Inference Chip Financing Emerges as Wall Street Bets Against GPU Dominance

General Compute, a startup building cloud infrastructure for AI inference, has secured a $400 million loan from Upper90, marking what appears to be the first major financing deal to use inference-specific chips as collateral. The agreement underscores mounting investor appetite for alternatives to the pricey, power-hungry processors that dominate AI model training.

The Shift From Training to Inference#

The financing represents a strategic pivot by early GPU investors. Instead of bankrolling the expensive hardware needed to train cutting-edge language models, financial firms are now targeting the less glamorous but potentially more profitable business of running already-trained models efficiently. General Compute, founded by Finn Puklowski, is building infrastructure around SN50 chips made by SambaNova, an Intel-backed chipmaker designed specifically for inference workloads.

Unlike traditional GPUs, these inference chips consume less power, don't require costly water-cooling systems, and can be rolled out more quickly across diverse data centers. The company claims its setup delivers 16 times faster inference performance than GPU-based alternatives.

Following the GPS-Backed Loan Playbook#

Upper90's Billy Libby, a former Goldman Sachs quantitative trader, pioneered the model of using advanced semiconductors as collateral. In 2021, his firm financed GPU purchases for Crusoe Energy, a data center startup focused on efficient computing. At the time, traditional lenders viewed such deals as too risky, given uncertainty about hardware depreciation and long-term value.

That calculus shifted as CoreWeave transformed chip-backed lending into a sustainable business model and ultimately went public—legitimizing the practice and opening the door for new investors to replicate it.

Why Inference Chips Matter Now#

The emergence of cheaper inference alternatives reflects broader market concerns about AI tool affordability. As open-source models improve—recent entrants like Kimi's K3 are matching proprietary benchmarks from Anthropic and OpenAI—enterprises increasingly realize they don't need the most powerful training infrastructure. What they need is affordable, reliable inference at scale.

Companies offering open-model access, including OpenRouter and Fireworks, have raised money at soaring valuations. Meanwhile, specialized chip designers like Groq and Cerebras are attracting acquirer interest and public market attention, signaling genuine diversification away from Nvidia's grip on the sector.