Another Y Combinator Demo Day took place on Thursday, but this batch looked a little different. While every cohort brings a fresh wave of founders, the startups presenting this week skewed far more toward deep tech than in past years.
As we do every quarter, TechCrunch asked early-stage VCs to name the hottest startups in this batch — both their top picks and the deals everyone was talking about. And as usual, we put together a list of startups flagged by at least two investors as the buzziest in the batch.
The tech in this batch felt “like science fiction,” as one investor put it. But beyond the wild ideas, the consensus was that valuations were far more grounded than in recent cohorts.
We’ve listed the top picks alphabetically below:
What it’s building: Nuclear-powered data centers floating at sea
Why it’s a fav: Power is in short supply and local communities are increasingly opposing the buildout of new data centers. Atomarine — co-founded by an MIT computer science and naval engineer and an MIT PhD in nuclear engineering — wants to solve the compute shortage by putting data centers on barges at sea, where seawater could provide near-free cooling. The startup plans to launch a gas-powered pilot by 2028 and plans to transition to floating nuclear power ships in 2032. Atomarine claims it has already secured over $4 billion in customer interest through letters of intent. That revenue potential has helped make it one of the highest-valued startups in the batch, one VC told TechCrunch.
What it’s building: Effective and energy-efficient high-speed optical networking hardware for AI data centers
Why it’s a fav: The idea is that GPU clusters waste a lot of compute time simply waiting for data to move between each chip. Within the networking layer, data is converted from light to electricity and back again in a process that burns lots of power and yields lots of heat. But Dipole Labs says it built an optical switch that skips the conversion so the data can stay as light and go directly where it needs to go. It’s a timely problem to tackle as GPUs are crazy expensive and data centers want the most out of their compute hardware without leaving any time to spare.
What it’s building: Locally producible, jet-powered strike and counter-drones
Why it’s a fav: Isengard aims to mass-produce jet-powered attack and counter-drones directly within allied countries, at a fraction of what prime contractors charge to build them in the U.S. Co-founded by a former Australian Army officer and a defense entrepreneur who previously scaled another Ukraine-focused drone startup to $60 million in revenue, Isengard is already generating $10 million in revenue itself. The startup has generated strong VC buzz, garnering one of the loftiest valuations in this YC batch, according to two investors.
What it’s building: Custom inference chips with hardcoded AI model weights
Why it’s a fav: Traditional AI chips burn massive amounts of energy during inference, fetching model weights from memory. Co-founded by an Imperial College London AI Ph.D. and an Oxford theoretical physicist, Lamb Labs wants to build ultra-efficient chips by hardcoding AI model weights directly into silicon. Dubbed “Model Processing Units” (MPUs), these custom chips eliminate memory-bandwidth bottlenecks.
What it’s building: Collecting real-world data on which to train robots
Why it’s a fav: This company partners with businesses to collect videos and data of humans doing work, which it then turns into training material for companies building robots. It says it’s already working with publicly traded companies and has captured video data in more than 150 different environments. This company could become important as businesses start experimenting more with which tasks are best for humans — and, as AI continues to refine itself, which tasks are better left to a robot.
What it’s building: Creating affordable robots to handle everyday tasks
Why it’s a fav: Launched just six weeks ago, this company already touts almost half a million in sales. That’s not shocking — it’s a humanoid robot that is promising to help clean and fold clothes. Users can also operate their robot via a laptop app. It’s priced at around $1,600, which is a steal compared to other humanoids like Neo, priced around $20,000. One of the biggest questions in robotics is if it is possible to create an affordable at-home robot that can actually stack a dishwasher and do everyday home tasks. Nori is another swing at taking on the task.
What it’s building: Autonomous robots that can lift heavy objects
Why it’s a fav: The founders of this company want to build a city on Mars. The first step to that? Building robotic technology that performs heavy-duty tasks. It says its tech is already installing solar panels across the U.S. and that it has $25 million in contracts through 2027. The vision is that this technology can help automate construction for colonizing Mars. The startup is racing with SpaceX’s timeline for building on Mars, as it hopes to begin an exploratory mission by 2028.
What it’s building: Training human brain cells to one day power compute
Why it’s a fav: Another company trying to find the best way to tackle how much power and energy is needed to run models. This time, Parasma is looking at using human brain cells as an effective, more energy-efficient alternative to today’s AI computing hardware.
What it’s building: An API layer that writes robot control code
Why it’s a fav: Investors and tech enthusiasts are hoping that the ChatGPT moment for robots is around the corner. That excitement is driving different approaches to creating a general robotics model. Instead of training foundation models on raw video or human teleoperation data, Waddle Labs uses a layer of LLM agents to write code and control robots directly. The startup, founded by Harvard graduates, is positioning itself as “Claude Code for robotics.” The startup claims that by plugging any hardware into Waddle’s API, developers will be able to tell the robot what to do in natural language, and its AI agents will autonomously generate executable control code, check that it worked, and set up the robot in about 20 minutes.
