Written by Admin Alex · Fact-Checked by M.Ali · Info Verified September 2026
We review and update this article regularly as new information becomes available.
TL;DR: Y Combinator’s newest Demo Day batch leaned heavily into “deep tech” rather than another wave of AI wrapper apps. Investors flagged nine standouts ranging from a startup building nuclear-powered data centers on floating barges to one experimenting with human brain cells as computing hardware. Valuations this round looked more grounded than recent cycles, even with a few eye-popping numbers.
Every Demo Day produces its share of forgettable pitches. This one didn’t feel like that. VCs who sat through the presentations kept using the same word afterward: science fiction. Not as a knock, but as a genuine description of what they were looking at.
Power and chips, the unsexy backbone of the AI boom
Automarine wants to put nuclear reactors on floating barges to power data centers, an idea that sounds absurd until you remember how badly the AI industry needs electricity it doesn’t currently have. The founding team combines MIT-trained computer science, naval engineering, and a nuclear physics PhD, and they’re not starting with reactors. A gas-powered pilot comes first, targeted for 2028, with the nuclear transition planned for 2032. The company says it already has more than $4 billion in customer interest through letters of intent, numbers that helped make it one of the highest-valued companies in the batch.
Dipole Labs is chasing a narrower but equally real problem: the massive energy cost of converting data from optical signals to electrical ones inside AI data centers. Their optical switches keep data as light the whole way through, which sounds like a small engineering detail until you consider how much heat and power gets wasted on that conversion at scale across a GPU cluster.
Lamb Labs is going after a similar inefficiency from a different angle, building what they call Model Processing Units, chips with model weights hardcoded directly into the silicon rather than pulled from memory on the fly. The founders, out of Imperial College London and Oxford, are betting that eliminating the memory-bandwidth bottleneck matters more than general-purpose flexibility for certain AI workloads.
Robots that actually generate revenue
A lot of robotics startups show up to Demo Day with a prototype and a dream. Isengard Industries showed up with $10 million in actual revenue from locally-producible jet-powered strike and counter-drones, co-founded by a former Australian Army officer. That kind of traction is rare enough at this stage that it landed the company among the batch’s loftiest valuations.
Nori took the opposite approach on price. Their humanoid robots handle household tasks and sell for around $1,600, compared to roughly $20,000 for competing hardware. In six weeks after launch, Nori pulled in nearly $500,000 in sales, proof that affordability might matter more than raw capability for a first wave of home robotics buyers.
Cosmic Robotics builds autonomous heavy-lifting machines for construction, already installing solar panels across the U.S. under $25 million in contracts through 2027. Their long-term pitch is even bigger: supporting Mars colonization, with an exploratory mission targeted for 2028. Ambitious doesn’t begin to cover it.
The genuinely strange ones
Praxis AI collects real-world training data for robotics companies, capturing footage and sensor data across more than 150 different environments, already working with publicly traded clients. Waddle Labs built an API layer that turns natural language commands into executable robot control code, pitching itself bluntly as “Claude Code for robotics,” claiming it can generate working code within 20 minutes of a plain-English request.
And then there’s Parasma, which is trying to use actual human brain cells as a computing substrate for AI, aiming for energy efficiency that silicon can’t match. It’s the kind of pitch that either sounds like a breakthrough or a fever dream depending on who you ask, and honestly, it might be both.
What this batch says about where VC money is heading
The shift toward deep tech isn’t random. After a couple of years dominated by thin AI wrapper startups chasing quick exits, investors seem hungry for companies solving problems that require years of technical groundwork and actual physical infrastructure. That’s a slower, riskier bet, but it’s also harder to copy overnight, which matters a lot when your competitive moat used to be “we called OpenAI’s API first.”
Bottom Line: This Demo Day batch bet on hard problems instead of easy ones, and investors rewarded that with real interest rather than just hype. Whether floating nuclear reactors or brain-cell computers actually ship is a different question entirely, but the appetite for trying has clearly shifted.
