infrastructure around them. Of the other $20 million, he wrote, "I'm not aware of what UT did with the rest." SCN's questions to UT included what the balance funded.
The Horizon user guide, last updated August 12, still carries a July 24 note restricting the system to internal users. Stanzione described the current phase as "Very early user" (emphasis his), with about 15 teams on the system while TACC finishes benchmarking and tuning. He expects to expand to about 40 teams in October and add more slowly after that, and said the formal early-user period will probably start sometime in October. TACC had a site review with NSF the week before he wrote.
Frontera's retirement date has moved. A TACC notice dated September 16 sets October 15 as the last day of job submissions, two weeks past the September 30 target in Stanzione's May note. If the formal early-user period opens in October, GPU users get an overlap.
CPU-bound teams get no such overlap. TACC's July 7 release, which described the GPU racks as operational, expects Horizon's Vera CPU partition in winter 2026 or early 2027, and the May note acknowledged a CPU gap once Frontera shuts down. In the interim, it points users to CPU nodes on Vista, to Stampede3 through ACCESS, and to Lonestar6.
Stanzione told The Next Platform in February that TACC would produce Linpack results in both native and emulated FP64. On September 14, he wrote that TACC has run both and is now tuning NVIDIA's supported implementation of Ozaki-2, and that it will publish all its HPL numbers once tuning is done. "The Top 500 list is only accepting native and MXP (which we will submit), but I will publish the Ozaki2 outputs (with the residuals) just so the community can see them," he wrote.
The Ozaki scheme rebuilds a 64-bit matrix multiply from many low-precision products that run on a GPU's AI-oriented arithmetic units. Its second version does so with integer modular arithmetic. The TOP500's Linpack rules bar submissions that emulate floating-point arithmetic with integer operations in software, which is consistent with Stanzione's description of what the list accepts. HPL-MxP, the mixed-precision benchmark, is ranked separately. Residuals are Linpack's accuracy check, so publishing them lets readers judge how close the emulated answer is as well as how fast it came.
SCN has followed this argument from the field's adaptation of FP64 science to AI silicon through the joint Dongarra, Hoefler, and Matsuoka paper. The abstract of Satoshi Matsuoka's Ozaki 2.5 paper presents every result as a model projection that has yet to be measured, as SCN reported on September 16. A tuned Ozaki-2 run across about 4,000 Blackwells would be a system-scale measurement.
Both October milestones, the formal early-user period and the move to about 40 teams, are Stanzione's estimates, and the Linpack numbers follow whenever tuning ends.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
