Investors are shifting attention to founders tackling compute, power and heavy automation after Y Combinator’s summer Demo Day, a cohort that skewed heavily toward deep tech. The immediate consequence is a funnel of early-stage capital and scrutiny toward companies promising cheaper power, more efficient inference and practical robots for industrial and consumer use.
Early-stage VCs singled out nine startups as the buzziest in the batch, with several firms named by multiple investors. One investor described the technologies on show as "like science fiction," even as others noted that valuations felt more grounded than in recent cohorts.
Atomarine proposes to put data centers on barges and eventually on nuclear-powered ships, using seawater for near-free cooling to address local power constraints and community resistance to land-based facilities. The company plans a gas-powered pilot by 2028 and a transition to floating nuclear power ships in 2032, and it says it has secured over $4 billion in customer interest through letters of intent. Those commitments have helped make Atomarine one of the highest-valued startups in the batch, according to investors.
Dipole Labs is building optical networking hardware for AI data centers, an optical switch designed to keep data in light form and avoid the energy and heat costs of converting between light and electricity. The startup frames the work as a way to reduce wasted GPU time that results when chips wait for data to move between them.
Isengard aims to mass-produce jet-powered strike and counter-drones inside allied countries, undercutting prime contractors on price. Co-founded by a former Australian Army officer and a defense entrepreneur who previously scaled a Ukraine-focused drone startup to $60 million in revenue, Isengard is already generating $10 million in revenue itself and drew strong VC buzz, including one of the loftier valuations in the cohort.
Lamb Labs plans custom inference chips that hardcode AI model weights into silicon, a concept it calls Model Processing Units, which the founders say will eliminate memory-bandwidth bottlenecks that drive energy use during inference. The team includes an Imperial College London AI Ph.D. and an Oxford theoretical physicist.
Other flagged companies include a startup that collects video data of humans working in more than 150 environments to train robots, a humanoid maker that launched six weeks ago and claims almost half a million in sales while pricing its robot at around $1,600 compared with Neo at about $20,000, and a business building autonomous heavy-lift robots that says it is installing solar panels across the U.S. and has $25 million in contracts through 2027 while aiming for an exploratory mission by 2028.
Two more buzzy ideas drew attention: Parasma, which is exploring training human brain cells to power compute as a more energy-efficient substrate, and a startup building an API layer that writes robot control code, a move investors hope could be the ChatGPT moment for robotics. Together these companies illustrate why VCs are testing demand for pragmatic hardware and infrastructure bets rather than pure software plays.
The next test for the cohort is execution. Several startups outlined concrete near-term milestones, pilot launches, contracted work and multi-year deals, so investors will be watching whether those commitments convert into repeatable revenue and deployed systems.
