A survey of 700 practitioners building visual and physical AI systems, cited by IEEE Spectrum AI in its 2026 state-of-the-field report, converges on one blocking issue: teams are running into a data bottleneck that is slowing deployment more than any modeling breakthrough could fix. The bottleneck isn't abstract — it shows up as projects stalled at the data-collection and labeling stage, long before a model ever reaches a robot arm or a camera pipeline.

The report frames the slowdown as the product of two compounding constraints: limited compute capacity and an insufficient supply of training data. Neither problem is new on its own, but together they create a chokepoint that text-based AI, with its near-limitless supply of scraped web data, has never had to deal with.

For context, "visual AI" here covers computer vision systems built on image and video models, while "physical AI" refers to embodied systems — robots and autonomous machines — that need to act in the physical world, not just recognize what's in a frame. Both categories depend on data that's fundamentally harder to source than web text.

According to IEEE Spectrum AI, the survey's authors treat this as the central obstacle standing between current visual and physical AI systems and production-scale deployment — and they build the case from practitioner responses, not vendor claims.

Where the bottleneck actually forms

Two structural constraints show up repeatedly in the survey:

• Compute capacity — training and iterating on visual and physical AI models demands sustained compute budgets that many teams can't scale on demand, which slows experimentation cycles.
• Data scarcity — there isn't enough training data of the right kind available to reach the performance levels these systems need before they're trusted in production.

Neither constraint is fatal on its own. Combined, they mean teams can't simply compensate for thin data by throwing more compute at the problem, and they can't compensate for compute limits by training smaller, more data-efficient models if the data itself doesn't exist.

Why visual and physical AI can't scale the way language models did

Large language models scaled quickly in part because the internet already contained an enormous, freely available corpus of text. Visual and physical AI don't have that shortcut. Camera footage, sensor logs, robot telemetry, and real-world interaction data are expensive to capture, harder to label consistently, and far more constrained by privacy, safety, and physical access than a web crawl.

That structural difference is what turns "we need more data" from a minor complaint into a production-blocking constraint. A chatbot can be retrained on a larger text dump relatively cheaply; a robotics team can't retrain on a larger set of real-world manipulation episodes without literally running more physical trials.

What the bottleneck means for teams shipping these systems

For builders working on computer vision products or embodied/robotic systems, the survey's framing has direct operational consequences:

• Budget for data collection and curation as a first-class cost center, not an afterthought to model training.
• Treat compute allocation and data supply as a joint constraint when planning a training run, not two separate line items.
• Expect production timelines that look nothing like LLM release cadences — physical and visual data pipelines move on a slower, more capital-intensive clock.
• Treat labeling quality as a competitive advantage, since inconsistent labels compound the data shortage instead of offsetting it.

In our estimation, this is also why synthetic data and simulation environments have drawn so much investment across the robotics and computer-vision space recently — they're one of the few levers available for expanding data supply without expanding physical data-collection operations, though the survey itself doesn't quantify how much of the current data gap synthetic approaches actually close.

AiiN's takeaway

The practical lesson for teams building on visual and physical AI is to stop benchmarking their roadmap against language-model scaling curves. The bottleneck described here isn't a temporary rough patch waiting on the next model release — it's a structural feature of working with real-world sensor and interaction data, and it will keep shaping which products reach production first. Teams that treat data acquisition as core infrastructure, on par with model architecture decisions, are the ones best positioned to get through the chokepoint the survey describes.