Y Combinator’s latest Demo Day cohort appears to have produced a different kind of investor conversation: less focus on software wrappers and more on the energy, networking, chips, data and machines needed to make AI and automation work in the physical world.
TechCrunch asked early-stage investors to identify the most discussed companies in the batch, selecting nine startups mentioned by at least two investors. The list is not a measure of commercial durability, and several companies’ traction figures are self-reported. But it offers a useful view into where early-stage capital sees bottlenecks—and where founders are trying to build around them.
The AI infrastructure constraint is broadening
Three of the highlighted companies target different pressure points in AI compute.
Automarine proposes floating data centers, initially using gas power and ultimately nuclear-powered ships. Its premise is that power availability, cooling and local opposition can slow terrestrial data-center projects. The company says it plans a gas-powered pilot by 2028 and a transition to floating nuclear power ships in 2032; it also claims more than $4 billion in customer interest through letters of intent. The technical, regulatory and project-finance hurdles are substantial, especially for nuclear deployment at sea.
Dipole Labs is focused on networking inside AI data centers. It says its optical switching approach keeps data in the optical domain rather than repeatedly converting it between light and electricity. If the approach performs in production, it could address a costly issue for large GPU clusters: compute capacity sitting idle while data moves across the system.
Lamb Labs is taking aim at inference efficiency. The startup plans custom chips that hardcode model weights into silicon, reducing reliance on memory access. That could be valuable for stable, high-volume workloads, though it also raises a familiar hardware trade-off: specialized silicon can improve performance and energy use, but is less flexible when models or workloads change.
For operators, the important signal is that the AI buildout is creating opportunities well beyond model providers and cloud platforms. Power, cooling, interconnects, memory bandwidth and deployment timelines are all becoming strategic constraints.
Robotics needs data, control software and deployment economics
The rest of the investor-selected group shows how broad the robotics stack has become.
Praxis AI collects video and other real-world data from human work to help train robots. It says it has captured data across more than 150 environments and works with publicly traded companies. Its model reflects an emerging reality: useful automation depends on task-specific data, not only generalized models.
Waddle Labs is building an API layer intended to translate natural-language instructions into robot control code. The company describes its approach as using LLM agents to generate, test and configure executable controls across hardware. The proposition is attractive to developers because robot integration is still slow and specialized; the key test will be reliability and safety outside controlled demonstrations.
Two companies are pursuing physical deployment directly. Cosmic Robotics makes autonomous equipment for heavy lifting and says its technology is installing solar panels in the U.S., with $25 million in contracts through 2027. Nori is pursuing a lower-cost home robot for cleaning and folding clothes, priced at about $1,600; it says it had nearly $500,000 in sales within six weeks of launch. Both claims need to translate into sustained operation, support and unit economics.
Defense and frontier bets remain part of the mix
Isengard Industries aims to manufacture jet-powered strike and counter-drones locally in allied countries. It says it is already generating $10 million in revenue. Its presence on the list underscores continued investor interest in defense manufacturing capacity, particularly companies that claim they can move faster or manufacture closer to customers than traditional contractors.
The most speculative company is Parasma, which is exploring human brain cells as an energy-efficient compute substrate. It is a reminder that some venture attention remains reserved for long-horizon alternatives to conventional AI hardware, even when commercialization paths are unclear.
What to watch next
The near-term question is not whether these concepts are ambitious. It is whether they can clear the unglamorous gates: permitting, safety validation, manufacturing yield, integration, service and repeatable customer demand.
For founders, the cohort suggests that investors remain receptive to difficult technical problems when they map to a visible operational bottleneck. For enterprise buyers, it is a cue to separate compelling demos from products that can meet uptime, compliance and procurement requirements. The winners will be those that turn scarce compute and labor into measurable, dependable capacity.
