BQP's Orbital Object Classification AI Heads to SpaceWERX Contract, Quantum-Inspired Technique Under Scrutiny

• What happened: BQP has won its first U.S. federal contract from SpaceWERX to develop and validate PC-QAML, a system for classifying orbital objects and their behaviors.
• Why it matters: In the labor-intensive processing of uncorrelated observations, a low-power classification AI could help prioritize follow-up tracking more quickly.
• What to watch next: The contract phase, evaluation data, false alarm rates, and demonstration on space-qualified processors have not yet been disclosed.

BosonQ Psi Federal LLC (BQP) announced on July 17, 2026, that it had secured its first federal contract from SpaceWERX. The contract covers "PC-QAML," a system that combines physics-informed modeling with quantum-computing-inspired machine learning concepts to classify orbital objects and their behaviors. This marks a progression from ground-based testing at the 2025 Space Domain Awareness (SDA) TAP Lab to a government-contracted development and validation phase. However, this is neither a plan to put a quantum computer on a satellite, nor an announcement of demonstrated on-orbit performance.

The technology aims to enable inference using the limited computational resources of satellites or forward-deployed systems, before observation data is fully transmitted to large ground-based computing infrastructure. What matters here is not the label "quantum" but rather which part of the lengthy process of handling uncorrelated observations can actually be shortened—and whether BQP's claimed dramatic model compression holds up on orbital data as well.

What the Contract Moves Forward Is a Validation Process, Not Operational Deployment

What BQP secured is an Open Topic Small Business Innovation Research (SBIR) contract from SpaceWERX. The company's announcement states that it will "develop and validate" PC-QAML—using future tense—and does not claim that the U.S. Space Force has certified its performance. The milestone of a company's first federal contract should be read separately from adoption into an operational system.

Open Topic contracts come in two stages: Phase I, which explores feasibility and potential customers, and Phase II, which advances to prototyping and demonstration. According to AFWERX's explanation, Open Topic Phase I contracts typically run 90 days or less, while Phase II contracts are typically two-year R&D awards. Yet BQP has not disclosed which phase this is, nor the contract value. The contract period and contract number also remain unknown. The bare fact of having secured a contract does not tell us whether this is a short feasibility study or a prototyping stage partnered with an operational user.

The preceding SDA TAP Lab is an experimentation and acceleration hub under Space Systems Command's System Delta 85. It applies technologies from companies and universities to operational challenges, helping move them from proof-of-concept toward transition. According to BQP, the 2025 Mini-Accelerator detected on-orbit separation events, positioning the technology as a candidate for future uncorrelated track classification, threat simulation, and catalog integration. The current contract represents an entry point for advancing that candidate technology to the next stage of evaluation.

A UCT Is Not an "Unknown Object"—It's an Observation That Hasn't Yet Been Correlated

The U.S. Space Force's SDA doctrine defines an Uncorrelated Track (UCT) as "an observation that has not been linked to an object in the Department of Defense's maintained catalog of Earth-orbiting objects." In other words, a UCT is not a category of object such as an unidentified hostile satellite. It can include a newly launched satellite, an object separated from a spacecraft, debris generated by a breakup, or even a brief observation of an already-known object captured for only a short time.

The 18th Space Defense Squadron describes a UCT as a "short track segment," typically consisting of a series of observations gathered by a single radar over a brief period. The work of correlating multiple observations to the same object, estimating an orbit, and updating the catalog when necessary is labor-intensive and time-consuming. This involves several distinct tasks: correlating observations with one another, orbit determination, classifying objects and behaviors, and threat assessment.

What PC-QAML targets, according to public materials, is primarily the latter task—classification. If features suggestive of separation events or proximity operations can be narrowed down at the edge, this could support decisions about which observations to point sensors at, or which tracks to escalate quickly to analysts. However, even high classification accuracy does not guarantee the ability to correlate the same object across short observation arcs. BQP has not announced that PC-QAML automates orbit determination or catalog management.

