For four years, every large chip-backed loan in AI has rested on the same asset: Nvidia's GPUs. The template was set in 2021, when Upper90 Capital lent against advanced chips for the first time, and cemented in 2023 when CoreWeave borrowed $2.3 billion against a rack of H100s. CoreWeave now carries more than $21 billion in such debt, some of it rated investment grade. So it is worth pausing on what Upper90 did last week. It committed up to $400 million to a Boston startup called General Compute, and for the first time, the collateral is not Nvidia silicon at all. It is inference chips: SambaNova's SN50 accelerators, hardware built to do one narrow job as cheaply as possible.
The distinction is not a technicality. It is a wager on where the durable money in AI actually lives. Training a model is a burst of intense computation that happens once, then the cluster idles between runs. Inference, the work of answering every prompt and running every agent step, never stops. As AI seeps into ordinary software, serving those models starts to look less like a project and more like a utility, a stream of billable tokens that flows day and night. Goldman Sachs projects token consumption will grow roughly twenty-four-fold in the coming years. Upper90's chief executive, Billy Libby, a former Goldman trader, frames it plainly: lenders follow recurring revenue, and recurring revenue is moving from the training cluster to the inference layer.
SambaNova's chips are built for that layer. Generating text is bottlenecked not by raw compute but by how fast a chip can move data between memory and processor, and the SN50's design places the two side by side to cut that distance. General Compute pairs it with AMD cards that handle the front end of each request, and independent testing clocked the hybrid setup at 763 tokens per second on one model, several times faster than GPU-only rivals. The chips also run on air cooling rather than water, which means new capacity can be installed in weeks instead of the multi-year construction that liquid-cooled GPU halls demand. Each slice of the loan is drawn only as General Compute signs paying customers, so the debt grows in step with revenue rather than ahead of it.
The risk sits in one word: resale. Asset-backed lending assumes the collateral can be sold if the borrower fails. A used Nvidia GPU has a growing secondary market and loses about half its value over three years, a curve lenders can price. A used SN50 has no secondary market at all, because none has ever been sold. Worse, an inference chip tuned for today's model architecture could be left nearly worthless if that architecture shifts, the same fate that stranded Bitcoin mining chips when their algorithms changed. Upper90 has accepted a chip whose salvage value is genuinely unknown, on terms it has not disclosed.
Whether the bet pays off will take years to learn. What it signals arrives now. Two weeks after SambaNova raised money at an $11 billion valuation, five times its price in February, a major lender independently reached the same verdict from the credit side: inference silicon is an asset class of its own, separate from the training hardware that defined the last cycle. General Compute's founder calls it "the fragmenting of Nvidia's monopolistic dominance." That may be premature. But the debt markets, which are usually the last to move, have started to price a future in which the money follows the tokens, not the training run.