A batch that skewed deep tech

Y Combinator's Demo Day on Thursday looked different from recent cohorts. Far fewer software wrappers, far more founders building physical things.

How the list was built

TechCrunch asked early-stage VCs for their top picks and the deals everyone was gossiping about. Only startups named by at least two investors made the cut.

Wild ideas, saner prices

One investor said the tech felt like science fiction. The consensus was that valuations, unlike recent batches, were far more grounded.

“like science fiction”

Atomarine floats the data center

Nuclear-powered data centers on barges at sea, where seawater cools for almost nothing. A gas-powered pilot is planned for 2028, floating nuclear ships by 2032.

Four billion in interest

Atomarine says it has already signed letters of intent worth over $4 billion. That pipeline made it one of the highest-valued startups in the batch.

Dipole keeps data as light

GPU clusters idle while data converts from light to electricity and back, burning power and making heat. Dipole's optical switch skips the conversion entirely.

Isengard builds drones locally

Jet-powered strike and counter-drones mass-produced inside allied countries for a fraction of what US prime contractors charge. Already at $10 million in revenue.

Lamb Labs burns weights into silicon

Inference chips waste energy fetching model weights from memory. Lamb hardcodes the weights into the chip itself, calling the result a Model Processing Unit.

Praxis films humans working

It partners with businesses to record people doing real jobs, then sells that footage as robot training data. Already capturing in over 150 environments.

Nori sells a $1,600 humanoid

Six weeks old and nearly half a million in sales, for a robot that folds laundry and cleans. Rival humanoid Neo costs around $20,000.

Cosmic Robotics starts on Earth

Its heavy-lift robots are installing solar panels across the US under $25 million in contracts through 2027. The founders want the same tech building a city on Mars.

Parasma computes with brain cells

Another answer to the energy problem: train human brain cells to do the work today's AI hardware does, at a fraction of the power.

Waddle writes the robot code

Rather than training on video or teleoperation, Waddle uses LLM agents to generate control code. Plug in hardware, describe the task, and it sets up in 20 minutes.

“Claude Code for robotics”