Physical AI's real bottleneck isn't intelligence. It's knowing where you are.
Every major AI conversation right now is about models getting smarter. Mine is about something less glamorous: most AI systems built for the real world don't actually know where they are in it. That gap, not model capability, is what's holding physical AI back from the deployments everyone is promising.
We don't think this is a fringe opinion anymore. Gartner named physical AI a top strategic technology trend for 2026, describing it as tangible AI systems that sense and act in the real world. The market is moving fast enough that estimates diverge wildly depending on who's counting and how. Grand View Research, reporting in USD, puts the global physical AI market at $81.6 billion in 2025, growing to $960.4 billion by 2033. PwC's Strategy practice, reporting in EUR, models it differently, at roughly €430 billion by 2030, with aerospace and defense alone worth about €50 billion and industrial automation and warehousing another €69 billion. Different currencies, different methodologies, different endpoints, and still not remotely comparable numbers. When two credible firms disagree that fundamentally, it usually means the category is real but still figuring out its own shape. I'd rather take that as a signal of momentum than wait for consensus. Gartner (Top Strategic Technology Trends 2026: Physical AI)
Line chart showing the global physical AI market growing from 81.6 billion dollars in 2025 to a projected 960.4 billion dollars by 2033, according to Grand View Research
Only the 2025 and 2033 endpoints are directly sourced; intermediate years are interpolated at the reported CAGR to show trajectory, not independently reported data points.
Here's what none of that market math accounts for: the positioning layer underneath it is quietly broken.
GPS was never built for this
Everyone assumes GPS is a solved problem. It isn't, and the further you get from open sky, the worse it gets. Standard GPS achieves 2-4 meter accuracy in open environments, but that degrades to as much as 30 meters in dense urban areas with standard single-band receivers. Indoors, it's not a degradation problem anymore, it's an absence problem: GPS signals are weak and easily blocked by building materials like concrete, metal, and glass, which significantly reduces or eliminates signal reception entirely. Tripela, NEXUS HELP
Bar chart showing GPS positioning error in meters by environment: open sky about 3 meters, dense urban canyon up to 30 meters, with indoor environments frequently losing signal entirely
Indoor environments frequently lose fix entirely due to signal blockage by building materials, so are not shown on this scale.
That's not a chart about weak signal in unusual places. That's a chart about signal failing in exactly the environments where enterprises want to deploy physical AI: warehouses, plants, cities, contested terrain.
Defense figured this out before the rest of the market did
I'd argue defense planners are ahead of commercial enterprise on this because they had to be. GNSS denial isn't treated as an edge case in military planning, it's a standing assumption. Military unmanned systems are often deployed in anti-access/area-denial zones where GNSS spoofing and jamming are expected, forcing reliance on inertial navigation, terrain-referenced navigation, visual odometry, and SLAM just to keep flying and completing the mission. Ground systems face the identical wall: urban canyons cause multipath distortion or complete signal loss, so unmanned ground vehicles fall back on tactical-grade IMUs, LIDAR, radar, and SLAM to keep functioning. Defense Advancement
What's changing now is the tooling to solve it is finally consumer-adjacent instead of exotic. A recent example from the field: 3D vision-based navigation combines onboard cameras with high-resolution map data so a drone can continuously verify its position and hold its route even with GPS gone, using accurate, current 3D terrain data as the reference layer instead of a satellite signal. That's the shift we think matters most: positioning is moving from "trust a signal from space" to "understand the place you're actually in." That's a spatial intelligence problem, not a stronger-antenna problem. Breaking Defense
The upside is bigger than any one vertical
It's tempting to file this under defense or robotics niche and move on. We think that undersells it. McKinsey Global Institute estimates $2.9 trillion of economic value could be unlocked in the United States by 2030 if organizations redesign workflows, not just individual tasks, around people, agents, and robots working together. That number only materializes if the robots and agents in question can trust their own sense of place. An agent that reasons brilliantly but can't answer "where am I and what changed here" doesn't get to participate in that $2.9 trillion. It gets stuck in pilot. McKinsey & Company
Our take
The industry keeps competing on model intelligence because that's the visible, fundable part of the story. We think the winners in physical AI will be the teams that treat ground truth as seriously as they treat model architecture. Reconstruction, localization, and understanding aren't a nice-to-have stack, they're the precondition for everything else working outside a demo. Positioning reliability should be scrutinized in vendor evaluations with the same rigor as model benchmarks. Right now it mostly isn't. That's the gap we're betting on closing next.
Sources: Gartner (Top Strategic Technology Trends 2026: Physical AI), Grand View Research, PwC Strategy, Tripela, NEXUS Help, defenseadvancement.com, Breaking Defense, McKinsey Global Institute.