The system
Grace Blackwell GPUs and Vera CPUs, Liquid-Cooled
- GPU partition: 2,000 Dell nodes of Nvidia GB200 NVL4, two Blackwell GPUs and one Grace CPU each, about 4,000 GPUs in 28 racks.
- CPU partition: 4,752 Nvidia Vera nodes, about 836,000 cores, due between winter 2026 and early 2027.
- Performance: about 300 petaflops at FP64, roughly 10 times Frontera; 20 exaflops at BF16 and 80 at FP4.
- Storage and network: 400 PB of all-flash VAST Data storage at more than 10 TB/s, a Spectra Logic and Versity archive, and Nvidia Quantum-X800 InfiniBand at 800 Gb/s.
- Cooling: direct liquid cooling in Dell IR7000 racks. TACC has not published rack power; an independent systems wiki reports about 215 kW per GPU rack and over 100 kW per CPU rack.
- Status: phase one finished in July 2026 with GPU racks operating for early, internal users. NSF's $457M award covers the system, rent and operations; 10,000 to 15,000 researchers a year are expected.
The site
An NSF Supercomputer in a Commercial Colocation Building
TACC chose a Sabey Data Centers campus in Round Rock, Williamson County, in August 2024 instead of its J.J. Pickle Research Campus. Sabey replaced a former Sears Teleserv building with two high-density buildings; the Horizon data center is rated 15 to 20 MW. Distributed partners in NSF's Leadership-Class Computing Facility are NCSA, the Pittsburgh Supercomputing Center, the San Diego Supercomputer Center and the Atlanta University Center Consortium. See Horizon at Sabey Round Rock on AustinIO.
TACC's other systems
Frontera Retires as Horizon Arrives
- Frontera: Dell, about 8,000 Intel nodes, 23.5 petaflops, in service since 2019; its queues closed September 30, 2026, with data kept accessible for migration.
- Vista: Nvidia Grace Hopper GH200 and Grace CPU nodes, in production since September 2024 and a NAIRR resource.
- Stampede3: still running, offered as a CPU stopgap until Horizon's CPU partition arrives.
NAIRR
NAIRR Moves From Pilot to Permanent
The National AI Research Resource pilot, launched January 2024, has served about 900 research teams across all 50 states. In September 2026 NSF awarded $35M over five years to the San Diego Supercomputer Center, with TACC as partner, to run a NAIRR Operations Center with monthly allocations. Texas now holds both NSF flagship computing roles: leadership computing through Horizon and NAIRR operations through TACC.
Academic AI clusters
University AI Clusters Count GPUs in the Hundreds
| System | Scale | Status |
|---|---|---|
| DeltaAINCSA, Illinois | 608 H100-class (GH200) | Production since November 2024; NAIRR resource |
| NexusGeorgia Tech, Atlanta | More than 400 PF | Unveiled October 12, 2026; $20M NSF |
| TX-GAINMIT Lincoln Laboratory, Holyoke, MA | More than 600 GPUs; 2 AI exaflops | Online since summer 2025 |
| Kempner AI ClusterHarvard, Holyoke, MA | 1,144 GPUs incl. H200 and H100 | Operating |
| UT Center for Generative AIUT Austin and TACC | More than 1,000 GPUs | Doubled November 2025; part-funded by $20M from the Texas Legislature |
| Empire AI BetaUniversity at Buffalo, NY | 11 times Alpha's training capacity | Online August 2026; Gamma due end of 2027; $500M state program |
| GautschiPurdue, Indiana | 160 H100 | Dedicated October 2024 |
| Princeton AI clusterPrinceton, NJ | 300 H100 | Operating since 2024 |
| CalComputeUniversity of California | None yet | Framework due January 2027 under state law SB 53 |
UT Austin says it has more than 5,000 Nvidia GPUs in all, more than 4,000 of them Blackwell. No academic system publishes its power draw except Horizon's building rating.
Record
Dataset Record
- Dataset
- US academic and NSF AI systems: scale, funding and status
- Sources
- TACC, NSF, City of Round Rock, SDSC, NCSA, Georgia Tech, MIT, Harvard Crimson, Purdue, University at Buffalo, UT Austin, HPCwire, The Next Platform, DatacenterDynamics
- Last verified
- October 2026