A federal judge halted NSF's handoff of the NCAR-Wyoming Supercomputing Center. The problem wasn't legality; it was that NSF's process was 'arbitrary and capricious,' and the expertise running the machine can't be quickly replaced.
This analysis is based on the June 1 preliminary injunction order, NSF and UCAR statements, and public NWSC/Derecho documentation.
On June 1, a federal judge told the National Science Foundation it could not, for now, hand the keys to one of its national supercomputing facilities to somebody else. Senior U.S. District Judge R. Brooke Jackson, in the U.S. District Court for the District of Colorado, granted a preliminary injunction barring NSF from "divesting UCAR or NCAR of any rights, resources, or responsibilities related to the NWSC." The University Corporation for Atmospheric Research (UCAR), which has run the NCAR-Wyoming Supercomputing Center in Cheyenne, Wyoming since it opened in 2012, brought the suit (UCAR v. NSF, No. 1:26-cv-01061-RBJ).
The trade press has filed this under climate politics. That framing misses what makes the order matter to anyone who builds or operates large systems. The judge did not rule that NSF lacks the authority to change who operates a national supercomputer. He ruled that the way NSF tried to do it was unlawful. And the way he found irreparable harm should interest any practitioner: the court treated the people who keep an integrated supercomputer running as a part of the asset that does not transfer with a contract.
Read the order as a signal, not a verdict on the future. It is a district-court preliminary injunction, an early test of how NSF moved, not a final ruling on the merits and not binding precedent for the field. But it puts a marker down on a question the rest of the field is going to face: what does it actually take to hand off a working national system, and what breaks if you get the process wrong?
NWSC is operated by NSF's NCAR through its Computational and Information Systems Lab, with UCAR as the managing nonprofit consortium. NCAR is a federally funded research and development center, an FFRDC, that UCAR operates under an NSF cooperative agreement. That agreement is not a DOE-style management-and-operating (M&O) contract, and it is not a university-hosted leadership-class award like the ones behind TACC's systems. It grants the awardee substantial day-to-day autonomy, with NSF holding what the agency calls "substantial involvement." Moving or recompeting one is a legally and operationally distinct exercise from a DOE lab recompete, where the operator can change while the facility and its assets stay federal.
The sequence matters, because the court anchored its reasoning in it. NSF issued a Dear Colleague Letter on January 23, 2026, stating it was "exploring options to transfer stewardship of the NWSC to an appropriate operator" and inviting responses by March 13 (order, pp. 9, 24). Separately, on February 12 NSF announced that NWSC operations were expected to transition to a third-party operator. The named candidate was the University of Wyoming, already a major user: the Wyoming-NCAR Alliance allocates roughly 320 million core-hours, about 20% of the machine's capacity.
Here is the procedural problem the court seized on. The comment window the DCL opened did not close until March 13. Yet on March 11, NSF Program Director Steven Ellis met with NCAR and UCAR leadership and described transferring stewardship of NWSC as "a decision made." Asked why, the order records, Ellis "responded: 'Because NSF has said so'" (order, p. 11). The judge treated that, alongside earlier implementation steps, as evidence the outcome was settled before the public process it had announced was complete.
NWSC runs Derecho, an HPE Cray EX system whose CPU partition appeared at No. 59 on the June 2023 TOP500 list (Rmax 10.32 PF, Rpeak 12.40 PF) before the machine entered full production that September. Its headline numbers, per NCAR's CISL: 19.87 petaflops peak, and roughly 3.5 times the throughput of Cheyenne, the system it replaced. The configuration is a CPU partition of 2,488 dual-socket nodes of AMD EPYC 7763 "Milan" alongside an accelerated partition of 82 nodes carrying 328 NVIDIA A100 40GB GPUs, wired with HPE Slingshot v11 in a Dragonfly topology. It is the production backbone for weather, climate, and Earth-system modeling across the U.S. research community.
That community is the stake, and the order quantifies it: the NWSC ecosystem "served 3,752 researchers from over 400 public and private institutions" in the past year (order, p. 5). CISL's own FY2025 reporting puts the figure in the same range, around 2,700 unique users across more than 420 institutions. The court also found, on the record before it, that the U.S. military, federal agencies, and private-sector partners rely on the data NWSC produces to make operational decisions (order, p. 35). When the question is whether science can access national compute capacity, this is that question in its most literal form: who runs the machine, and does the handoff put continuity at risk?
The legal hook is the Administrative Procedure Act, and it is narrow. The court found NSF's decision arbitrary and capricious for two reasons (order, pp. 23-30): the agency offered no explanation for it, and it failed to abide its own announced process for considering public feedback before proceeding. On the first, the order is blunt: "NSF's failure to provide any explanation for its decision, let alone a reasonable one, thwarts meaningful judicial review and renders the challenged action arbitrary and capricious" (order, p. 24). On the second, the judge noted that the DCL drew more than 2,500 responses. NSF moved on the transfer before the March 13 deadline and, by its own admission, had not yet "duly considered" them (order, pp. 25-27).
