The debt markets that supercharged the GPU boom just placed their first big wager on life after Nvidia. New York investment firm Upper90 is lending $400 million to General Compute, an AI inference cloud, in a deal reported by TechCrunch on July 17 — with the loan collateralized not by Nvidia GPUs but by inference-specific chips.
Why the collateral matters
GPU-backed lending became a defining financial structure of the AI buildout — tens of billions in credit extended against Nvidia hardware, on the logic that H100s hold resale value like aircraft. This deal appears to be the first major facility secured by non-Nvidia inference silicon: SambaNova SN50 accelerators. If lenders will underwrite alternative chips, the single biggest financing moat around Nvidia's ecosystem starts to crack.
The parties
Upper90 knows the trade: CEO Billy Libby's firm financed Crusoe's early GPU purchases, making it one of the first GPU financiers anywhere. General Compute is young — a $15 million seed in May 2026 — and is building its cloud on SambaNova hardware, claiming 16x faster inference than GPU-based clouds, a company figure. SambaNova itself, backed by Intel, reached an $11 billion valuation this month.
The thesis
"We think open source models are going to be important... Everyone doesn't need a supercomputer, but they do need inference and AI," Libby told TechCrunch. The bet: training compute stays concentrated in a few frontier labs, but inference — serving models to actual users — fragments across specialized, cheaper hardware as open-weight models proliferate.
The quote of the deal
General Compute CEO Finn Puklowski was blunter: "this is the first signal of capital organizing itself and the fragmenting of Nvidia's monopolistic dominance." That is a $400 million claim resting on whether SN50 clusters hold their value the way H100s did — precisely what this loan's performance will now test in public.