Big Ideas 2026: Physical AI and the Industrial Stack

Erin Price-Wright introduces the concept of treating complex industrial challenges—from housing construction to mining to data centers—as factory problems. The key is decomposing society-scale problems into modular, repeatable processes using AI, autonomy, and skilled labor. Data centers serve as the testing ground for technologies that will later scale across infrastructure projects.

• •Apply assembly line modularity to non-traditional factory settings: housing, data centers, mines, and energy infrastructure
• •Use AI to map regulatory complexity and create formulaic, agentic processes without redesigning from scratch each time
• •Leverage rapid data center construction as a proving ground for autonomy, robotics, and AI deployment on large-scale physical assets
• •Scale learnings from data centers to manufacturing facilities, fabs, and defense sector production
• •Rebuild American industrial capacity by creating standard IP and designs that enable record-time deployment

Ryan McEntush explains that the real challenge isn't matching China's technology—it's building the industrial ecosystem to produce electrified components (batteries, power electronics, motors, compute) at scale and low cost. Success requires blending Silicon Valley software culture with industrial veterans, co-locating engineering and manufacturing, and building prestige around the mission to attract top talent.

• •The bottleneck is ecosystem development, not technology—America lacks the tier 1-3 supplier networks that exist in China
• •Companies like SpaceX vertically integrate by necessity, not strategy, because the supplier ecosystem doesn't exist yet
• •Combine Silicon Valley software talent with industrial veterans who know what's been tried before (e.g., SpaceX hiring from shuttle programs)
• •Co-locate engineering and manufacturing for faster 'designed for manufacturing' iteration cycles
• •Build prestige and purpose around the mission to compete for top talent against higher-paying alternatives
• •Owning these supply chains today determines economic and military power 50-100 years from now

Zabie Elmgren argues that physical observability—real-time visibility into physical environments using cameras, sensors, and AI—is essential for deploying robotics and autonomy safely. With over 1 billion networked cameras in the US, the challenge is fusing multimodal sensor data (thermal, RF, acoustic) to interpret what's happening, not just record it. Trust, privacy, and interoperability are fundamental design requirements, not add-ons.

• •Physical observability brings software-style real-time visibility to mines, data centers, construction sites, and critical infrastructure
• •Modern AI fuses thermal, RF, acoustic, and visual sensors to interpret context, not just capture images—like upgrading from a note-taking intern to an analyst
• •Construction sites are chaotic, remote environments where expensive materials go missing and robotics need accurate mental models of constantly changing layouts
• •The same systems that prevent wildfires or secure data centers can enable surveillance nightmares—privacy must be a design requirement
• •Winners will build the trusted perception layer that robots, infrastructure operators, and emergency responders all plug into
• •Success means becoming the 'Samsara for physical environments'—providing a live, multimodal understanding of entire sites

Will Bitsky argues that the pendulum is swinging from compute constraints back to data constraints, specifically messy, multimodal industrial data. The most defensible advantage is data collection at the source—industrial incumbents with installed bases, labor forces, and existing operations have lower marginal costs than startups building robotic farms or teleoperated consumer products to hack together datasets.

• •In 2026, the bottleneck shifts from compute/energy back to data constraints—specifically messy industrial data across manufacturing, defense, aerospace, energy, and mining
• •Foundation models need multimodal inputs: language from software platforms, spatial data, proprioceptive sensor feedback, and tactile inputs
• •Industrial incumbents have a structural advantage: they can collect data from existing operations at lower marginal cost than startups
• •Startups building robotic arm farms or selling teleoperated products to collect data face unsustainable marginal costs per data unit
• •The most defensible layer long-term is collection (top of funnel), not infrastructure or data cleaning
• •It's impossible to disintermediate industrial companies from their existing operations—they control the data walled gardens