Hardware Startups We’d Actually Fund

Lightmatter uses photonics to move data between AI chips, attacking the bandwidth and power bottlenecks that now limit large computing systems.

Cerebras builds wafer-scale processors instead of slicing silicon into ordinary chips. The engineering is difficult, distinctive and attached to real AI workloads.

Etched is designing chips specifically for transformer models, sacrificing flexibility for speed and efficiency. We like the focus, even if architectures move.

Celestial AI uses photonics to connect memory and processors, addressing a costly bottleneck that software alone cannot remove from AI systems.

Gecko’s climbing robots inspect boilers, tanks and industrial infrastructure while producing data operators already know how to use and value.

Generalist trains software intended to work across different robots. Selling intelligence rather than one machine keeps its potential customer base unusually broad.

Ayar Labs puts optical connections directly into chip packages. Moving data with light may matter as much as improving the processors themselves.

Groq designed its inference chips and compiler together, producing speed users can feel immediately. That clear advantage gives the hardware a reason to exist.

Machina Labs uses robots to form sheet metal without dedicated tooling, which could make small production runs faster and far less expensive.

Figure is placing humanoid robots in factories before attempting the home. Real production work gives the company a useful test beyond demonstrations.