Five AI startups land funding for recycling, health, mining, fuel, and construction

This month's column captures a significant inflection point in AI investment: the technology is increasingly moving out of software labs and into physical environments where it automates complex, dangerous, or labor-intensive work. Greyparrot's £20.3 million ($27 million) Series B exemplifies this shift—the London-based startup has transformed recycling from a sampling-based guessing game into a data-driven operation by mounting computer vision systems on conveyor belts. Its customers include major waste-processing firms like Waste Management and Veolia, and its data platform Deepnest now serves consumer brands including Unilever, L'Oréal, and Kenvue seeking to redesign packaging and comply with Extended Producer Responsibility rules.

The broader market context reinforces the trend. Physical AI startups raised nearly $47.3 billion in the first half of 2026—up nearly 80% year over year—and robotics funding overall hit $15 billion in 2025, a record that has already been surpassed partway through 2026. This capital surge reflects industries' urgent need to address labor shortages and dangerous working conditions. Emesent's $17 million raise (combining $10 million equity and $7 million venture debt) targets an undersupplied niche: autonomous navigation in GPS-denied environments like mines and tunnels, where the company's technology is already deployed at more than 200 mine sites worldwide.

Even narrower verticals are attracting capital. SoundHealth's $12.25 million Series A sits at the intersection of AI and regulated medical devices—a category that raised more than $629 million in the first half of 2026, up about 33% year over year. Meanwhile, Cascade's $3.5 million seed round and Forge Industries' $3.85 million round show that capital is also flowing to less capital-intensive applications: construction bid prediction and waste-to-fuel conversion. Together, these five deals illustrate how AI's economic value is increasingly realized not in language models or image generators, but in domain-specific automation where regulatory approval, hardware integration, or domain expertise create defensible positions.