Deep tech dominates YC's latest Demo Day, from floating nuclear data centers to brain cells

A deep-tech tilt at Y Combinator

Y Combinator's latest Demo Day, held this past Thursday, marked a departure from the accelerator's software-heavy tradition, according to TechCrunch. The presenting companies skewed toward hard science — nuclear engineering, optics, custom silicon, defense manufacturing and robotics — and when TechCrunch polled its regular panel of early-stage VCs, nine startups were flagged by at least two investors as the batch's most discussed deals. One investor described the technology on display as feeling "like science fiction," yet the same survey found valuations were noticeably more grounded than in recent cohorts.

The nine startups investors flagged

Atomarine wants to float data centers on ocean barges, using seawater for near-free cooling, with a gas-powered pilot planned for 2028 and floating nuclear power ships targeted for 2032. Backed by a founding team spanning computer science, naval engineering and nuclear engineering at MIT, the company claims more than $4 billion in customer interest via letters of intent — enough, one VC told TechCrunch, to make it one of the highest-valued startups in the batch.

Dipole Labs targets a quieter bottleneck: GPU clusters burning time while data converts between light and electricity inside the network layer. Its optical switch lets data stay in optical form and travel directly to its destination, cutting the power draw and heat that conversion generates.

Isengard Industries mass-produces jet-powered strike and counter-drones inside allied countries, at a fraction of what US prime contractors charge. Its founders include a former Australian Army officer and a defense entrepreneur who previously scaled a Ukraine-focused drone startup to $60 million in revenue. Isengard already generates $10 million in revenue and, according to two investors, commands one of the loftiest valuations in the batch.

Lamb Labs attacks inference energy costs by hardcoding AI model weights directly into silicon. Its "Model Processing Units" sidestep the memory-bandwidth bottleneck of fetching weights during inference; the founders come from AI research at Imperial College London and theoretical physics at Oxford.

Robotics accounts for four of the nine. Praxis AI partners with businesses to record humans doing real work — footage from more than 150 environments, including at publicly traded companies — and turns it into training data for robot builders. Nori, founded just six weeks before Demo Day, reports nearly half a million dollars in sales of a roughly $1,600 humanoid robot that cleans and folds clothes and can be operated from a laptop app, a price TechCrunch sets against the roughly $20,000 Neo. Cosmic Robotics builds autonomous machines for heavy lifting; its systems are already installing solar panels across the US under $25 million in contracts through 2027, with an exploratory Mars mission targeted for 2028 to stay aligned with SpaceX's timeline for building there. Waddle Labs, founded by Harvard graduates, exposes an API in which LLM agents write robot control code directly instead of training foundation models on video or teleoperation data. Positioning itself as "Claude Code for robotics," the company claims that plugging any hardware into the API lets a developer describe a task in natural language, with agents generating executable control code, verifying it and bringing the robot online in about 20 minutes.

Parasma sits furthest out: it is training human brain cells to serve as a more energy-efficient alternative to today's AI computing hardware.

Why it matters

A YC batch is a rough proxy for where early-stage capital believes the next decade of returns live, and this one points away from pure software toward energy, silicon, defense and the physical world. Notably, several of these companies attack the same underlying constraint — the power and data-movement costs of AI compute — from completely different directions, while robotics draws the widest single share of attention. The combination of ambitious hardware and saner valuations suggests investors are now funding deep tech with more discipline than in the frothier cohorts of recent years. For founders and engineers, the signal is clear: capital-intensive ideas once considered too slow or too hard for an accelerator are exactly what Demo Day crowds came to see.