General Compute, a startup that rents out inference capacity, borrowed $400 million from Upper90, a tech-focused investment firm, putting up its own chips as collateral rather than cash flow or equity. The security behind the loan is SambaNova’s SN50 hardware, silicon built to run already-trained models cheaply rather than train new ones. Lenders backing non-Nvidia silicon is the actual news here.
The structure matters because it reassigns risk. Equity investors absorb a company’s losses in proportion to their stake. A secured lender expects repayment first, ahead of everyone else, which means Upper90 is not betting on General Compute’s business so much as on whether the SN50 chips themselves would fetch a decent price if the company defaulted.
Upper90 chief executive Billy Libby has run something like this play before. A former quantitative trader at Goldman Sachs, he financed GPU purchases for Crusoe, the data center operator built around cheap power, back in 2021. He describes it as the first loan collateralized against advanced chips. Banks stayed away from that kind of deal at the time, unsure how fast a GPU would lose value once installed. CoreWeave later turned chip-backed lending into a business model, then into the foundation of its IPO.
Nvidia GPUs now have a track record lenders can price. SambaNova’s SN50 does not: no public depreciation curve, no liquid resale market the size of Nvidia’s, no history of holding value through a borrower’s default. Upper90 is treating that gap as an opportunity rather than a reason to pass, the same calculation it made on GPUs five years earlier.
Libby frames the shift as early-mover economics repeating itself. “When we financed Nvidia GPUs as the first group to do that, the market was inefficient,” he told TechCrunch. GPU collateral is priced efficiently now, which is another way of saying the excess returns are gone. Inference silicon, he argues, is where that inefficiency still sits.
General Compute’s pitch rests on a specific performance claim: the company says SN50 hardware runs inference roughly 16 times faster than GPU clouds, without the water-cooling buildout Nvidia’s chips typically demand. That figure comes from the company, not an independent benchmark. The same cheaper-inference thesis is behind recent high-valuation rounds for OpenRouter and Fireworks, and behind Kimi K3, the coding model from Chinese lab Moonshot AI that has drawn comparisons to releases from Anthropic and OpenAI.
General Compute is not the only startup pitching lenders on non-Nvidia hardware. TensorWave is pursuing a comparable structure built on AMD chips, and Groq and Cerebras have each attracted acquisition and public-market interest as inference alternatives to Nvidia. General Compute itself, led by chief executive Finn Puklowski and chief technology officer Jason Goodison, closed a seed round of $15 million in May: a thin equity cushion beneath $400 million in new secured debt.
Puklowski calls the deal a sign that capital is starting to fragment Nvidia’s dominance. That reading assumes the SN50 chips will still have buyers three years from now if General Compute cannot service the loan, and nothing disclosed about the deal answers that question. Operators weighing inference contracts should watch whether other lenders follow Upper90 into non-Nvidia collateral over the next two quarters; if none do, this loan is a single bold bet, not a new asset class.
TechCrunch’s Tim Fernholz reported this story on July 17, 2026.
