Deep Tech and Hardware Take Center Stage at Y Combinator's Latest Demo Day

The latest Y Combinator Demo Day concluded on Thursday with a distinct shift in direction, as participating startups leaned heavily into deep tech, hardware, and advanced science concepts. While recent cohorts focused heavily on software wrappers and conventional artificial intelligence applications, early-stage investors surveyed by TechCrunch Venture noted that the newest batch featured projects resembling science fiction. Despite the ambitious scope of the technologies on display, venture capitalists reported that financial valuations across the cohort remained more grounded than in recent years.

Among the startups commanding the highest valuations in the batch was Atomarine, a venture aiming to address severe electrical power shortages and local pushback against land-based data infrastructure. Founded by two Massachusetts Institute of Technology alumni—one holding a doctorate in nuclear engineering and the other specializing in naval engineering and computer science—the company designs ocean-floating data centers cooled directly by seawater. Atomarine plans to deploy a gas-powered pilot vessel by 2028 before transitioning to floating nuclear power platforms in 2032. The company claims to have collected over $4 billion in customer interest via letters of intent.

Defense technology also stood out as a high-valuation category with Isengard, a startup manufacturing jet-powered attack and counter-drones for allied nations. Co-founded by a former officer in the Australian Army alongside a defense entrepreneur who previously scaled a Ukraine-focused drone startup to $60 million in revenue, Isengard is already generating $10 million in revenue. The company aims to produce hardware locally at a small fraction of the prices charged by traditional defense contractors in the United States.

To resolve energy and latency bottlenecks in modern artificial intelligence infrastructure, several founders presented novel chip and networking architectures. Dipole Labs introduced an optical switch for AI data centers that routes signals purely as light, eliminating the traditional conversions between light and electrical current that generate excess heat and waste graphics processing unit cycles. Meanwhile, Lamb Labs—launched by an Oxford theoretical physicist and an Imperial College London AI doctorate graduate—is designing custom silicon known as Model Processing Units. These chips hardcode AI model weights into the silicon itself, bypassing traditional memory bandwidth limits during inference.

Biological computing and specialized software frameworks also emerged as key themes. Startup Parasma is investigating the use of cultivated human brain tissue as an energy-efficient alternative to conventional silicon hardware for running AI workloads. Taking a software-first approach to physical automation, Waddle Labs introduced an API platform founded by Harvard graduates that uses large language model agents to author autonomous control code. Functioning as a development suite for robotics, the system generates executable code, runs verification checks, and configures connected hardware within 20 minutes using natural language commands.

In physical robotics, consumer and industrial applications showed early commercial momentum. Six-week-old Nori reported generating roughly $500,000 in early sales for its $1,600 humanoid home assistant, which folds laundry and cleans spaces under user guidance via a laptop application. On the industrial side, a separate heavy-lifting robotics firm revealed $25 million in signed contracts through 2027 for installing domestic solar panels. That company aims to adapt its heavy autonomous equipment for space exploration, targeting an uncrewed mission to Mars by 2028 in line with timelines outlined by SpaceX.

Rounding out the cohort's robotics push is a startup focused on training data, which has captured video across more than 150 distinct operational environments. By contracting with enterprise clients and publicly traded corporations to record human workers, the venture turns real-world video footage into training datasets for autonomous machines. According to the TechCrunch Venture report, the surge in hardware and deep tech ideas reflects a growing investor interest in solving foundational compute, energy, and physical automation constraints.