Key Takeaways

  • Upper90's $400 million loan to General Compute marks the first time inference-specific chips — not training GPUs — have served as collateral, signaling a structural shift in how AI infrastructure gets financed.
  • The deal bets on a simple thesis: most companies don't need supercomputers to train models; they need cheap, fast inference to run open-source models that are rapidly closing the gap with frontier labs.
  • SambaNova's SN50 chips dodge the water-cooling and power constraints that slow GPU deployments, letting a neocloud scale across ordinary data centers instead of hyperscaler bottlenecks.
  • The same financiers who pioneered GPU-backed loans in 2021 are now rotating into inference because GPUs have become "over-bought" and well-understood — the risk premium has moved downstream.

The first financiers to treat advanced chips as bankable assets are done with GPUs. Upper90, the firm that backed Crusoe's Nvidia purchases in 2021 before chip-backed lending became a CoreWeave IPO playbook, just dropped $400 million on General Compute — collateralized by SambaNova SN50s built solely for inference. That detail matters more than the headline number. It tells you where the smart money thinks the next margin lives: not in training frontier models, but in serving the open-source alternatives that are already eating their lunch.

Finn Puklowski and Jason Goodison founded General Compute a year ago with a $15 million seed round and a conviction that inference would outgrow training. The SN50s they're deploying don't need water-cooling. They sip power. They slot into ordinary data centers instead of the hyperscaler bottlenecks that constrain GPU clouds. General Compute claims 16 times faster inference than GPU-based alternatives. If that holds, the economics flip: open models like Kimi's K3, which already match Anthropic and OpenAI on coding benchmarks, become cheaper to run than the proprietary APIs they compete with.

Upper90's Billy Libby puts it bluntly. "Everyone doesn't need a supercomputer, but they do need inference and AI." The former Goldman quant watched the GPU lending market mature from inefficient risk into a crowded trade. Now he's hunting the next inefficiency. He found it in a neocloud — purpose-built for AI workloads, unlike the general-purpose slop AWS and Azure still push — running silicon outside Nvidia's moat. That last part is deliberate. Groq and Cerebras are drawing acquirer interest for the same reason: diversification of supply chain matters when one vendor controls the calendar.

The open-source inference thesis has teeth. OpenRouter and Fireworks have raised at valuations that would have looked absurd eighteen months ago. Developers are voting with tokens. They'll take a model that costs a fraction of GPT-4o if it solves the coding task. General Compute's job is to make that fraction shrink further. The SN50's power profile lets them deploy density where GPU clouds can't — colocation facilities, regional data centers, anywhere with a rack and a power drop. Speed of deployment becomes a competitive weapon when demand is elastic and supply is constrained.

Skepticism has a seat at this table. General Compute is a year old. SambaNova, for all its Intel backing, has yet to prove volume manufacturing at the scale Nvidia takes for granted. Sixteen-times-faster claims live in marketing decks until real workloads stress them. Upper90's loan structure — presumably senior secured against the chips — still prices in depreciation risk that nobody has modeled for inference silicon. The Crusoe precedent worked because GPU demand outran supply for years. Inference demand is real but less proven at scale.

Yet the directional signal is unmistakable. The first movers in chip finance are rotating capital from training infrastructure to inference infrastructure. They're not hedging. They're concentrating. TensorWave and others are watching. If General Compute executes, the neocloud model — specialized, chip-agnostic, open-source-native — becomes the default architecture for the next phase of AI deployment. The $400 million isn't a bet on a startup. It's a bet on where the compute economy's center of gravity shifts when models commoditize and serving them becomes the business.