Upper90 just handed General Compute $400 million secured by SambaNova inference chips, not Nvidia GPUs, betting the same trick that built CoreWeave will work a second time.
General Compute has never designed a chip, and it doesn't own a data center either. What it has is $300 million worth of SambaNova's new SN50 processors on order and, as of this month, a $400 million credit facility from Upper90 to help pay for them. The loan is notable for one reason. The collateral isn't a GPU.
Upper90, the New York investment firm run by former Goldman Sachs trader Billy Libby, structured the facility around inference chips rather than the Nvidia hardware that has backed nearly every AI infrastructure loan to date. General Compute calls itself an inference neocloud. That's a company that rents out processing power for the moment a trained model actually answers a user, not the more publicized business of training one. "When we financed Nvidia GPUs as the first group doing so, the market was inefficient," Libby said of the deal. Now he's making the same play again. He put the broader bet plainly: "Not everyone needs a supercomputer, but everyone does need inference and AI."
The chips themselves are the pitch. SambaNova's SN50 is air-cooled. That means General Compute can drop it into existing data centers and repurposed crypto-mining facilities, without the liquid-cooling retrofits that GPU clusters increasingly require. General Compute says the chips run at 600 to 700 tokens per second, well ahead of the roughly 250 tokens per second it gets out of comparable GPU setups. That's General Compute's own figure, not an independently audited benchmark. Worth treating it that way until a neutral test confirms it.
General Compute is barely a year old. It raised just $15 million in seed funding in May, led by FUSE VC with Carya Venture Partners and Village Global Ventures, at a $60 million post-money valuation. A $400 million loan against a company that small only makes sense if the lender believes the collateral, not the balance sheet, carries the risk.
Upper90 has run this play before. In 2021, the firm financed GPU purchases for Crusoe Energy in what was, at the time, an odd bet: banks wouldn't touch chip-backed lending because nobody knew how fast a GPU's resale value would fall. Crusoe's cloud business grew. The structure Upper90 pioneered became the template CoreWeave scaled into a full financing strategy, one that eventually helped take the company public and, this year, into an $8.5 billion GPU-backed facility of its own. Libby is now betting the same arc repeats with inference silicon instead of training hardware.
General Compute's chief executive, Finn Puklowski, frames the loan as more than a funding round. "This is not just a startup that got some money for compute," he said. "This is the first sign that capital is organizing and fragmenting Nvidia's monopolistic dominance."
Everyone Wants In on Inference #
Money already agrees that inference is where the fight is. SambaNova itself just closed the first tranche of a $1 billion raise at an $11 billion valuation, led by General Atlantic with T. Rowe Price and Capital Group joining in. That's a fivefold jump in five months from the $350 million round it closed in February. Groq took a different route to the same market: it licensed its LPU chip design to Nvidia for $20 billion, then turned around and raised $650 million of its own to build a competing inference cloud. Etched, another inference-chip startup, has raised $800 million and says it already has $1 billion worth of orders on the books. Cerebras is building dedicated inference data centers under a plan to expand capacity twentyfold. None of that guarantees General Compute wins its bet on SambaNova specifically, but it tells you the capital chasing inference silicon is real, not speculative noise around one deal.
What nobody in this market has yet is a track record. GPU-backed lending works, when it works, because there are now years of resale data on how fast a Nvidia chip loses value and how deep the secondary market runs. Inference ASICs like the SN50 have none of that history. If SambaNova's chips hold their performance edge and General Compute keeps landing colocation deals, Upper90's bet looks prescient. If the tokens-per-second numbers don't survive contact with real customer workloads, the firm is holding collateral nobody has ever had to resell.
Also read: Xi Jinping tells the world AI should be a symphony, not a solo performance • Apple Hands China's AI Brain to Alibaba to Get Apple Intelligence Approved • Hackers hijacked Brian Chesky's X account to push an AI-written crypto tokenization thread