General Compute, an AI inference cloud startup, has secured a $400 million loan from Upper90 using inference-specific chips as collateral. It is the first major financing deal of its kind: a loan backed by chips designed to run already trained AI models efficiently, rather than the expensive Nvidia GPUs used to train them. The deal signals that the capital markets are responding to a structural shift in the AI industry from training-dominated compute spending to inference, where models actually run in production. And it marks the moment when GPU financiers, who made billions financing Nvidia chip purchases during the training gold rush, are pivoting to the next wave.
General Compute was founded by CEO Finn Puklowski and CTO Jason Goodison. The company raised a seed round in May to build what they call an inference neocloud, a purpose-built cloud designed specifically for running AI models rather than training them. The company uses silicon from SambaNova, an Intel-backed chipmaker, specifically the SN50 chip. The SN50 is designed for inference: it is power-efficient, does not require expensive water-cooling systems, and can be deployed across a wider variety of data centers than Nvidia's H100 or B200 GPUs. General Compute claims the SN50 delivers 16 times faster inference than GPU-based clouds.
How the Deal Works
The financing structure is itself noteworthy. Upper90 co-founder and CEO Billy Libby, a former Goldman Sachs quantitative trader, pioneered chip-backed lending in 2021 when his firm financed GPU purchases by Crusoe, an energy-focused data center startup. At the time, traditional lenders avoided chip-backed loans because of concerns about GPU depreciation and resale value. But as CoreWeave made chip-backed loans into a core business model and then the basis of a blockbuster IPO, the financing category became established. Now Upper90 is applying the same playbook to inference chips, betting that the collateral value of inference hardware will follow the same trajectory as training GPUs.
Libby articulated the thesis directly: 'We think open source models are going to be important, and we went and looked for a player last year that was in inference. Everyone does not need a supercomputer, but they do need inference and AI.' This is a bet that the AI infrastructure market is bifurcating. Training GPUs will remain a concentrated market driven by the handful of frontier labs that train the largest models. But inference compute will be vastly larger in total volume, serving thousands of companies running AI applications in production. The capital markets are beginning to recognize that inference, not training, is where the long-term compute demand lies.
Why Inference Chips Matter Now
The timing of this deal reflects several converging trends. Open-source models have proven increasingly competitive with proprietary frontier models, with new releases matching or exceeding the performance of Anthropic and OpenAI models on key benchmarks like coding. As open models improve, more companies are choosing to self-host or use inference clouds rather than paying per-token API fees to frontier labs. This drives demand for inference-optimized hardware that can run these models cost-effectively at scale.
New chipmakers are entering the inference market as well. Groq and Cerebras have drawn interest from acquirers and public markets. SambaNova, Intel, AMD, and a growing list of companies are producing alternatives to Nvidia's dominance. General Compute's ability to access chips outside of the Nvidia ecosystem is strategically important. As more alternatives emerge, compute providers that are not locked into Nvidia pricing may have a structural advantage in providing cost-efficient inference. General Compute's CEO Puklowski framed it bluntly: 'There are a bunch of chips that are starting to scale that have amazing total cost of ownership, or that can operate much faster than Nvidia, but there are not too many buyers for them. By getting together with Upper90, this is the first signal of capital organizing itself and the fragmenting of Nvidia's monopolistic dominance.'
What This Means for Founders
For founders building AI applications, this deal carries a clear signal: inference costs are going to come down significantly over the next 12 to 24 months. The entry of specialized inference hardware, combined with the capital markets now willing to finance it, means the unit economics of running AI models in production will improve. Companies like OpenRouter and Fireworks, which provide access to open models, have already raised significant rounds at huge valuations. The inference infrastructure layer is attracting serious capital, and that capital will translate into lower prices for end users.
Founders should take three actions. First, evaluate inference-optimized hardware for production workloads. If General Compute delivers on its 16x speed improvement claim, the cost savings could be transformative for AI applications with high inference volume. Second, track the open-source model ecosystem closely. As open models improve and inference costs fall, the economic case for using frontier model APIs weakens. Startups that build flexibility into their model selection can benefit from the rapidly improving cost-performance curve of open models. Third, watch the financing environment. The fact that chip-backed loans are expanding from training GPUs to inference hardware means the capital markets are betting on massive inference demand. That is a positive signal for anyone building AI applications that will need inference compute at scale.
The $400 million General Compute deal is a small number compared to the tens of billions flowing into GPU infrastructure. But as a signal, it is significant. The first GPU financiers are now inference financiers, and that shift tells you where the AI infrastructure opportunity is heading next.