NVIDIA highlights AI sorting systems aimed at scaling recycling and cutting landfill waste

With climate and resource pressure mounting, recycling remains stuck at the first step: identifying what is worth saving out of mixed waste streams fast enough to be practical. Ahead of COP31, NVIDIA used a sustainability-focused blog post to lay out how AI vision and robotics are being deployed at recycling plants to recover materials that would otherwise be landfilled, and to make the economics of reprocessed inputs more competitive with virgin materials.

The post frames the challenge in broad terms. NVIDIA says less than 7% of the materials entering the global economy come from recycled or reused sources, while municipal solid waste generation is already over 2 billion tons per year and could nearly double by 2050. The company argues that a more circular economy depends on improving how waste is sorted, automated, and scaled beyond labor-intensive manual sampling.

NVIDIA also points to a core constraint that affects recycled plastics in particular: recycled material often fails on cost and quality. According to the post, incoming waste can be inconsistent, sorting typically depends on hard-to-find manual labor, and mixed scrap can be downcycled into low-grade products, limiting how often materials can circulate in closed loops.

One example NVIDIA highlights is AIRECO, a Taiwan-based startup focused on plastic sorting for “closed-loop recycling,” where used items are reprocessed into new versions of the same product. The company says it uses vision AI to sort plastics more finely so that materials can be picked from mixed scrap piles for reprocessing. “The finer you sort, the more you can recycle,” AIRECO founder and CEO Chun-Feng Liu said in the post, arguing that precise sorting increases the odds a plastic bottle can be recycled into a bottle again rather than moving into harder-to-recycle product categories.

AIRECO’s pilot system, running at a recycling plant in Tainan, uses NVIDIA Jetson Orin NX modules for inference across its line. The post says the system also drives robotic arms that pick targets off a conveyor belt, and that recognition accuracy on the test line tops 95%. For scale-up, AIRECO plans a full production line next year with an aim to nearly triple annual waste processing capacity from 18,000 to 48,000 tons, while cutting operating costs by 85%.

NVIDIA then shifts from plastics to real-time sorting operations that must keep pace with fast conveyor belts. It describes EverestLabs AI, a U.S. startup deploying computer-vision-guided robots for material recovery facilities. NVIDIA says about 150 robots are already deployed across North America and Australia, and it quotes CEO R. Paul Singh describing the task as time-critical: unlike robots that can take “infinite time,” the systems must classify and act before material moves out of position.

In its description, NVIDIA ties EverestLabs’ performance to edge computing and software acceleration. The post says EverestLabs runs AI inference on NVIDIA RTX GPUs at the edge and trains models on NVIDIA data center GPUs, optimized using NVIDIA CUDA, a widely used programming and acceleration layer for GPU computing. NVIDIA adds that the company estimates deployed systems have identified about 40 billion objects, recycled roughly 25,000 tons of material, and avoided about 30,000 metric tons of carbon dioxide emissions. It also says EverestLabs is building “agentic AI” tools that deliver actionable data to recycling facility staff, from floor managers to executives overseeing multiple sites.

Another NVIDIA example is Greyparrot, a U.K. startup that builds an “Analyzer” platform to identify each passing object by material and brand while mounted above conveyor belts. NVIDIA says the system is designed to work even in dusty environments or poor lighting and with belts moving up to 4 meters a second. In the post, Greyparrot cofounder Ambarish Mitra says operators previously relied on manual sampling of less than 1% of the stream, and that the Analyzer changes this by providing visibility into what is actually on the belt.

NVIDIA adds measurable outcomes from Greyparrot deployments. It says more than 300 Analyzer systems operate in over 20 countries and have produced over a trillion bounding box detections to date, reducing the time needed to check 15 tons of waste from 375 hours to six. The post claims facilities using the real-time data report efficiency gains of 10% to 30%. It also cites an example at packaging company Amcor, where NVIDIA says the solution helped increase the yield of food-grade plastics by 12%, and notes that brands including Coca-Cola, Colgate-Palmolive, and Kenvue use the technology to measure how packaging is recycled at sorting plants.

Separately, NVIDIA extends the recycling theme into the used-textiles market, where it says work is slow, costly, and difficult to staff. It points to Europe generating nearly 7 million tons of used textiles a year, with about a quarter collected separately and sorted by hand into hundreds of categories. NVIDIA highlights Berlin startup as a driver of automation using high-resolution computer vision that assigns garments to their most valuable category by identifying product type, defects, color, size, brand, fabric construction, season, and style.

The post acknowledges the broader reality that recycling outcomes depend on more than sensors and software, including downstream manufacturing acceptance and material demand. It provides several performance targets and deployment counts, but it does not offer independent verification of the cited emissions reductions, accuracy figures, or projected capacity and cost results beyond the companies’ statements. It also leaves room for questions about how these pilots scale across different waste compositions and facility constraints, including maintenance, data quality, and operator adoption.

What to watch next is whether the pilot-to-production transition described by the startups becomes repeatable across multiple plants and geographies. For investors and industry observers, the key announcement will be whether edge systems using NVIDIA hardware can keep classification accuracy high under shifting inputs, and whether real-time analytics translate into measurable improvements in yield and recovery that can sustain investment in circular supply chains.