Our Take
YC's batch shift toward physical AI and robotics infrastructure is real, but the real move is deeper: startups are now building the training data and simulation layers that physical systems cannot do without, not the robots themselves.
Why it matters
Venture investors and corporate strategy teams use YC cohorts as leading indicators of where technical problems are hardest and most fundable. A 100% increase in industrials and defense companies signals that physical AI deployment bottlenecks are now obvious enough to attract founder attention and capital.
Do this week
Robotics teams: audit your training data sourcing strategy this week before you lock in a simulation vendor, because startups like Asimov and Fern are positioning proprietary datasets and evaluation infrastructure as defensible layers in the robotics stack.
YC Winter 2026 batch tilts hard toward physical systems and supporting infrastructure
Y Combinator's Winter 2026 cohort of 199 companies includes 25 focused on physical AI: robots, drones, wearables, and space hardware. That represents roughly 1 in 8 startups, a material concentration for a program historically dominated by software. The shift extends beyond the hardware count itself.
Industrials and defense companies doubled from 17 in the previous cohort to 35 in W26, making it the third-largest category. AI infrastructure became the second-largest category at 39 companies, with a visible sub-specialization emerging: instead of full-stack agent platforms, startups are now targeting single bottlenecks. Terminal Use focuses on frontier agentic systems; Salus on guardrail validation; Carrot Labs on continuous learning.
Physical AI and its supporting layers dominate the trend analysis. Companies like Servo7, RoboDock, and Hetherington Robotics are building warehouse robots, autonomous depot infrastructure, and robotics tooling stacks. But the infrastructure layer tells the harder story: Asimov is collecting real-world human activity data at scale to annotate datasets for robot learning. One Robot is building world models that simulate physical interactions like gripping and assembly. Fern is running high-fidelity simulations on proprietary training data.
Training data scarcity, not hardware, is the binding constraint
Earlier robotics waves assumed that training data would follow hardware adoption into the field. YC W26 companies are betting the opposite: data scarcity is acute enough now that founders are building infrastructure to solve it directly.
This distinction matters because it points to where venture capital and technical talent will concentrate over the next 18 months. Startups that control proprietary, real-world datasets in robotics are positioned to capture outsized value. If hardware deployment is the end product, data sourcing and simulation validation are the defensible layers. The batch signals that the field knows it.
Secondary trend: agent infrastructure is maturing past the platform layer into granular, problem-specific modules. This typically indicates that production deployments are real enough to expose specific failure modes. The production-readiness gap for enterprise agents is closing faster than most buyers expect.
Energy infrastructure emerged as a surprise category, with Squid, Voxel Energy, and Condor Energy each targeting AI-driven optimization in constrained systems. The thesis is self-referential: AI demand is generating the infrastructure bottleneck that this cohort is solving. They're not building against macro trends; they're building against demand their own industry is creating.
Audit your data and simulation strategy before lock-in
If you are building or deploying physical AI systems, your training data sourcing and simulation infrastructure choices are now moving faster than your hardware roadmap. Startups like Asimov and Fern have momentum and defensible positions in those layers. Evaluate them alongside or before you commit to a hardware platform, because data control is where the real leverage sits.
For corporate venture teams and M&A strategists, physical AI infrastructure startups (data, simulation, evaluation) are higher-leverage acquisition targets than individual robot or drone companies. They sit closer to the bottleneck.