The first US$400 million financing secured by inference chips reveals new trends in AI investment
AI inference cloud startup General Compute has received a $400 million loan from technology investment firm Upper90. This may be the first deal to use inference-specific chips as collateral - chips used to run trained AI models quickly and efficiently, rather than the more expensive training chips used to build the models. The financing is the latest response to market concerns about AI tools and token prices—investors are turning to infrastructure using open source models that cost far less to run than cutting-edge labs’ latest big language models. General Compute, founded by CEO Finn Puklowski and CTO Jason Goodison, raised $15 million in seed funding in May to build around Supported by chipmaker SambaNova's silicon-based inference cloud. The company uses the SN50 chip, which is specially designed for inference, has low power consumption and does not require expensive water cooling systems, and can be deployed in a variety of data centers faster than GPUs. The inference speed of the new chip is said to be 16 times that of the GPU cloud. The challenge, especially for new companies, is acquiring large quantities of these chips. Upper90 co-founder and CEO Billy Libby provided financing for Crusoe's GPU purchase in 2021 in what is believed to be the first loan targeting the value of advanced chips. At the time, traditional lenders were reluctant to get involved due to the risk of GPU depreciation, but as CoreWeave developed chip mortgage loans into a business model and successfully IPOed, such financing has become increasingly common. Libby said: "When we took the lead in financing NVIDIA GPUs, the market was not very efficient. As an early participant, we were able to design solutions and get returns from the risks." Now that GPUs are fully recognized and even overbought, Upper90 turns to General Compute to capture the next stage of the AI wave. Libby pointed out: "We believe that open source models will become important, and last year we started looking for players in the inference field. Not everyone needs a supercomputer, but they do need inference and AI." This trend is continuing to gain momentum, and companies that provide access to open source models such as OpenRouter and Fireworks are completing new rounds of financing at high valuations. New models such as the Kimi K3 have proven to compete with the latest offerings from Anthropic and OpenAI on programming benchmarks. New chipmakers such as Groq and Cerebras are also attracting attention from acquirers and public markets. General Compute's ability to acquire chips outside of the Nvidia ecosystem is also critical. Another AI infrastructure company, TensorWave, is working with Collaborate on similar layouts. As alternatives to Nvidia continue to emerge, computing power providers that are not bound by Nvidia's agreement may have an advantage in cost-benefit reasoning. Puklowski said: "There are a number of chips that are scaling up, have excellent total cost of ownership or run faster than NVIDIA, but there are not many buyers. The cooperation with Upper90 is not only a start-up company obtaining funds to purchase computing power, but also the first signal of capital organization, marking the fission of NVIDIA's monopoly pattern."