Y Combinator's S26 Demo Day on 13 September presented a cohort that investors consistently described as "skewing far more toward deep tech" than recent batches, with technology characterised as "like science fiction" in investor briefings. Five companies drew the strongest VC attention. Lamb Labs: custom AI inference chips with hardcoded model weights — the architecture eliminates the separation between model and silicon, improving inference efficiency by training the chip specifically for a given model; co-founded by an Imperial College AI PhD and an Oxford theoretical physicist. Parasma: replaces GPU cluster compute with human brain cell cultures as a more energy-efficient substrate for certain AI workloads — an approach with obvious regulatory and scaling questions but potentially transformative power-per-operation economics. Waddle Labs: an API layer that generates robot control code via LLM agents, positioned explicitly as "Claude Code for robotics" — Harvard founders targeting the gap between LLM reasoning capability and physical robot execution. Automarine: nuclear-powered floating data centres claiming more than $4 billion in letters of intent; the floating form factor addresses land acquisition and cooling constraints simultaneously. Dipole Labs: high-speed optical networking specifically designed for AI data centre interconnects, addressing the bandwidth bottleneck between GPU clusters that limits distributed training. The shift from application-layer software to hardware and infrastructure at the seed stage is consistent across the batch: founders and institutional investors are now betting that durable AI advantages will be built at the hardware layer — chips, power, networking, physical execution — not at the prompt or fine-tuning layer.