Y Combinator Demo Day on Thursday showcased a batch that leaned heavily into deep‑tech concepts, prompting early‑stage venture firms to single out nine startups they consider the most talked‑about.
Floating nuclear data centers aim to solve compute shortage
Atomarine proposes placing data centers on barges at sea, using seawater for near‑free cooling and eventually switching to floating nuclear reactors. The co‑founders include an MIT computer‑science engineer, a naval engineer, and a nuclear‑engineering Ph.D.
The firm says a gas‑powered pilot will debut by 2028, with a nuclear‑powered ship targeted for 2032. It claims over $4 billion in customer interest via letters of intent, a figure that helped push its valuation among the highest in the cohort, according to a venture partner.
Nuclear power offers the promise of abundant energy, though the plan sounds almost like a sci‑fi movie, yet it’s grounded in contracts.
Optical switches promise faster AI networking
Dipole Labs is building high‑speed optical networking hardware designed for AI data centers. Its switch keeps data in light form, avoiding the electrical‑to‑light conversion that typically wastes power and adds heat.
By eliminating that step, the startup hopes to cut latency and reduce electricity use, a timely improvement as GPU costs climb and operators seek to squeeze every ounce of performance from their clusters.
Jet‑powered drones target defense markets
Isengard focuses on mass‑producing jet‑powered strike and counter‑drones within allied nations, aiming to lower costs compared with U.S. prime contractors. One co‑founder previously grew a Ukraine‑focused drone firm to $60 million in revenue.
The company reports $10 million in sales already and has attracted strong VC attention, earning one of the loftiest valuations in this YC batch, according to two investors.
Hard‑coded AI chips cut inference power
Lamb Labs plans to embed AI model weights directly into silicon, creating “Model Processing Units” that avoid the memory‑bandwidth bottlenecks of conventional inference chips. The founders include an Imperial College London AI Ph.D. and an Oxford theoretical physicist.
By hard‑coding weights, they claim the chips will consume far less energy during inference, a benefit for data centers battling rising electricity bills.
Robot training data collected from real workplaces
A yet‑unnamed startup partners with businesses to capture video of humans performing tasks, then packages that footage as training material for robot developers. It says it has recorded data in more than 150 environments.
Read Also: Mecka AI secures funding near $500M valuation
The firm already works with publicly traded companies, positioning itself as a potential linchpin as firms decide which jobs remain human‑centric and which shift to automation.
Affordable home robot tackles laundry
Another newcomer launched six weeks ago with a humanoid robot that can clean and fold clothes. Priced at roughly $1,600, it undercuts competitors like the $20,000 Neo model.
Early sales approach half a million dollars, suggesting consumer appetite for low‑cost assistants that can handle everyday chores, a market segment long considered out of reach.
Heavy‑lifting robots built for Mars construction
A startup aiming to support a future Martian city is developing autonomous robots capable of moving large objects. Its technology is already being used to install solar panels across the United States.
The venture says it holds $25 million in contracts through 2027 and hopes to align its timeline with SpaceX’s plans for a 2028 exploratory mission.
It targets a future on Mars.
Brain‑cell computing explores new AI power source
Parasma is investigating whether cultured human brain cells could serve as a low‑energy alternative to traditional AI hardware. The approach seeks to address the growing power demands of large models.
While still experimental, the effort reflects a broader search for biologically inspired computing methods that might one day complement silicon‑based systems.
LLM‑driven API writes robot control code
Waddle Labs offers an API that translates natural‑language commands into executable robot code using large language model agents. Founded by Harvard graduates, the startup describes its service as “Claude Code for robotics.”
According to the company, developers can connect any hardware to the API, describe a task in plain English, and receive working control code within about 20 minutes, potentially accelerating robot deployment.
These nine ventures illustrate a shift toward ambitious engineering solutions that blend cutting‑edge research with commercial traction. Investors appear drawn to projects that promise tangible revenue streams—whether through letters of intent, early sales, or sizable contracts—while also pushing the boundaries of what technology can achieve.
