One investor described the tech at Thursday’s Y Combinator Demo Day as feeling “like science fiction.” Then the same crowd told me the valuations were more grounded than they’ve been in recent batches. Both things were true at once, which doesn’t happen often.
This cohort skewed hard toward deep tech. Not another CRM wrapper, not another vertical SaaS play for dentists. Floating nuclear data centers. Chips with model weights burned into silicon. Human brain cells doing compute.
As we do every quarter, we asked early-stage VCs which startups they were chasing and which ones the whole room was whispering about. What follows are the nine that at least two separate investors flagged, listed alphabetically.
The one with $4 billion in letters of intent and no reactor yet
Atomarine wants to put data centers on barges at sea, powered by nuclear reactors. Its co-founders are an MIT computer science and naval engineer and an MIT PhD in nuclear engineering.
The pitch makes uncomfortable sense. Power is scarce, and local communities keep fighting new data center construction. Park the thing offshore and seawater handles cooling for close to nothing.
The timeline is where you should pay attention. A gas-powered pilot is planned for 2028, with a transition to floating nuclear power ships in 2032. That’s a long road, and Atomarine claims it has already secured over $4 billion in customer interest through letters of intent. Letters of intent aren’t revenue. But that potential helped make it one of the highest-valued startups in the batch, one VC told me.
Light that never becomes electricity
Dipole Labs is going after a problem anyone who’s watched a GPU cluster idle already knows about. Chips spend real time waiting on data to move between them.
Inside the networking layer, data gets converted from light to electricity and back again. That conversion burns power and throws off heat. Dipole says it built an optical switch that skips the step entirely, keeping data as light and routing it straight to where it’s going.
Timing works in its favor. GPUs cost a fortune, and nobody running a data center wants to pay for compute that sits there waiting.
Jet drones built abroad, not in the U.S.
Isengard Industries plans to mass-produce jet-powered strike and counter-drones inside allied countries, at a fraction of what prime contractors charge to build them domestically.
The founding team has receipts. One is a former Australian Army officer. The other is a defense entrepreneur who previously scaled another Ukraine-focused drone startup to $60 million in revenue.
Isengard itself is already generating $10 million in revenue, which puts it ahead of most of this list on the only metric that’s hard to fake. Two investors said it carries one of the loftiest valuations in the batch.
Chips that don’t go looking for weights
Lamb Labs is building custom inference chips with AI model weights hardcoded directly into silicon. It calls them Model Processing Units, or MPUs.
The argument is about memory. Conventional AI chips burn enormous energy during inference just fetching weights from memory. Hardcode the weights and the memory-bandwidth bottleneck goes away.
Co-founded by an Imperial College London AI PhD and an Oxford theoretical physicist. The obvious tradeoff sits right in the name: a chip built around one set of weights is a chip built around one set of weights.
Somebody has to film the humans first
Praxis AI partners with businesses to record video and data of people doing actual work, then converts that into training material for robotics companies.
It says it’s already working with publicly traded companies and has captured video data in more than 150 different environments. That number matters more than it sounds like it should, because the variety of settings is what separates usable robot training data from a warehouse demo reel.
The business gets more interesting as companies start sorting out which tasks humans should keep and which ones belong to a machine.
A $1,600 humanoid against a $20,000 one
Nori launched six weeks ago and already claims almost half a million in sales. It’s a humanoid robot that promises to clean and fold clothes, and you can operate it from a laptop app.
Price is the whole story here. Nori runs around $1,600. Neo, one of the humanoids it’s up against, sits around $20,000. That’s not a discount, that’s a different product category.
Whether it works is the open question, and it’s the same question that’s dogged home robotics for a decade. Can anyone build an affordable robot that genuinely stacks a dishwasher? Nori is another swing at it.
Solar panels now, Mars later
Cosmic Robotics builds autonomous robots that lift heavy objects. The founders want to build a city on Mars, and heavy-duty robotics is step one.
Here’s the part that keeps it honest: the company says its tech is already installing solar panels across the U.S., with $25 million in contracts through 2027. Terrestrial revenue funding an interplanetary thesis.
It’s racing SpaceX’s Mars timeline and hopes to begin an exploratory mission by 2028.
Brain cells as a power strategy
Parasma is training human brain cells to one day power compute. Same underlying problem as Lamb Labs and Dipole Labs, wildly different answer.
The bet is that human brain cells could serve as a more energy-efficient alternative to today’s AI computing hardware. This one is the furthest from anything you could buy, and the “one day” in the description is doing real work.
Claude Code, but for robot arms
Waddle Labs is skipping the foundation model approach entirely. Rather than training on raw video or human teleoperation data, it uses a layer of LLM agents that write code and control robots directly through an API.
The founders, Harvard graduates, are positioning the company as “Claude Code for robotics.” The claim is that you plug any hardware into Waddle’s API, tell the robot what to do in natural language, and its agents generate executable control code, verify it worked and set the robot up in about 20 minutes.
Twenty minutes is a specific enough number that it’s testable, which is more than you can say for most robotics pitches.
What the list tells you
Five of the nine are fighting over the same bottleneck: the energy and hardware cost of running AI models. Optical switching, hardcoded silicon, offshore nuclear, brain cells. Four different bets, one shared assumption that compute stays scarce and expensive.
If you’re an investor looking at this batch, the ones with revenue already on the board are Isengard at $10 million and Cosmic Robotics with $25 million in contracts through 2027. Everything else is a timeline and a thesis.
