Upper90 has issued a $400 million loan to AI inference cloud startup General Compute, collateralized by SambaNova's SN50 inference chips rather than Nvidia GPUs. It appears to be the first major financing to treat inference-specific silicon as loan collateral, and it lands just months after General Compute closed a $15 million seed round in May. Upper90, run by former Goldman Sachs quantitative trader Billy Libby, is the same firm that in 2021 financed Crusoe's GPU purchases in what it believes was the first loan ever written against the value of advanced AI chips.
General Compute, founded by CEO Finn Puklowski, is building a neocloud — infrastructure purpose-built for AI workloads rather than the general-purpose fleets of AWS or Azure — around SambaNova's SN50 chips. SambaNova is backed by Intel, and its silicon is designed for inference, meaning it runs already-trained models rather than training new ones. The company says the SN50-based cloud delivers 16 times faster inference than GPU-based clouds.
The chips have a second structural advantage: they are power-efficient and do not require water cooling, which shortens deployment timelines and widens the pool of data centers that can host them. That matters because access to Nvidia GPUs has become one of the tightest constraints in AI infrastructure, and every non-Nvidia deployment path expands supply.
Key facts
- 01Upper90 issued a $400 million loan to General Compute, collateralized by SambaNova SN50 inference chips rather than Nvidia GPUs.
- 02General Compute raised a $15 million seed round in May to build an inference neocloud around SambaNova silicon.
- 03The startup claims its SN50-based cloud runs inference 16 times faster than GPU-based clouds.
- 04Upper90 previously financed Crusoe's Nvidia GPU purchases in 2021, in what it says was the first loan against advanced chip value.
- 05SN50 chips are power-efficient and skip water cooling, allowing faster deployment across more data center footprints.
Libby's pattern-recognition here is worth naming. In 2021, traditional lenders wouldn't touch GPU-backed loans because chip depreciation was too uncertain. Upper90 wrote the deal to Crusoe anyway, and by the time CoreWeave built the model into an entire business and rode it into a blockbuster IPO, chip-backed lending had become mainstream.
Now GPUs are comparatively well understood — and, in Libby's read, arguably over-bought. Upper90's move into inference-chip financing is a bet that the next inefficient market is the one for silicon that runs open-source models cheaply, rather than the silicon that trains frontier ones.
The thesis has supporting evidence. OpenRouter and Fireworks, which provide access to open models, have raised new rounds at large valuations. Moonshot's Kimi K3, released this week, has posted coding-benchmark results competitive with the latest releases from Anthropic and OpenAI, a story AI Chat Daily covered earlier alongside Moonshot's $31.5B valuation. And chipmakers Groq and Cerebras have drawn interest from both acquirers and public markets.
TensorWave is making a parallel bet with AMD silicon. The through-line: compute providers not locked into Nvidia may have a structural cost advantage in inference, where margins are thinner and workloads are more price-sensitive than in training.
Puklowski frames the deal as more than a startup financing round. He argues it signals that the capital stack itself is starting to organize around Nvidia alternatives, which is a different claim than the usual startup pitch that a new chip is technically superior. Superior chips have existed before; the question was always whether anyone would finance them at scale.
The risks are real. SambaNova has been shipping silicon for years without becoming a household name in AI infrastructure, and SN50 adoption depends on General Compute converting customers who are already tooled for CUDA. A $400 million loan against a young chip line is a bigger balance-sheet bet than a $15 million seed, and if inference demand consolidates around a handful of hyperscaler-hosted frontier models rather than open-source workloads, the collateral thesis weakens quickly.
The interesting move here is not the loan size but the collateral class. Every non-Nvidia AI infrastructure play has faced the same financing gap: institutional lenders wanted GPU-backed paper because that's what CoreWeave normalized. If Upper90's SN50-backed loan performs, it opens a template for financing AMD, Groq, Cerebras, and SambaNova buildouts at scale — which is the actual mechanism by which Nvidia's pricing power gets challenged. Chips don't fragment a monopoly. Capital does.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
