Inside Flyer, the Air Force's new AI supercomputer at Wright-Patterson

Why it matters

Flyer shows how military AI adoption is becoming an infrastructure race: the Air Force is buying secure compute to run its own models, simulations and vendor tests.

The Air Force Research Laboratory showed off Flyer on June 18 at Wright-Patterson Air Force Base, putting a $20 million high-performance computing system into the center of the Air Force's AI and simulation push.

Flyer is not a consumer-facing AI product, and it is not a hyperscale cloud cluster. It is more important than that for the Air Force's internal R&D machine: a secure, in-house computing system built to compress modeling, simulation and machine-learning workloads that would otherwise move through slower cycles of lab work, range time and contractor infrastructure.

The public specs are clear enough to separate the hardware from the ceremony. The DoD High Performance Computing Modernization Program systems summary lists Flyer as an HPE Cray XD system at the AFRL DoD Supercomputing Resource Center with 8.7 petaFLOPS of peak performance, 884 compute nodes, 169,782 AMD EPYC "Genoa" cores, 64 Nvidia H100 GPUs and 16 Nvidia L40 GPUs. The system is connected by a 200 Gbps Cray Slingshot network and backed by 738.9 TB of memory and 19.4 PB of usable disk storage.

That makes Flyer a specialized defense research machine, not a public-cloud AI factory. The H100s matter because they give AFRL dedicated accelerator capacity for AI and machine-learning workloads. The large CPU footprint and storage matter because much of the Air Force's hardest work is still computational science: fluid dynamics, materials, aircraft design, sensor processing, digital twins, and hypersonic modeling. The Air Force's AI story is not just model training. It is the steady conversion of engineering work into simulation work, then into faster acquisition decisions.

AFRL officials framed Flyer in those terms. In a June 18 local television report carried by K99.1FM, Brig. Gen. Douglas Wickert, AFRL commander, tied the system to the lab's role in aviation research and said the Nvidia chips and GPUs would help AFRL "invent the future." Kelly Dalton, identified in the report as director of the DoD High Performance Computing Program, said Flyer would run nonstop for five years and that the $20 million investment would save the department more than $800 million over its lifetime.

Treat the savings figure as an official claim, not an audited return calculation. The same is true of the headline comparison that Flyer can do in a day what an average laptop would take 500 years to calculate. That comparison is useful as a public demonstration of scale, but it is not the benchmark that matters. The harder number is the procurement choice: AFRL is buying dedicated CPU and GPU capacity because defense modeling and AI workloads cannot always move into commercial cloud environments, and because speed inside the lab changes the cadence of research programs.

Flyer is part of a replacement cycle, not a one-off splash

Flyer follows Raider, the AFRL supercomputer that arrived at Wright-Patterson in 2023. In its 2023 announcement for Raider, AFRL said Raider delivered about 12 petaFLOPS and replaced Thunder, a 2015 system rated at 3.1 petaFLOPS. AFRL also said at the time that Flyer and Raven were the next two systems to be installed, with Flyer supporting unclassified work and Raven supporting classified work.

That older Raider announcement is useful because it explains the operating model behind Flyer. Bryon Foster, then division chief in AFRL's Digital Capabilities Directorate, said AFRL orders the next machine while the current one is still being installed because the systems take so long to build, move into place and bring online. That is the procurement lesson buried under the AI language: high-performance computing is a rolling infrastructure program. The Air Force does not buy one large machine and pause. It stages refreshes before old systems age out.

The HPCMP computation centers page says the program delivers high-performance computing capabilities to research, development, test and evaluation, and acquisition engineering communities through five DoD Supercomputing Resource Centers. AFRL's center at Wright-Patterson is one of those hubs. The practical customer set is not limited to a single Air Force lab team; these systems are meant to serve defense scientists, engineers, testers and acquisition programs that need large-scale compute.

The cloud gap is the point

Flyer also shows the limit of the simple "AI moves to the cloud" story. AFRL Chief Information Officer Alexis Bonnell said in the Raider announcement that not every AFRL, Air Force, Space Force or DoD mission is appropriate for cloud, and that in-house high-performance computing gives AFRL secure networks and scale for its hardest questions. That line reads even more directly in 2026, after two years of federal agencies pushing AI adoption while still trying to protect sensitive data, models and mission workflows.

The AFRL Digital Capabilities Directorate describes its role as accelerating the lab's digital transformation through faster research, better decisions, streamlined transitions and reduced toil. Flyer is infrastructure for that agenda. It gives AFRL more control over where sensitive research data sits, how simulation jobs are scheduled, and how AI workloads are paired with legacy scientific computing code.

For defense startups, the signal is concrete. The Air Force is not only buying finished autonomy software, targeting tools or AI-enabled sensors. It is also expanding the compute layer that lets government labs test, simulate and evaluate those systems before they reach programs of record. That changes where leverage sits. A startup selling into defense AI has to assume the government buyer is becoming more computationally literate, more capable of running its own experiments, and less dependent on vendor demos as the only evidence that a model works.

Flyer's scale will not impress commercial AI labs racing to assemble GPU clusters by the tens or hundreds of thousands. That is the wrong comparison. The Air Force is buying enough dedicated accelerator and simulation capacity to shorten engineering loops inside a controlled environment. In defense procurement, that can matter more than raw bragging rights: a model that runs on secure data, a simulation that eliminates a failed physical test, or a hypersonics workflow that moves from months to weeks can change which program gets funded next.