Why the first GPU financiers are turning to inference chips in a $400 million deal
TechCrunch Tim Fernholz
General Compute got a $400M loan backed by inference chips, not GPUs. It's the first sign lenders see cheap AI chips as the next big bet.
Based on reporting by TechCrunch, Tim Fernholz — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Every AI infrastructure story until now has been about GPUs — buying them, financing them, hoarding them. This one is different. General Compute, a young inference cloud startup led by CEO Finn Puklowski, just pulled in a $400 million loan from Upper90, a tech investment firm, and the collateral isn't Nvidia silicon. It's chips built specifically to run already-trained models, not to train new ones.
The company is building what's called an inference neocloud, purpose-built infrastructure rather than the general-purpose data centers hyperscalers like AWS or Azure run. General Compute raised a $15 million seed round back in May to do it, centered on silicon from SambaNova, the Intel-backed chipmaker. The SN50 chips at the heart of the plan are power-efficient, skip the need for expensive water cooling, and can be dropped into a wider range of data centers faster than GPUs. General Compute says the setup delivers inference sixteen times faster than GPU-based clouds.
Getting a pile of specialized chips as a brand-new company is the hard part, and that's where Upper90 comes in. Billy Libby, the firm's co-founder and CEO and a former Goldman Sachs quantitative trader, has done this before. Back in 2021, Upper90 financed GPU purchases for Crusoe, the energy-focused data center startup, in what Libby believes was the first loan ever made against advanced chip value. Traditional lenders wanted nothing to do with GPU depreciation risk at the time. But CoreWeave later turned that same playbook into a business model and eventually an IPO, and now chip-backed lending is simply how things work.
Libby argues GPUs are now well understood, maybe even over-bought, which is why Upper90 went looking for the next inefficiency. "Everyone doesn't need a supercomputer, but they do need inference and AI," he told TechCrunch. That bet lines up with what's happening elsewhere: companies giving access to open models, like OpenRouter and Fireworks, are raising rounds at big valuations, and newer releases like Kimi's K3 are already trading blows with the latest from Anthropic and OpenAI on coding benchmarks. Chipmakers outside Nvidia's orbit, Groq and Cerebras among them, have started drawing serious interest from acquirers and public markets.
General Compute's access to non-Nvidia chips fits that same pattern, and it isn't alone — TensorWave, another infrastructure company, is making a similar wager with AMD. Puklowski frames the deal as something bigger than a funding milestone. "This is the first signal of capital organizing itself and the fragmenting of Nvidia's monopolistic dominance," he said, pointing out that plenty of chips now offer strong cost of ownership or faster operation than Nvidia's, they just don't have enough buyers yet.
My take — AI-written commentary, not fact-checked reporting
Nvidia's dominance was never going to last forever once the money people figured out how to price the alternatives, and this loan is the clearest evidence yet that they have. Betting on inference chips over training GPUs is really a bet that most of the AI economy will be about running models cheaply at scale, not building bigger ones — and that's probably the correct bet. The irony is that it took a former Goldman trader repeating his own old GPU playbook to make that case credible to a lender.
Read more about this at: TechCrunch