Deep Tech and Floating Nuclear Data Centers Dominate Y Combinator Demo Day

Y Combinator's latest Demo Day showcased a distinct pivot toward deep tech, featuring ventures tackling power shortages, optical networking bottlenecks, and advanced robotics with valuations described by investors as far more grounded than previous cohorts. While the cohort featured science-fiction concepts ranging from biological computing to Martian infrastructure, several early-stage startups emerged as the buzziest picks among venture capitalists due to substantial customer interest and early revenue generation.

Among the highest-valued and most discussed projects in the batch is Atomarine, which plans to deploy nuclear-powered data centers on ocean barges where seawater provides near-free cooling. Co-founded by MIT alumni, the startup aims to launch a gas-powered pilot by 2028 before shifting to floating nuclear power ships in 2032. According to venture capital investors participating in TechCrunch's quarterly survey, Atomarine has already secured over $4 billion in customer interest through letters of intent, driving significant valuation interest.

Addressing the massive power consumption and latency challenges plaguing modern artificial intelligence infrastructure, several hardware startups drew intense investor attention for redesigning chip-to-chip communication and inference processing. Dipole Labs developed an optical switch designed to keep data in the form of light across the networking layer, bypassing the traditional power-intensive conversion between light and electricity that generates significant heat in GPU clusters.

At the silicon level, Lamb Labs is tackling the energy consumed by fetching model weights from external memory during inference. Co-founded by an Imperial College London AI PhD and an Oxford theoretical physicist, the startup is developing Model Processing Units that hardcode AI model weights directly into silicon. This architectural approach aims to eliminate memory-bandwidth bottlenecks that drive up operational costs for large-scale AI deployments.

Defense technology and industrial robotics also commanded significant attention from investors evaluating early-stage traction and sovereign production capabilities. Isengard, co-founded by a former Australian Army officer and a defense entrepreneur, is mass-producing jet-powered strike and counter-drones within allied nations. The company has already reached $10 million in revenue, positioning itself as a lower-cost alternative to prime defense contractors.

In the robotics sector, Waddle Labs introduced an API layer that utilizes large language model agents to write control code directly for hardware. Billed as a developer tool for robotics, the system allows engineers to connect hardware to an API and translate natural language instructions into executable control code within twenty minutes. Meanwhile, consumer-facing humanoid robotics developer Nori launched a domestic robot priced at approximately $1,600, reporting nearly $500,000 in sales within six weeks of its debut.

Pushing the boundaries of compute architecture and long-term infrastructure, other cohort participants targeted biological systems and extraterrestrial engineering. Parasma began exploring the use of cultivated human brain cells as an energy-efficient alternative to conventional silicon computing hardware for running complex workloads.

On the heavy-machinery front, an unnamed Mars-focused robotics startup is building autonomous machinery designed to perform heavy-duty construction tasks. The firm reported $25 million in contracts through 2027 and stated its technology is already being used to install solar panels across the United States as a proving ground for future planetary exploration.

Atomarine plans to launch its initial gas-powered marine data center pilot by 2028, with the full transition to floating nuclear power ships scheduled for 2032.