How Early GPU Investors Are Shifting Focus to Inference Chips in a Groundbreaking $400 Million Agreement
General Compute, a startup specializing in AI inference cloud services, has secured a substantial loan of $400 million from Upper90, a technology investment firm. This financing may mark a pioneering instance of using inference-specific chips as collateral—chips designed to quickly and efficiently run pre-trained AI models, unlike the pricier chips used for model development.
This recent funding reflects a market shift towards infrastructure that supports open-source models at a lower cost, responding to rising concerns regarding AI tool and token pricing compared to advanced models from leading research labs.
General Compute, led by CEO Finn Puklowski, previously raised $15 million in seed funding in May to develop an inference-focused cloud platform utilizing silicon developed by SambaNova, a chipmaker backed by Intel. These purpose-driven clouds are tailored for AI tasks, contrasting with the general-purpose infrastructures offered by established providers like AWS and Azure.
The startup’s SN50 chips are specifically engineered for inference tasks. They are energy-efficient and do not need costly water-cooling systems, allowing for quicker deployment than GPUs in a wider range of data centers. General Compute asserts that these new chips can deliver inference speeds up to 16 times faster than those based on GPU infrastructure.
A significant hurdle for a nascent company is acquiring sufficient quantities of these chips.
Billy Libby, co-founder and CEO of Upper90 and a former quantitative trader at Goldman Sachs, had a strategy in mind: In 2021, his firm facilitated GPU purchases for Crusoe, a data center focused on energy solutions, making it one of the first businesses to secure a loan based on the value of advanced chips.
At that time, conventional lenders were hesitant to engage in such arrangements due to the uncertainties surrounding GPU value depreciation. However, as CoreWeave successfully established chip-backed loans as a viable business model and subsequently went public, this type of financing has become increasingly mainstream.
“When we financed Nvidia GPUs, we were entering an inefficient market,” Libby shared. “We managed to create a unique opportunity as early participants and were rewarded for the associated risks.”
With GPUs now more widely understood, albeit possibly over-invested in, Upper90 is now seeking to capitalize on the next phase of the AI surge with companies like General Compute. “We believe that open-source models will be significant, and last year, we searched for players focused on inference,” Libby noted. “Not everyone requires a supercomputer, but there is a clear need for inference and AI capabilities.”
This perspective gains traction as firms providing access to open models, such as OpenRouter and Fireworks, continue to secure substantial funding at impressive valuations. New entrants like Kimi’s K3 have demonstrated competitive performance against leading models from Anthropic and OpenAI in coding tests. Additionally, emerging chipmakers like Groq and Cerebras are garnering interest from buyers and public investors alike.
General Compute’s access to chips outside the Nvidia ecosystem is crucial for similar reasons. Another AI infrastructure firm, TensorWave, is pursuing a partnership with AMD, indicating a potential shift away from dependence on Nvidia. Compute providers that are not exclusively tied to Nvidia may have a competitive edge in delivering cost-effective inference solutions.
“There are several new chips coming to market with impressive total costs of ownership and the capability to operate significantly faster than Nvidia, yet there’s a limited pool of buyers for them,” Puklowski explained. “Partnering with Upper90 signals not just funding for a startup’s computing needs, but a pivotal moment in the capital landscape, potentially heralding a shift away from Nvidia’s monopolistic hold.”
