💾 The First GPU Financiers Are Now Backing Inference Chips in a $400M Deal

Some of the investors who first turned GPUs into a financeable asset class are now placing their next bet a layer down the stack: Upper90, led by former Goldman Sachs trader Billy Libby, has extended a $400 million loan to AI inference startup General Compute — reportedly the first deal to use inference chips as loan collateral.

General Compute, founded by Finn Puklowski and Jason Goodison, is building its clusters around SambaNova's SN50 inference chips, which the company claims deliver “16 times faster inference than GPU-based clouds” while drawing less power and needing less water-cooling. Structuring the financing as debt against the hardware itself is the same playbook that made GPU-backed lending a multibillion-dollar market — now pointed at a different silicon target.

The reasoning is a read on where AI demand actually sits. “Everyone doesn't need a supercomputer, but they do need inference and AI,” Libby said. Training frontier models is a game for a handful of labs; serving those models to millions of users is the workload that scales — and inference-optimized chips can be cheaper and more efficient at exactly that job.

It is also a small crack in Nvidia's gravitational pull. As lenders grow willing to underwrite non-Nvidia hardware, capital starts to fragment across the AI compute market, and alternative chipmakers gain a financing on-ramp they previously lacked. The deal won't dent Nvidia's lead on its own, but it shows the money is beginning to hedge.

Key details

• The financier: Upper90, led by ex-Goldman Sachs trader Billy Libby, provided a $400 million loan.
• The borrower: General Compute, an inference startup founded by Finn Puklowski and Jason Goodison.
• The hardware: SambaNova SN50 inference chips, claimed at 16x faster inference than GPU-based clouds.
• Why it's notable: Reportedly the first deal to use inference chips — not GPUs — as loan collateral.

Why it matters

Financing follows conviction. When lenders start accepting inference silicon as collateral, it's an early signal that the market believes serving AI — not just training it — is where durable, financeable demand lives, and that the compute supply chain is diversifying beyond a single vendor.

Source: TechCrunch. Reporting cross-referenced by Tech Bytes on Jul 17, 2026.