On March 23, 2026, Sikorsky delivered the U.S. Army’s first UH-60MX Black Hawk fully integrated with the MATRIX™ autonomy suite—a fly-by-wire, optionally piloted helicopter capable of executing missions with no crew aboard. That same month, Anduril locked in a $20 billion, ten-year enterprise contract for its Lattice AI platform, while Munich-based Helsing partnered with Mistral AI to co-develop Vision-Language-Action models for European strike drones already deployed in Ukraine. These are not incremental upgrades. They are architectural inflection points—proof that the defense sector is undergoing the same disaggregation-and-software-recomposition cycle that transformed the automotive industry over the past decade.

I call this convergence Neo Defense Tech: the application of Software-Defined Vehicle (SDV) principles, autonomous-vehicle (AV) data architectures, and embedded high-performance AI compute to platforms that fly, swim, crawl, and fight. What makes Neo Defense Tech distinct from traditional defense modernization is not any single technology; it is the way three reinforcing patterns—disaggregated control systems, data flywheel architectures, and end-to-end VLA models validated on real-world deployment data—combine to compress the innovation cycle from decades to months.

A. Disaggregated Control Systems: From Monolith to Modular

The autonomous Black Hawk illustrates the first pattern vividly. Sikorsky’s MATRIX system replaces the helicopter’s legacy mechanical flight controls with a full-authority fly-by-wire backbone and layers an open-architecture autonomy mission manager on top. Crucially, the ALIAS Optimally Piloted Vehicle kit includes a software development kit that invites third-party sensors and algorithms—turning a 1970s airframe into a software-defined platform. This is precisely the zonal-controller disaggregation that the automotive world pioneered with domain and zone ECU architectures over the past five years.

In automotive SDV, NVIDIA DRIVE AGX Thor delivers up to 2,000 TOPS of FP4 performance (1,000 INT8 TOPS) on a single Blackwell-based system-on-chip, centralizing ADAS, infotainment, and cockpit functions. A full-size autonomous helicopter like the Black Hawk requires comparable compute density—multiple sensor modalities (LIDAR, radar, EO/IR), terrain-relative navigation, obstacle avoidance in degraded visual environments, and real-time mission replanning—demanding an estimated 500–1,000+ TOPS of aggregate onboard inference, potentially serviced by multi-SoC configurations. These are safety-critical, ASIL-D-class workloads running at the edge, with no round-trip to the cloud allowed during a contested logistics insertion.

At the opposite end of the size spectrum, tactical drones like Anduril’s Ghost and Helsing’s HX-2 operate on far leaner compute budgets—typically 15–275 TOPS, using NVIDIA Jetson-class modules (Xavier at ~32 TOPS, AGX Orin at ~275 TOPS) or comparable edge SoCs from Qualcomm. Weight, power, and thermal constraints dominate: a loitering munition carrying a 2 kg warhead cannot afford a 60-watt compute module. Helsing’s HX-2, for instance, uses onboard AI and stored map data to navigate and engage targets without GPS—relying on vision-based inference that must run within a power envelope of roughly 10–20 watts.

The key insight borrowed from SDV architecture is disaggregation of the control plane from the platform plane. Whether the vehicle is a 10-ton helicopter or a 15 kg strike drone, the autonomy stack separates into perception, planning, and actuation layers that can be independently updated, tested, and scaled. MATRIX on the Black Hawk, Lattice on Anduril’s fleet, and Altra on Helsing’s drones all embody this principle. The disaggregated architecture also enables what Sikorsky calls “optionally piloted” modes—seamless transition between crewed, remote-supervised, and fully autonomous operation—mirroring the L2–L4 spectrum that automotive OEMs navigate today.

Compute Spectrum: TOPS Across Neo Defense Platforms

B. The Data Flywheel: Deployments Drive Intelligence

Anduril’s $20 billion Army enterprise contract is, at its core, a data infrastructure deal. Lattice consolidates more than 120 previously separate procurement pathways into a single open-architecture C2 backbone connecting sensors, effectors, and AI decision engines. Every sensor feed, every engagement, every counter-UAS intercept generates labeled training data that flows back into the model pipeline. This is the defense equivalent of Tesla’s shadow-mode data flywheel or Waymo’s fleet learning loop—the more you deploy, the smarter your models become, the faster the next iteration ships.

Helsing has operationalized this pattern in the most demanding test environment imaginable: the Ukrainian battlefield. With more than 10,000 HF-1 and HX-2 airframes ordered for Ukraine, each sortie returns real-world sensor logs from GPS-denied, electronically jammed environments—data no simulation can replicate. Helsing’s Altra reconnaissance-strike platform fuses multi-drone sensor feeds in a ground station with high compute capacity, producing annotated engagement datasets that accelerate the retraining of targeting and navigation models. Germany’s subsequent €9 billion loitering-munition program, split between Helsing and Stark Defence, ensures this flywheel will scale under sovereign European production.

