NVIDIA unveils Cosmos 3 Edge for robots; General Compute raises $400M backed by inference chips

NVIDIA continued its physical-AI offensive with Cosmos 3 Edge, an AI model designed for robots and vision AI agents running at the edge, and announced that Japan's robotics and manufacturing leaders are building on its Cosmos, Isaac, Metropolis and Jetson platforms. It complements the same-week LeRobot expansion with Hugging Face and reinforces NVIDIA's strategy of owning the full embodied-AI stack from training GPUs to edge inference models.

The more novel signal came from financing. Inference-cloud startup General Compute secured a $400 million loan from Upper90 — reportedly the first deal to use inference-specific chips as collateral. TechCrunch framed this as a milestone in how AI infrastructure gets financed: as inference rises to roughly 70% of total AI compute cost, the chips running production workloads are becoming financeable, cash-flowing assets rather than depreciating capex.

This matters competitively because it deepens NVIDIA's moat in a subtle way — its silicon is now bankable collateral, lowering the cost of capital for the cloud providers and startups that buy it, which in turn drives more NVIDIA demand. It also mirrors the week's compute-scarcity theme running through the Anthropic-Meta lease talks and DeepSeek's custom-chip ambitions.

The Japan ecosystem push is a geographic hedge, positioning NVIDIA at the center of a manufacturing powerhouse racing to deploy physical AI. Caveats: collateralizing rapidly-depreciating, fast-obsoleting AI chips carries real risk if a hardware generation shift craters resale value, and edge robotics models like Cosmos 3 still face the hard sim-to-real generalization problems dogging the whole field. What to watch is whether inference-chip-backed lending becomes a broader financing norm.