General Compute has secured a $400 million loan from Upper90, according to TechCrunch, in a deal that points to changing assumptions about how AI infrastructure gets financed.

The startup is building an inference-focused cloud rather than a training cluster. Inference is the stage where already-trained models respond to users, and it is becoming a larger cost center as coding agents, voice agents and other real-time AI products move into production.

The notable part of the financing is the collateral. TechCrunch reported that the loan may be the first to use inference-specific chips, rather than GPUs, as the asset backing the debt. General Compute is using chips from SambaNova, whose SN50 hardware is designed for inference workloads. The company says those systems are more power-efficient and easier to deploy than water-cooled GPU clusters, which could matter for providers trying to add capacity outside the most constrained data centers.

Upper90 is not new to asset-backed startup credit. The firm says it provides debt and equity to technology companies with predictable revenue or collateral, and its portfolio page now lists General Compute.

The transaction does not prove that specialized inference chips will replace GPUs. It does show lenders are beginning to underwrite a broader set of AI compute assets as demand shifts from model training toward high-volume model serving.