Y Combinator Demo Day Highlights 9 Deep Tech Startups Praised by VCs

During Thursday’s Y Combinator Demo Day, a noticeable shift toward deep technology was observed compared to previous cohorts. TechCrunch surveyed early-stage venture capitalists to identify the most prominent startups in this batch, resulting in a list of nine companies highlighted by at least two investors. While one investor described the technology on display as resembling “science fiction,” there was a general consensus that startup valuations were more conservative than in recent quarters.

Atomarine is addressing energy shortages and local opposition to new data centers by developing floating nuclear-powered facilities. Co-founded by experts in computer science, naval engineering, and nuclear physics from MIT, the company plans to deploy gas-powered pilot barges by 2028 before transitioning to nuclear vessels in 2032. Utilizing seawater for cooling, Atomarine reports over $4 billion in customer interest via letters of intent, making it one of the highest-valued startups in the current batch.

Dipole Labs is tackling efficiency bottlenecks in AI data centers with high-speed optical networking hardware. The startup claims its optical switches eliminate the need to convert data between light and electricity, a process that typically consumes significant power and generates heat. This innovation aims to maximize the utility of expensive GPU clusters by reducing idle time caused by data transfer delays.

Isengard Industries focuses on the mass production of jet-powered strike and counter-drone systems within allied nations, aiming to undercut the costs of prime U.S. contractors. Backed by a former Australian Army officer and a defense entrepreneur who previously scaled a Ukraine-focused drone company to $60 million in revenue, Isengard is already generating $10 million in sales and has attracted top-tier valuations from investors.

Lamb Labs is developing custom chips designed to drastically reduce energy consumption during AI inference. By hardcoding AI model weights directly into silicon, the company’s “Model Processing Units” (MPUs) aim to bypass memory-bandwidth bottlenecks. The startup was founded by an AI PhD from Imperial College London and a theoretical physicist from Oxford.

Praxis AI acts as a data provider for robotics companies, partnering with businesses to capture video and operational data from human workers. Having recorded footage in over 150 distinct environments, including engagements with publicly traded companies, Praxis AI provides the training materials necessary for developers to refine robotic automation as industries determine which tasks are best suited for machines versus humans.

Nori, launched just six weeks prior to the Demo Day, has already secured nearly $500,000 in sales. The company offers an affordable humanoid robot priced at approximately $1,600, designed to perform household chores such as folding clothes and loading dishwashers. It can be controlled via a laptop application, offering a budget-friendly alternative to competitors like Neo, which carries a price tag around $20,000.

Cosmic Robotics is building autonomous heavy-lifting robots with the long-term goal of supporting construction on Mars. The technology is currently being used to install solar panels across the United States, with $25 million in contracts secured through 2027. The founders intend to align their development timeline with SpaceX’s Mars colonization efforts, hoping to launch an exploratory mission by 2028.

Parasma is exploring biohybrid computing by training human brain cells to serve as a more energy-efficient alternative to traditional AI hardware. The startup aims to address the immense power demands of running large AI models by utilizing biological neural networks for computation.

Waddle Labs is creating an API layer that allows developers to control robots using natural language. Founded by Harvard graduates, the company positions itself as “Claude Code for robotics,” using LLM agents to autonomously write, verify, and execute control code. This approach enables developers to program robots in approximately 20 minutes without needing to train foundation models on raw video or teleoperation data.