First Inference-Chip-Backed Loan Signals Shift from GPU Dominance in $400 Million Deal

General Compute, a Boston-based AI inference cloud startup, has secured a $400 million loan from investment firm Upper90 — marking what is believed to be the first major financing deal backed by inference-specific chips as collateral, rather than the Nvidia GPUs that have dominated the AI infrastructure lending market.

What makes this deal different from GPU-backed loans

Traditional AI infrastructure financing has centered on Nvidia GPUs, which are used to train large language models and have become a widely accepted asset class for lenders. Upper90 co-founder and CEO Billy Libby, a former Goldman Sachs quantitative trader, pioneered this approach in 2021 when his firm financed GPU purchases for data center startup Crusoe. Since then, companies like CoreWeave have turned chip-backed loans into a billion-dollar business model.

General Compute’s loan instead uses chips from SambaNova, an Intel-backed chipmaker, that are purpose-built for inference — the process of running already-trained AI models. These SN50 chips are more power-efficient than Nvidia GPUs and do not require expensive water-cooling systems, allowing faster deployment across a wider range of data centers. General Compute claims the chips provide 16 times faster inference than GPU-based clouds.

Why inference chips are attracting capital now

The deal reflects a broader market shift toward cost-efficient AI infrastructure. As of November 2024, the cost of running AI models has become a central concern for enterprises and developers, with many turning to open-source models that can run on cheaper hardware. Companies like OpenRouter and Fireworks, which provide access to open models, have raised new funding rounds at high valuations. New models such as Kimi’s K3 have recently matched the coding benchmarks of releases from Anthropic and OpenAI.

“Everyone doesn’t need a supercomputer, but they do need inference and AI,” Libby told Bitcoin World. “We think open source models are going to be important.”

Implications for Nvidia’s market position

General Compute CEO Finn Puklowski framed the deal as more than a financing event. “This is the first signal of capital organizing itself and the fragmenting of Nvidia’s monopolistic dominance,” he said. The company’s ability to access chips outside the Nvidia ecosystem gives it flexibility in pricing and deployment. Other infrastructure providers, such as TensorWave, are making similar bets with AMD chips.

As more chipmakers like Groq and Cerebras gain traction, compute providers not locked into Nvidia deals may have a cost advantage in inference workloads. “There are a bunch of chips that are starting to scale that have amazing total cost of ownership, but there’s not too many buyers for them,” Puklowski said.

Conclusion

The $400 million loan to General Compute represents a maturing of the AI infrastructure financing market, moving beyond Nvidia GPUs to a broader set of hardware assets. For enterprises seeking cheaper ways to run AI models, this shift could accelerate the availability of cost-efficient inference services. The deal also signals that lenders are gaining confidence in alternative chip architectures, potentially reshaping the competitive landscape of AI cloud computing.

FAQs

Q1: What is an inference chip?
An inference chip is a processor designed specifically to run already-trained AI models quickly and efficiently, as opposed to training chips like Nvidia GPUs that are used to build models. Inference chips typically consume less power and cost less than training hardware.

Q2: Why is this loan considered a first?
This is believed to be the first major loan where inference-specific chips, rather than Nvidia GPUs, were used as collateral. Previous chip-backed loans in the AI sector have almost exclusively used Nvidia GPUs.

Q3: How does this affect the cost of AI services?
If inference chips become more widely financed and deployed, cloud providers can offer lower prices for running AI models. This could make AI tools more affordable for startups and enterprises that don’t need the massive computing power of training clusters.