Physical AI and robotics are forcing a rethink of enterprise infrastructure, according to Evan Helda, Head of Physical AI at Nebius. The shift from lab to production deployment exposes limitations in cloud platforms designed for batch-style LLM workloads.
Three infrastructure problems
First, training data. Physical AI can't train on internet text. It requires context-specific sensor streams, LiDAR, video, and motion data mapped to physical outcomes. Real-world data collection is slow and expensive, making simulation critical for scale. But running thousands of parallel simulations demands GPU orchestration optimized for throughput, not latency, a distinct workload from training or inference.
Second, data usability. Once robots deploy, teams face massive volumes of noisy, time-sensitive multimodal data. Object storage doesn't cut it. Data must be indexed, synchronized, and searchable through automated pipelines. Without purpose-built platforms for multimodal ingestion, more data doesn't mean better models.
Third, data movement. Robotics generates continuous sensor streams requiring millisecond responses. This rules out centralized batch processing. Physical AI increasingly relies on fast edge inference paired with cloud-based planning models, operating as a single system. Existing platforms struggle with sustained, high-throughput multimodal data. Scaling GPUs alone fails if data can't move efficiently between devices, storage, and compute.
Industry context
The timing matters. Global industrial robot installations hit $16.7 billion in 2025, and Deloitte predicts sales will exceed 1 million units and $20 billion revenue by 2030. CES 2026 showcased robot-as-a-service models from Hyundai and Boston Dynamics, signaling infrastructure demands will intensify.
The skepticism is warranted. Businesses struggle with deployment complexity, maintenance, and skills gaps. Security experts highlight cyber risks in cloud-connected robots. The question isn't whether robotics needs new infrastructure, it's whether enterprises can deploy it without operational disruption.
Hardware reliability becomes a first-order concern when simulation failures derail training cycles. Price-performance ratio and mean time to failure matter more than raw specs. History suggests early movers who solve data movement and simulation orchestration will have an edge.