Why the first GPU financiers are turning to inference chips in a $400 million deal
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- General Compute secured a $400 million loan from Upper90, potentially the first deal to put up inference-specific chips as collateral rather than Nvidia's training-focused GPUs.
- The startup runs SambaNova's SN50 chips, which it claims deliver 16x faster inference than GPU-based clouds, are power-efficient, and skip expensive water-cooling systems.
- Upper90 co-founder Billy Libby pioneered GPU-backed lending in 2021 by financing Crusoe's purchases — a model later validated by CoreWeave's blockbuster IPO — and now calls GPUs 'well understood and perhaps over-bought.'
- The deal reflects a market shift toward cheaper open-source inference, with OpenRouter and Fireworks raising at huge valuations and Kimi's K3 model competing with Anthropic and OpenAI on coding benchmarks this week.
- TensorWave is making a parallel non-Nvidia bet through an AMD partnership, while chipmakers Groq and Cerebras have drawn interest from acquirers and public markets.
Why it matters: By accepting inference silicon — rather than training-focused GPUs — as collateral for the first time, Upper90 is opening a new lending category that directly funds non-Nvidia AI infrastructure. General Compute's SambaNova-based cloud and TensorWave's AMD partnership become early beneficiaries as capital flow shifts toward cost-efficient inference alternatives to Nvidia's compute monopoly.