The Black Hawk program feeds its own flywheel differently. The Army’s DEVCOM will use the UH-60MX to develop techniques, tactics, and procedures (TTPs) for autonomous logistics, CASEVAC, and contested resupply. Hundreds of flight hours of autonomous operation under controlled test conditions produce perception failure cases, edge-case terrain maps, and degraded-visual-environment datasets that loop directly into MATRIX model updates delivered through the open SDK. Near Earth Autonomy, which received a separate $15 million Army contract for Black Hawk retrofit kits in 2025, brings over 10,000 prior autonomous flight logs across 140+ airframes into this data pool.

The flywheel dynamic creates a powerful moat. Incumbents who deploy first—and deploy at scale—accumulate proprietary training corpora that new entrants cannot replicate without equivalent operational exposure. This is why Anduril’s Lattice marketplace presence and Helsing’s Ukraine deployment cadence matter strategically as much as any hardware specification.

C. End-to-End VLA Models: The SDV–Defense Convergence

The most consequential technology transfer from autonomous vehicles to defense is the emergence of Vision-Language-Action (VLA) models—architectures that perceive the environment through multimodal sensors, reason about mission objectives in natural language, and output continuous control actions in real time. In automotive, NVIDIA’s DRIVE AGX Thor developer kit is explicitly optimized for VLA and LLM workloads, and companies like Kodiak AI are deploying reasoning VLA models on dual-Thor configurations for Level 4 autonomous trucking in the Permian Basin today.

Helsing and Mistral’s partnership, announced at the 2025 Paris AI Action Summit, targets exactly this architecture for defense: Vision-Language-Action models that enable strike drones and reconnaissance platforms to interpret their surroundings, communicate naturally with operators, and execute faster decisions in contested scenarios. In robotics, Figure AI’s Helix VLA demonstrates the paradigm’s maturity—a single neural network controlling a humanoid’s full upper body at 200 Hz, running entirely on embedded low-power GPUs. The defense application is a direct analog: an onboard VLA that fuses camera, radar, and electronic-warfare sensor data into a world model, accepts mission-level language commands (“engage armored vehicle at grid reference”), and generates continuous flight-control or guidance outputs.

What makes VLA models transformative for defense is their ability to generalize from real-world deployment data in ways that hand-coded autonomy stacks cannot. Traditional waypoint-following and rule-based obstacle-avoidance systems require extensive manual engineering for each new environment. A VLA trained on thousands of Ukrainian battlefield sorties can transfer learned representations—camouflage detection, jamming-resilient navigation, multi-target prioritization—to entirely new theaters without reprogramming. This is the defense-domain equivalent of a self-driving model trained on San Francisco streets performing competently in snowy Munich.

The SDV–Defense Convergence: A Perspective

The pace of Neo Defense Tech advancement is startling. Helsing went from founding in 2021 to battlefield-deployed AI strike drones and a €5 billion valuation in under four years. Anduril’s Lattice went from border surveillance towers to a $20 billion enterprise C2 backbone in under nine years, reaching a $60 billion valuation in its latest round. The Black Hawk—first flown in 1974—now operates autonomously with a retrofit kit that separates the autonomy brain from the airframe entirely. In every case, the velocity comes from borrowing and adapting patterns proven at scale in commercial autonomy.

Three forces will define the next phase. First, compute density at the tactical edge will continue to climb. The jump from Jetson Xavier (32 TOPS) to AGX Orin (275 TOPS) to DRIVE Thor (1,000 INT8 TOPS) happened in roughly four years; the next generation will bring transformer-optimized, safety-certified inference to platforms that today run on 15-watt budgets. This unlocks VLA-class reasoning on expendable drones—not just piloted helicopters.

Second, data flywheel compounding will separate leaders from followers. The companies with large-scale, real-world deployment data—Anduril through JIATF-401 and allied programs, Helsing through Ukraine and the Bundeswehr—will retrain and redeploy models at cadences that traditional defense primes simply cannot match. Lattice’s presence on the AWS GovCloud marketplace further democratizes access but also centralizes the data gravity around Anduril’s ecosystem.

Third, end-to-end VLA validation on real deployment data will become the gold standard for autonomous defense certification. Just as automotive L4 programs demand billions of miles of combined real and simulated testing, defense regulators and operational commanders will require VLA models to demonstrate robustness across GPS-denied, EW-contested, and multi-domain environments before fielding. The programs generating that validation data today—the UH-60MX test campaign, Helsing’s Ukrainian sorties, Anduril’s counter-UAS engagements—are building the certification corpus of the future.

Neo Defense Tech is not a buzzword. It is the structural convergence of SDV architecture, AV-class data pipelines, and embedded AI compute into defense platforms that learn, adapt, and deploy at startup speed. The Black Hawk’s MATRIX brain, Anduril’s Lattice backbone, and Helsing’s VLA-powered strike drones are the early proof points. The organizations—and nations—that master disaggregated compute, close the data flywheel loop fastest, and validate VLA models on real-world engagements will define the next era of autonomous defense.