The ruling does not say NSF can never change who operates NWSC. It says NSF cannot do it with no stated rationale and no real comment period. The concept of an operator transition survives the order intact; what the court rejected was this particular execution of it.
Then comes the part that should interest anyone who has ever stood up or torn down a large system, the irreparable-harm analysis (order, pp. 30-34). To grant a preliminary injunction, the court had to find that the transfer would cause harm money cannot later undo. It found exactly that, and it located much of the harm in people. The order accepted that UCAR "cannot easily replace employees with the level of education, specialized training, and institutional knowledge necessary to operate and maintain the NWSC's 'highly integrated, high-performance supercomputing system'" (order, p. 33), and that recruiting and training for the recent vacancies could take from six months to three years per role. UCAR had already lost staff across CISL and other NWSC-dependent labs, with departures citing the transfer's uncertainty. On the public-interest factor, the court added that any degradation in forecasting or modeling "carries real-world consequences, including potential harm to property and human life" (order, p. 35). The court framed that as the stake riding on the system, not as a prediction of what the transfer would cause.
Note what the court did not say. It did not hold that "HPC operations expertise is non-fungible"; that is not its formulation. What it did was treat specialized operational expertise and institutional knowledge as difficult to replace, and weight that difficulty heavily in the harm analysis. For a practitioner, the logic is familiar. The people who keep the Slingshot fabric healthy, who tune the scheduler, who know why a particular MPI job stalls on this interconnect and not another, who maintain the storage stack and the domain codes carry knowledge that does not transfer with a contract. It walks out with the staff, and it does not reconstitute on a procurement timeline. ORNL has been making a version of this argument institutionally: its Next-Generation Data Center Institute treats operating exascale-class infrastructure as a specialized discipline rather than a facilities-management line item. The order does not endorse that thesis as law, but it lands close to it on the facts.
Set this against NSF's other recent moves and a real governance question surfaces.
While the agency fights in court to off-load one operating role, it is standing up new capacity under a different structure entirely. NSF's $457 million Leadership-Class Computing Facility, which will host the Horizon supercomputer, runs on a university-hosted leadership-class model, distinct from the FFRDC cooperative agreement that governs NWSC. That facility is led by the Texas Advanced Computing Center at UT Austin; Morehouse College is a partner that received $5 million to begin construction on a project-tied site, not the recipient of the whole system. So NSF is assigning a major leadership-class operating role one way and, at the same time, trying to reassign an existing one another way, under a different legal instrument and on a different timeline. Two facilities, two governance instruments, two theories of who holds operational expertise and how it is awarded.
Changing operators is not itself novel, and on the bare concept the law leaves NSF room. DOE periodically recompetes its management-and-operating contracts and the operator does change; the 2025 award of the Savannah River National Laboratory M\&O contract to Battelle Savannah River Alliance is a recent example. NSF, for its part, has rotated lead institutions across its supercomputing programs for decades, through the lineage that ran from the original Supercomputer Centers program to PACI, TeraGrid, XSEDE, and today's ACCESS. A transfer is hardly unheard-of. The harder question is who controls a nationally funded compute asset and under what instrument, and it is the domestic, governance-instrument cousin of a question playing out internationally, where some countries are deliberately keeping ownership and operation of flagship systems in domestic hands, as Finland did when it stood up its sovereign Roihu supercomputer alongside EuroHPC's LUMI at Kajaani. What the Colorado court faulted was not the idea but the execution: a national supercomputing facility reassigned with no rationale on the record and a comment process the agency moved past before it closed.
For the practitioners and procurement officers across those 400-plus institutions, the operational worry is continuity: the allocation machinery, the user support, and the data-platform stability that a clean operator transition is supposed to preserve and a botched one can rupture. The order pauses the handoff to the University of Wyoming while the case proceeds. It does not resolve who operates NWSC in the long run.
There is a broader and more contested backdrop here, and it belongs in the record as context. UCAR's complaint alleges the NWSC transfer is one piece of a larger plan to restructure or break up NCAR, associated with OMB and a campaign the suit characterizes as retaliation against Colorado. The judge noted that the timing "may" support an inference of retaliation but rested the order on APA and irreparable-harm grounds, not on motive. A separate suit brought by the State of Colorado runs on its own track and is not the case this injunction came from. The scope of this order is narrow: NWSC stewardship, nothing more.
What it sets up is a test the rest of the field will watch. Federal supercomputing facilities change hands, and they will keep changing hands. This order does not stop that. It puts conditions on it: a documented rationale, a real comment process, and a record that treats the expertise running an integrated system as part of the asset rather than an interchangeable input. Whether the order becomes influential beyond this case depends on the merits ruling, any appeal, and whether NSF documents a more coherent operating model.
