Project directory / United States
Frontera Supercomputer, TACC
Retired October 2026; historical research case
Frontera entered production at TACC in 2019 and closed its queues in October 2026. Its acceptance study documents application performance above earlier projections; this is historical evidence, not current operating AI capacity.
Published and evidence reviewed November 9, 2026.
Compare this project · Evidence checkpoints · FJ Intelligence
Investment takeaway
Frontera shows how computing projects can exceed application-level projections—and why the explanation matters. Higher delivered processor frequency, networking, storage and changed software contributed to the differences. This is a lesson in procurement and useful work, rather than a current commercial AI-campus investment.
Project at a glance
- Location
- Austin, Texas, United States
- Status
- Retired October 2026; historical research case
- Capacity scope
- Historical 23.5-petaflop HPL result; application acceptance tests separately compared projected runtimes with measured results.
- Named participants
- TACC / University of Texas at Austin; NSF; Dell EMC; Intel; Mellanox; DDN
- Map precision
- Approximate city/locality marker, not a surveyed facility, building or parcel.
FJI source-qualified: four-document gate met, including project-specific government evidence. Qualification evidence and decisions →
A completed operating life
NSF identified Frontera as a leadership-class computing resource before deployment. TACC records production from September 2019 and queue closure in October 2026. The system has been decommissioned; its historical performance is not included here as current service capacity. [1, 2]
Projected work versus measured work
TACC’s acceptance paper reports projected-to-measured runtime ratios. At unchanged projected node counts, ratios were 1.03 for AWP, 1.22 for CACTUS, 1.64 for MILC, 1.03 for NAMD and 1.15 for Caffe. For fixed work these imply processing rates about 3%, 22%, 64%, 3% and 15% above projection. They are application tests, not annual production totals. [3]
Why the tests beat projections
The paper identifies stronger-than-expected delivered hardware and workload-specific changes. It also documents storage and application-code effects. The largest differences cannot be assigned to one supplier or generalized to every workload. The primary system combined Dell EMC servers, Intel processors and Mellanox networking; the wider installation involved other partners. [3, 4]
Financial interpretation
Frontera was publicly funded research infrastructure. An application finishing earlier can free resources for other science, but that is not a disclosed cash return. Supplier revenues, margins, support terms and the economic value of additional research cannot be inferred from the runtime ratios.
Readiness boundary
The retired status applies to Frontera. Repurposed racks or successor systems require their own evidence. A historic TOP500 entry and an accessible project webpage do not establish that the machine remains in service. [2, 5]
What remains unverified
No current production capacity is attributed. Historical application tests do not establish annual output or a supplier-specific causal advantage; project cash returns are not disclosed.
Next evidence checkpoint
Preserve historical acceptance evidence and track published successor research separately; do not count retired hardware as current operating capacity.
Related project research
Sources and claim boundaries
- National Science Foundation
Predeployment Frontera allocation plans, TACC location and DellEMC / Intel system identity.
Frontera section · Document date: 2018-12-12 · Read November 9, 2026.
- TACC
Production timeline, October 2026 queue closure and decommissioned status.
System overview · Document date: Not established · Read November 9, 2026.
- Richard Todd Evans / TACC
Projected versus measured runtimes and node counts; explanations of deviations.
Tables 5–6; acceptance results discussion · Document date: 2020-11-13 · Read November 9, 2026.
- University of Texas at Austin
System launch, 23.5-petaflop HPL result and technology partners.
Article body · Document date: 2019-09-03 · Read November 9, 2026.
- TOP500
Independent predeployment account of the planned system; historical expectations only.
Article body · Document date: 2018-08-29 · Read November 9, 2026.
AI-assisted public-source research by Fjord Journal. No site visit, private contract review or independent metering. Institutional, supplier and benchmark claims retain their stated scope.
Company investor information & quarterly statements
Financial-disclosure review remains incomplete for one or more participants. The company-level limitations below remain unresolved; this section does not establish full corporate-disclosure verification.
Corporate disclosures provide company-wide context. They do not establish this project’s revenue, spending, profit, ownership share or operating capacity. Disclosure verification records →
