According to a report on TechCrunch, AI inference cloud startup General Compute has secured a $400 million loan from tech investment firm Upper90. This transaction may be the first of its kind in the market to use dedicated inference chips as collateral to obtain credit. These chips are designed specifically for the fast and efficient execution of already-trained artificial intelligence models, in contrast to the far more expensive chips used in the initial stages of building and training the models themselves. The new financing represents a significant signal that markets are responding to growing concerns over the high costs of AI tools and tokens, redirecting resources toward infrastructure that enables running open-source models more cheaply and cost-effectively than the newest large language models (LLMs) from frontier labs in the field.
A Shift in the AI Infrastructure Market Trend
The shift by financial institutions to financing hardware dedicated solely to inference reflects a deep change in the AI market dynamics. While in recent years most capital was directed toward acquiring powerful and expensive graphics processing units (GPUs) to train massive models, there is a growing understanding today that running these models on a day-to-day basis requires far more efficient and cheaper computing solutions. General Compute, founded by CEO Finn Puklowski, is attempting to address this exact need. This past May, the company raised a $15 million seed round to build a neocloud dedicated to inference workloads. This cloud is based on silicon chips manufactured by SambaNova, a chipmaker backed financially and technologically by chip giant Intel. Unlike traditional, general-purpose cloud providers (hyperscalers) such as Amazon's AWS or Microsoft's Azure, neoclouds are purpose-built from the ground up for the unique workloads and computing demands of artificial intelligence systems.
The Technological Advantages of SN50 Chips
The SN50 chips manufactured by SambaNova, which are utilized by General Compute, were developed specifically to provide an optimal solution for the inference phase. These chips are highly energy-efficient with low power consumption, and what sets them apart even further is that they do not require expensive and complex water-cooling systems to operate properly. This advantage allows them to be deployed and installed very quickly and with relative ease across a wide variety of data centers, without requiring the complex infrastructure adaptations necessary for traditional GPUs. According to data from General Compute, these new chips are expected to offer inference speeds that are 16 times faster than currently existing cloud solutions based on graphics processing units (GPUs). The key challenge facing brand-new, young companies like General Compute is the ability to acquire a large and sufficient quantity of these chips in a highly competitive market.
The History of Chip-Backed Financing: From Crusoe to CoreWeave
Billy Libby, co-founder and CEO of investment firm Upper90 and a former quantitative trader at investment bank Goldman Sachs, already has proven experience and a clear playbook for executing these types of transactions. In 2021, his firm was the first to provide financing for the purchase of GPUs for Crusoe, an energy-focused data center startup. Libby believes this deal represented the first loan in history granted against the physical value of advanced chips as collateral to obtain credit. At the time, traditional financiers and lenders completely avoided entering such transactions due to the high risks and high uncertainty surrounding the rapid rate of depreciation and value loss of GPUs. However, the situation changed dramatically after CoreWeave turned the chip-backed loan model into a central and thriving business model, even using it as the foundation for a highly successful initial public offering (IPO). Consequently, this type of financing has become common and widely accepted in the industry.
Shifting from Nvidia GPUs to Dedicated Inference Chips
"When we financed Nvidia GPUs as the first group to do that, the market was inefficient," Billy Libby shared in an interview with TechCrunch. "We could really put together something as an early participant, and kind of get compensated for the risk." Now that the market thoroughly understands the value of traditional GPUs and may have even over-bought them, Upper90 is choosing to direct its resources to companies like General Compute to ride the next wave of the AI boom. "We think open source models are going to be important, and we went and looked for a player last year that was in inference," Libby explained the rationale behind the investment. "Everyone doesn't need a supercomputer, but they do need inference and AI."
The Rise of Open-Source Models and Chip Alternatives
The thesis driving Upper90 has received significant reinforcement on the ground recently. Companies providing convenient and direct access to open models, such as OpenRouter and Fireworks, have recently completed new funding rounds at exceptionally high valuations. Furthermore, newly released open models, such as Kimi's K3 model released just this week, have proven capable of competing in coding benchmarks and performance metrics against the most advanced versions from leading companies in the field, including Anthropic and OpenAI. In parallel, innovative chipmakers like Groq and Cerebras are generating immense interest from both potential acquirers and public markets. General Compute’s ability to access and procure advanced chips outside of chip giant Nvidia's closed ecosystem is of critical importance for this very reason. Another AI infrastructure company, TensorWave, is making a very similar bet through a strategic partnership it established with AMD. As more viable alternatives to Nvidia emerge in the market, compute providers that are not bound by rigid, exclusive agreements with it may hold a massive competitive advantage in delivering incredibly cheap and efficient inference services.
Breaking Nvidia's Monopolistic Dominance
Finn Puklowski, CEO of General Compute, explains the historic importance of this move: "There are a bunch of chips that are starting to scale that have amazing [total cost of ownership] (TCO), or that can operate much faster than Nvidia, but there’s not too many buyers for them at this stage in the market. By partnering financially with Upper90, this is not just 'a cool startup got some money to buy some compute.' This is actually the first significant signal of capital organizing itself, marking the beginning of the fragmentation of Nvidia's monopolistic dominance in AI computing." The shift toward financing deals based on inference chips as collateral thus signals the beginning of a new, multipolar era in the AI hardware market.