“Like science fiction” was how one investor summed up the technology on display at Thursday’s Y Combinator Demo Day. The same room then turned around and told me valuations looked saner than they have in several recent batches. Rarely are both of those true in the same afternoon.
Deep tech dominated this cohort. No CRM wrappers, no vertical SaaS built for dental offices. Instead: data centers floating on the ocean, silicon with model weights baked in, and compute running on human brain cells.
Following our quarterly routine, we canvassed early-stage VCs on which companies they were pursuing and which names the room kept murmuring about. Below are the nine that drew a flag from at least two separate investors, in alphabetical order.
The one with $4 billion in letters of intent and no reactor yet
Atomarine’s plan is to mount data centers on seagoing barges and run them on nuclear reactors. The co-founding pair consists of an MIT computer science and naval engineer alongside an MIT PhD in nuclear engineering.
Uncomfortably, the logic holds up. Power supply is tight, and new data center projects keep running into local opposition. Move the facility offshore and cooling comes courtesy of seawater at almost no cost.
Watch the schedule, though. The company intends to run a gas-powered pilot in 2028 before shifting to floating nuclear power ships in 2032. That is a distant horizon, and Atomarine says it has already locked in more than $4 billion in customer interest via letters of intent. Letters of intent are not revenue. Even so, one VC told me that upside helped push it to among the highest valuations in the batch.
Light that never becomes electricity
Dipole Labs is attacking something familiar to anyone who has watched a GPU cluster sit idle: chips lose meaningful time waiting for data to travel between them.
Within the networking layer, information is translated from light into electricity and then back. Each conversion consumes power and generates heat. Dipole says its optical switch cuts that step out altogether, leaving data as light and steering it directly to its destination.
The moment suits the pitch. GPUs are extraordinarily expensive, and no data center operator wants to pay for compute that is simply waiting around.
Jet drones built abroad, not in the U.S.
Isengard Industries intends to mass-manufacture jet-powered strike drones and counter-drones within allied nations, at a sliver of what prime contractors charge for domestic production.
Its founders bring a track record. One served as an officer in the Australian Army. The other is a defense entrepreneur who had previously grown a separate Ukraine-focused drone company to $60 million in revenue.
Isengard is already pulling in $10 million in revenue itself, putting it ahead of most of this list on the one metric that resists spin. Two investors described its valuation as among the loftiest in the batch.
Chips that don’t go looking for weights
Lamb Labs is designing bespoke inference chips that hardcode AI model weights straight into the silicon. The company brands them Model Processing Units, or MPUs.
Memory is the crux of the pitch. Standard AI chips expend enormous amounts of energy during inference purely on retrieving weights from memory. Bake the weights in and the memory-bandwidth bottleneck disappears.
The co-founders are an Imperial College London AI PhD and an Oxford theoretical physicist. The catch is right there in the concept: a chip designed around a single set of weights is a chip designed around a single set of weights.
Somebody has to film the humans first
Praxis AI teams up with businesses to capture video and data of people performing real work, then turns that footage into training material for robotics firms.
According to the company, it already counts publicly traded businesses among its partners and has gathered video data across more than 150 distinct environments. That figure carries more weight than it appears to, since the range of settings is precisely what distinguishes genuinely useful robot training data from a warehouse demo reel.
The opportunity grows as companies begin deciding which tasks stay with people and which get handed to machines.
A $1,600 humanoid against a $20,000 one
Six weeks after launching, Nori says it has already booked nearly half a million in sales. The product is a humanoid robot pitched on cleaning and folding laundry, controllable through a laptop app.
Everything here comes down to price. Nori goes for roughly $1,600. Neo, among its humanoid competitors, lands near $20,000. That gap is not a discount so much as a separate product category.
Does it actually work? That remains unanswered, and it is the very question that has haunted home robotics for ten years. Is an affordable robot that can truly load a dishwasher buildable? Nori is taking another swing.
Solar panels now, Mars later
Cosmic Robotics makes autonomous robots designed for heavy lifting. Its founders are aiming at a city on Mars, and rugged robotics is the opening move.
What grounds the story: the company says its technology is already deploying solar panels around the U.S., backed by $25 million in contracts running through 2027. Earthbound revenue paying for an interplanetary thesis.
It is competing against SpaceX’s Mars schedule and aims to launch an exploratory mission by 2028.
Brain cells as a power strategy
Parasma is cultivating human brain cells with the goal of eventually powering compute with them. It is the same core problem Lamb Labs and Dipole Labs are chasing, approached from a radically different direction.
The wager is that human brain cells might prove a more energy-efficient substitute for the AI computing hardware in use today. Of everything on this list, it sits furthest from any purchasable product, and the phrase “one day” is carrying a lot of weight.
Claude Code, but for robot arms
Waddle Labs has abandoned the foundation model route altogether. In place of training on raw video or human teleoperation data, it deploys a layer of LLM agents that write code and drive robots directly via an API.
Its Harvard-educated founders describe the company as “Claude Code for robotics.” Their claim: connect any hardware to Waddle’s API, describe the task to the robot in plain language, and the agents produce executable control code, confirm it ran correctly and have the robot configured in roughly 20 minutes.
Twenty minutes is a concrete enough figure to be tested, which is rarer than it should be among robotics pitches.
What the list tells you
Five of these nine are competing over one bottleneck: what it costs in energy and hardware to run AI models. Optical switching, hardcoded silicon, nuclear power at sea, brain cells. Four distinct wagers resting on a single shared assumption — that compute remains scarce and pricey.
For an investor sizing up this batch, the companies with revenue already booked are Isengard at $10 million and Cosmic Robotics with $25 million in contracts through 2027. The rest is a schedule and a thesis.