Our Take
Anduril is building a weapon interface, not a productivity tool, and the company's confidence that soldiers will adopt it outpaces evidence that cognitive overload won't just move from radio chatter to AR displays.
Why it matters
This is the first serious attempt to fuse AI-driven threat detection with frontline decision-making at scale. The stakes are high: if it fails, the Army wastes billions (as it did with Microsoft); if it works, every allied military copies it.
Do this week
Defense contractors: audit your supply chain dependencies now before the $20 billion Lattice integration forces Chinese-component phase-outs.
Two military AR projects, one Army contract, one self-funded bet
Anduril secured a $159 million prototyping contract last year to build augmented-reality glasses for the Army's Soldier Born Mission Command (SBMC) program, in partnership with Meta. The glasses attach to existing military helmets and will overlay tactical information onto a soldier's field of view: compass data, maps, drone locations, and AI-identified targets.
In parallel, Anduril is self-funding a second project called EagleEye, announced in October, that integrates the same technology into a standalone helmet-and-headset combo. The company insists the Army will prefer EagleEye despite the Pentagon's preference for the SBMC glasses.
Both systems rely on voice commands and eye-tracking to control the interface. A soldier speaks plain-language orders like "scan that sector" or "recommend next move," and a large language model (per company testing: Google's Gemini, Meta's Llama, or Anthropic's Claude) translates speech into software commands. Multi-step tasks are possible: a soldier sends a drone to search for a target, and the system recommends a strike—subject to approval by the chain of command.
The backbone is Anduril's Lattice software, which the Army announced in March it would spend $20 billion to integrate across its infrastructure (company-reported). The glasses will run generative AI and machine learning algorithms locally, since battlefield conditions often lack 5G connectivity.
Both prototypes include a new digital night-vision system using electronic sensors and algorithms to amplify low light. Component procurement began in March; unlike Meta's commercial Ray-Bans, these parts avoid Chinese supply chains to comply with federal military contracting rules.
Production timelines remain distant. The Army is not expected to pick a final SBMC vendor before 2028, if at all. (Microsoft led the previous iteration and was set to receive a $22 billion production contract, which was cancelled after the Pentagon audited testing failures.)
Two competitors are also prototyping: Rivet (a wearable-sensor specialist) won a $195 million SBMC contract, and Israel's Elbit secured a $120 million contract in March.
Cognitive overload, not just interface design, is the unsolved problem
The core challenge is not technical—it is human. Jonathan Wong, a former US Marine and RAND policy researcher tracking Army technology adoption, notes that even simple dual-channel radio communications overloaded his situational awareness as a platoon commander. Adding AR-driven AI recommendations to a soldier already carrying 100+ pounds of gear, in dust and smoke, with no cell backup, creates a new failure mode: the soldier ignores the system because it demands more mental bandwidth than it saves.
Anduril's bet is that eye-tracking and voice commands are sufficiently low-friction to break that pattern. But Wong cautions that field validation will take years; computer vision and AI decision support have been used by militaries in drone operations and in the Iran conflict, but not yet at the individual soldier level, where miscalibration or false positives have immediate lethal consequences.
The supply-chain retooling is also a long-term forcing function. The $20 billion Lattice integration means legacy Chinese components will be phased out across the Army's infrastructure—a structural shift, not a procurement detail.
Validate the cognitive load assumption before scaling
If you are evaluating AR or AI-driven command-and-control systems for field personnel, do not assume that voice and eye-tracking interfaces solve information overload. Run cognitive load tests with end-users carrying full operational gear in high-stress conditions (noise, fatigue, divided attention). Measure not adoption rate but task-completion time and error rate under realistic scenarios. Field trials will reveal what lab prototypes hide: whether the system is faster than the radio-plus-map baseline, or just a different kind of slow.