AI passed the bar exam. It writes code, diagnoses skin conditions, and summarises legal contracts in seconds. By almost every benchmark, it keeps getting smarter. And yet, show it a ten-second video of your warehouse floor and it has almost nothing useful to tell you.
At Acquinox, we view this gap, between what AI can read and what it can actually see, as one of the most underappreciated dynamics in technology investing right now. And it's where the next decade of infrastructure spending is quietly heading.
The conference room goes quiet as the COO pulls up the video feed. "Right there," she says, pointing at the warehouse floor on the large screen. "See how the loading crew hesitates before placing that pallet? That's the third time this week. Something's wrong with our workflow." She turns to her head of digital transformation. "Can we run this through the new AI system? Get some insights?"
The answer, though dressed in technical language, is essentially no. This is what is called the AI information gap: the gap between what AI can process and what actually matters in reality. AI models write code and self-improve their own models. But show them a video feed? They can't tell you what an experienced operations manager spots in three seconds. The loading crew's hesitation, the subtle timing issue, the physical tell that something in the process is broken. The algorithm sees nothing.
Where AI lives and what it cannot reach: The “90% problem”
Modern artificial intelligence, particularly large language models, operates in a world of tokens. This architecture has delivered remarkable results, but it has built an iron wall.
Writing for the World Economic Forum, Jack Hidary, CEO of SandboxAQ, put it bluntly: "The well of untapped data that fuelled the last wave of AI breakthroughs is running dry."
According to industry consensus, roughly 90% of all enterprise-generated data is unstructured. Video, audio, physical interactions, and all the things people do rather than document. Most of it never gets fed into an AI system because the architectures we rely on simply weren't built for it.
The great divide: What AI can and cannot process today
AI's text and image capabilities are real and improving fast. Voice transcription works reliably in clean conditions. But add live video, and the physics of data processing break down.
A 1080p stream at 30 frames per second carries exponentially more data per second than spoken language. Processing video in real time through centralised cloud infrastructure runs straight into a hard wall of bandwidth, latency, and cost.
The result is two parallel information economies: one AI can read fluently, and one it can barely touch. The companies building the infrastructure to bridge this divide are where strategic capital is now being deployed.
Physical AI and the pioneers forcing infrastructure to evolve
The push to close the AI information gap is not being driven by consumer applications; it is being forced by industrial and autonomous systems that must perceive the physical world in real-time.
Autonomous vehicles and embodied AI pioneers are currently running the world's largest live experiments in low-latency perception. Companies like Waymo have championed complex sensor fusion (combining cameras with radar and LiDAR to capture rich semantic context alongside precise 3D geometry). Meanwhile, innovators like Wayve are pushing the boundaries of end-to-end embodied AI, teaching models to understand physical environments through continuous real-world interaction.
Beyond vehicles, leaders like Prometheus are attempting to build foundational models specifically for physical AI, aiming to give robotic systems the "common sense" reasoning required to operate safely in unstructured environments like warehouses and factories.
The underlying challenge for all these systems remains latency. Camera-only pipelines require heavy, time-consuming inference to reconstruct depth.
The commercial answer to this latency bottleneck lies in next-generation hardware like event cameras (neuromorphic sensors). Unlike standard cameras that process heavy, continuous frames, these sensors only report changes in the environment, drastically reducing bandwidth. A Nature paper demonstrated that combining a "20 frames per second RGB camera with an event camera" can achieve the same latency as a 5,000-fps camera, with the bandwidth of just a 45-fps camera, without compromising accuracy.
For investors, this is the critical unlock: it represents precisely the kind of high-moat, foundational hardware required to make physical AI commercially viable without crippling cloud inference costs.
The investment case for closing the AI information gap
At Acquinox, we see four sub-sectors that solve problems in the AI information gap:
Sector | Thesis | Representative Companies |
Edge and on-device inference | Heavy vision models running locally so only events, summaries, and embeddings travel upstream. The necessary response to the bandwidth, latency, and privacy ceiling on centralised inference. | Hailo, Axelera AI, SiMa.ai, Kneron |
Sensor fusion and autonomy stacks | Multi-sensor architectures combining cameras, radar, and LiDAR to deliver direct geometry alongside semantic context. The infrastructure layer for commercial autonomy. | Applied Intuition, Nuro |
Physical AI foundation models | The reasoning layer between sensor input and action. Models that let machines understand space, motion, and physical environments. | Waymo, Wayve, Physical Intelligence, Skild AI, Project Prometheus |
Curated and synthetic multimodal data | Training fuel for the next generation of physical and multimodal models as web-scale text data runs dry. | Surge AI, Odyssey, Applied Intuition |
These sectors are building the physical and regulatory infrastructure that determines what information AI can access, how fast, and at what cost. The parallel to cloud infrastructure a decade ago is hard to ignore: a small number of companies built the data centres, chips, and networks that every software company now depends on. The same structural shift is playing out again, one layer deeper in the stack.
For sophisticated investors and family office capital allocation strategies, the infrastructure layer presents a different risk-return profile than application-layer AI investments. These are opportunities that require technical diligence, regulatory awareness, and patience, yet they offer potential structural defensibility once deployed.
Published by Samuel Hieber