From 14,000,000 to 2,000: How to Read the 7,000x Claim

The comparison figures BQP presents are striking. According to the company, PC-QAML reduces a conventional model's 14,000,000 (14 million) parameters down to 2,000, while maintaining over 99% classification accuracy. By simple arithmetic, that's a 7,000x reduction—approximately 99.9857%. The company further claims up to a 10x reduction in inference latency, roughly 90% lower power consumption, and faster retraining.

However, the announcement presents these figures immediately following a reference to "similar results demonstrated in other applications." The dataset used for orbital object classification, the number of classes, the train/test split, and the configuration of the conventional model being compared against are not disclosed. Accuracy above 99% also cannot be given operational meaning without precision, recall, and false alarm rate figures. NIST itself has noted that in a dataset where one of two classes accounts for 98% of the data, even a naive model that always outputs that single class would achieve 98% accuracy. The more imbalanced the data—as is typical when searching for rare anomalies—the harder it becomes to judge performance from simple accuracy alone.

BQP's previously published research offers a comparison with more visible conditions than the current case. A 2024 IEEE Quantum Week paper and the company's published PCT patent evaluated a Physics-Informed Neural Network (PINN) combining quantum-inspired hidden layers with classical layers, using the 1D Burgers' equation. In configurations equivalent to 4 and 5 qubits, the model reduced trainable parameters by 20% compared to a classical PINN while maintaining accuracy. However, in the 3-qubit-equivalent configuration, there were regions where performance fell short of ground truth, and the experiments were run using PennyLane's local simulator in a cloud environment.

This 20% figure and the current ~99.9857% claim cannot be directly compared, since the tasks and models differ. If anything, what the published research shows is that adding a quantum-assisted layer does not always yield the same compression ratio. For PC-QAML as well, the practical value of the claimed 7,000x reduction can only be assessed once a confusion matrix on orbital data, generalization to unseen conditions, and the hardware used to measure power consumption are made public.

Before a Quantum Computer, a 5W-Class Edge Implementation

The "physics constraints" in PC-QAML refer to the concept of embedding known relationships—such as orbital dynamics—into the training loss function or model architecture, narrowing the search space that a purely data-driven model would otherwise explore. In BQP's published patent, the QA-PINN incorporates physical equations and initial/boundary conditions into the loss function, passing the output of a quantum-circuit-inspired hidden layer to classical layers. However, it remains unclear whether the current orbital PC-QAML shares the same architecture, or which physical laws it enforces.

On the implementation side, BQP states that it ran PC-QAML on an NVIDIA Jetson Nano at the BMC3I TAP Lab. The Jetson Nano is a compact device with a 4-core Arm CPU, 4GB of memory, 5W and 10W power modes, and an onboard 128-core Maxwell GPU. There is a gap in explanation here: while BQP states that the system does not depend on cloud or GPU resources, it has not clarified whether it avoided using the Jetson Nano's integrated GPU, or whether it simply means that no external data-center-class GPU was required.

Another boundary lies between ground-based development boards and space-qualified computers. While BQP states that its future design is intended to run directly on space-qualified processors, the target processor, radiation-environment fault mitigation, thermal design, and on-orbit evaluation plans remain undisclosed. Running on the TAP Lab demonstrates the potential for edge implementation, but it does not equate to completing satellite-flight qualification.

The next opportunity for disclosure is close at hand. At the AMOS Conference, running September 15–18, 2026, BQP's Alex Khan is scheduled to present related research titled "Edge-Deployable Physics-Informed and Quantum-Assisted Neural Networks for Real-Time Autonomous Orbit Prediction." Will that presentation reveal the model architecture, data, error metrics, and execution hardware? Combined with disclosure of the SpaceWERX contract phase, performance metrics, and reproduction on space-qualified processors, PC-QAML could move beyond the "quantum-inspired" label to become a tool that genuinely shortens the time needed to process uncorrelated observations.