AI for machines that cannot phone home.

The AI Cowboys is a Service-Disabled Veteran-Owned AI lab in San Antonio, Texas, founded in 2024 by retired Air Force leader Michael J. Pendleton. We work on AI and neuromorphic computing, physical and embodied AI, cybersecurity, and quantum computing, and we teach all of it as an NVIDIA Deep Learning Institute and IBM SkillsBuild partner. Everything we build can deploy on-premises, air-gapped, or in your own cloud.

Federal agencies and the defense community, critical-infrastructure operators, universities and research institutions, healthcare, and enterprise. Named work includes the FBI, the NSA, the WWE and the NFL Players Association. We partner with NVIDIA, Amazon Web Services, and the University of Texas at San Antonio.

A briefing first, then a scoped pilot against one real problem before anyone commits to scale. You get a roadmap with timelines and deliverables out of the first conversation, whether or not you proceed. Email contact_us@theaicowboys.com to set one up.

Yes. We are a Service-Disabled Veteran-Owned Small Business registered in SAM.gov, which makes us eligible for SDVOSB set-asides and sole-source awards. Our work is aligned to NIST, CMMC and FedRAMP, with SBOM and EO 14028 practice built in.

San Antonio, Texas, downtown at the UT San Antonio College of AI, Cybersecurity and Computing (San Pedro I, 506 Dolorosa St). That places us inside one of the densest cyber and national-security communities in the country: Joint Base San Antonio, the NSA's Texas Cryptologic Center and 16th Air Force are all here, about an hour from Austin. We work with clients across the United States and on-site for data-sensitive projects.

Yes, and it is a substantial part of what we do. We are an NVIDIA Deep Learning Institute partner and an IBM SkillsBuild partner, running hands-on workshops in AI, machine learning, quantum computing and cybersecurity for federal agencies, enterprise teams, universities and school districts. Sessions are built around live exercises and real use cases rather than slides, and the curriculum is tailored to your sector and skill level. We also run free community workshops, including AI and the Job Market and Building Financial Readiness Through AI Tools.

Because in all three, energy is now the binding constraint rather than a line item. Data center operators are limited by the megawatts they can get to the building, so anything always-on that can run at a fraction of the draw frees capacity for the workloads that genuinely need a GPU. In defense acquisition the constraint has a name, size, weight and power, and it decides what can be fielded on a drone, a satellite or an unattended sensor at all. In robotics it is simply runtime: perception is what drains the battery, so the power budget sets how long the machine is useful between charges. Neuromorphic silicon is event-driven, drawing power only when something happens, which is why it changes the arithmetic for continuous sensing specifically.

In our San Antonio lab, a hybrid deployment pairing a BrainChip Akida processor with an NVIDIA Jetson held the same accuracy at 9 watts that the Jetson alone required 120 watts to reach, roughly a 90% reduction on that workload. Across sensory and pattern-recognition tasks generally, spiking neural networks on neuromorphic silicon are commonly measured at 100x to 1,000x lower energy per inference. The saving is architectural: a neuron consumes power only when it has something to report, rather than every clock cycle.

Yes, and we have. The Akida handles always-on sensory processing at very low power while the Jetson stays idle until the Akida signals something worth escalating. That split is what produced the 9W-versus-120W result above at matched accuracy. Be clear about the cost: the model has to be converted to spiking form, and the toolchains are younger than PyTorch. But the hybrid pattern works today on hardware you can buy this quarter, which is why we recommend a bounded pilot on one real workload rather than a platform decision.

The terms overlap heavily and are often used interchangeably. Physical AI usually refers to systems that sense and act on the physical world: robots, drones, vehicles, unattended sensors. Embodied AI emphasizes that the intelligence sits in a body whose sensors and actuators shape what it can learn and do. In practice the engineering problem is the same: perception, decision and actuation have to close the loop on the platform, under a power budget, often with a degraded or absent network link.

Start with a model whose weights you are allowed to hold, because anything reached by API is off the table. Size it to the hardware you can actually put inside the enclave; a well-tuned Small Language Model on your own corpus usually beats a larger general model there. Then plan the parts people forget: model and dependency supply chain with an SBOM, an update path that works without outbound network, evaluation harnesses that run inside the boundary, and audit logging your accreditor will accept.

Continuous verification reads behavioral signal, such as how a person types, moves and hesitates, and re-establishes identity throughout a session rather than once at sign-in. It matters because agentic AI can now operate any interface a person can, so a credential checked at the login screen proves nothing about who is acting ten minutes later. Our product MagenTrust does this with entropy-based behavioral biometrics: the signal is derived and scored locally, cannot be phished, and needs no second device in the operator's hand.