The Hardware Bottleneck AI Can’t Fix

What It Covers

Jason Hock, CEO of Nominal, explains why hardware engineering lacks the observability and tooling software teams take for granted. Nominal builds a data platform managing high-frequency sensor data from physical assets, covering real-time control room monitoring, post-test analysis, and simulation correlation for aerospace, defense, and energy hardware programs.

Key Questions Answered

• •Hardware data loss at ingestion: One Nominal customer discovered only 10% of test data reached long-term storage due to misconfigured schema versioning when operators forgot to update config values after software downgrades. Automating version sync between aircraft software and ground control eliminated this silent data loss entirely — a fix requiring no operator action.
• •Hot/cold pipeline architecture for hardware: Supporting both real-time control room monitoring and deep post-test analysis requires maintaining synchronized hot and cold data paths. Nominal's architecture lets engineers write logic once and apply it across both use cases, abstracting away the fact that sub-second latency streaming and OLAP-style batch analysis are fundamentally different infrastructure problems.
• •Edge agents with local buffering over dropping data: In network-constrained test environments, Nominal deploys edge agents that prioritize safety-critical data for immediate transmission while buffering everything else locally for later upload. Since high-value tests like rocket fires produce dense data in short bursts — seconds of firing, hours of downtime — amortized upload catches up without losing a single data point.
• •Asset hierarchy and tagging as foundational infrastructure: Correlating a test result against its corresponding simulation takes two clicks when data is properly tagged and cataloged, versus writing custom SQL queries when it is not. Hardware organizations should define asset hierarchies — aircraft, engines, subsystems — and tag data at ingestion, because component swaps, like moving an engine between aircraft, make lineage tracking operationally critical.
• •AI agents cannot yet close the hardware feedback loop: Software agents iterate because code execution provides near-instant feedback. Hardware lacks this — a single rocket fire test represents years of work and tens of millions of dollars. Closing this gap requires accumulating labeled datasets where domain engineers tag regions of interest, anomalies, and expected behavior, building the training foundation that could eventually support hardware design agents.