Key facts
- General Compute has secured a $400 million loan from Upper90.
- The loan is backed by AI inference chips, a novel form of collateral.
- General Compute's SN50 chips are designed for efficient AI model inference.
- Upper90 previously pioneered chip-backed financing with Crusoe's GPU purchases.
- The deal suggests a growing market for AI inference solutions and alternatives to Nvidia's GPUs.
AI inference cloud startup General Compute has secured a $400 million loan from tech investment firm Upper90, marking a significant development in AI infrastructure financing. This deal is notable as it reportedly uses AI inference chips as collateral, a first for such specific hardware. The financing underscores a growing market trend towards more cost-effective AI solutions, particularly for running open-source models, as concerns about the expense of cutting-edge AI tools and tokens persist.
Founded by CEO Finn Puklowski, General Compute aims to build a 'neocloud' for AI workloads using silicon from SambaNova. The company's SN50 chips are specifically designed for inference, offering power efficiency and avoiding the need for expensive water-cooling systems, which General Compute claims will deliver 16 times faster inference speeds compared to GPU-based clouds. Accessing a substantial quantity of these chips is a key challenge for a new company.
Upper90's co-founder Billy Libby, a former quantitative trader at Goldman Sachs, has prior experience in chip-backed financing, having financed GPU purchases for Crusoe in 2021. This type of financing, initially seen as risky due to potential GPU depreciation, has become more common following the success of companies like CoreWeave. Libby noted that Upper90 was an early participant in the then-inefficient market for GPU financing.
With the GPU market now more understood and potentially saturated, Upper90 is shifting its focus to companies like General Compute, anticipating the next phase of the AI boom. Libby believes that while not everyone requires a supercomputer, there is a widespread need for AI inference capabilities. This thesis is supported by the rising valuations of companies providing access to open models, such as OpenRouter and Fireworks, and the competitive performance of new models like Kimi's K3.
The emergence of alternative chipmakers like Groq and Cerebras, and companies like TensorWave partnering with AMD, indicates a diversification beyond Nvidia's ecosystem. Puklowski highlighted that these alternative chips offer compelling total cost of ownership and performance advantages, but face limited buyers. He views the General Compute-Upper90 deal as a signal of capital organizing to challenge Nvidia's dominant market position.