A $400 Million Bet That Cheaper AI Chips Will Win

Key points

• General Compute raised $400 million in debt financing from investment firm Upper90 in 2025.
• The loan is believed to be the first ever secured against inference chips, the specialised hardware that runs finished AI models rather than training new ones.
• General Compute's SambaNova SN50 chips claim 16 times faster inference speeds than comparable GPU-based cloud services.
• The deal follows growing investor interest in open-source AI models as a cheaper alternative to frontier models from labs like OpenAI and Anthropic.

A small AI startup has just borrowed $400 million, and the unusual part is not the number. It is what the lender accepted as collateral.

General Compute, founded by CEO Finn Puklowski and barely past its seed stage, secured the loan from Upper90, a tech-focused investment firm. The collateral: inference chips. These are specialised processors built to run a finished AI model quickly and cheaply, as opposed to the far more expensive chips used to train a model from scratch in the first place. Think of training chips as the factory that builds a car, and inference chips as the engine that actually drives it.

Upper90 co-founder Billy Libby told TechCrunch he believes this is the first loan ever backed by inference chips specifically. He should know. In 2021 his firm financed GPU purchases, the graphics processing units that became the backbone of the AI boom, for a data centre company called Crusoe. Back then, traditional banks would not touch chips as collateral because no one knew how fast they would lose value. Since then, cloud giant CoreWeave built an entire business around chip-backed loans and went public, and that type of financing is now commonplace.

Now Libby thinks inference is the next wave. "Everyone doesn't need a supercomputer, but they do need inference and AI," he said.

General Compute is building its service around chips from SambaNova, an Intel-backed chipmaker. The specific chips, called the SN50, are designed to be power-efficient and do not need the expensive water-cooling systems that high-end GPUs typically require. That means they can be installed in more standard data centres, more quickly. The company claims the SN50 delivers 16 times faster inference than GPU-based cloud rivals.

The broader pitch is cost. Running AI tasks on open-source models, AI software whose underlying code is publicly available, is significantly cheaper than paying per query to a frontier lab. Startups like OpenRouter and Fireworks, which provide access to open models, have recently raised money at high valuations, and new models from teams outside the US are now matching OpenAI and Anthropic on technical benchmarks.

What does this mean for people and businesses using AI tools?

It could mean lower prices. If infrastructure companies can run capable AI models on cheaper, more efficient chips, the cost of AI tools and services should fall for everyone downstream: businesses building apps, developers writing code assistants, or anyone paying a subscription that relies on AI.

General Compute is not the only startup making this kind of bet. TensorWave is pursuing a similar strategy built around AMD chips rather than Nvidia hardware. The common thread is a belief that Nvidia's near-total grip on AI computing will loosen as alternatives mature.

Puklowski put it plainly: "This is the first signal of capital organising itself and the fragmenting of Nvidia's monopolistic dominance."

Whether the SN50 chips deliver on their specs at scale is a question the next year will answer.