There is a cost problem building quietly inside most businesses that have started using AI seriously. Every time an AI tool answers a question, drafts a document, or handles a customer interaction, it costs money. That cost scales with usage, and it has been catching companies off guard. Uber burned its entire 2026 AI budget in four months. A Goldman Sachs report found that in one software company, AI inference costs were approaching 10 percent of total headcount costs.
The companies charging those fees are mostly OpenAI, Anthropic, and Google, running their most capable models. But a wide and growing group of open-source models, built by labs like Meta, DeepSeek, and others, now come within a few percentage points of those premium models on most standard business tasks, at a fraction of the price. Researchers at MIT found that the cost of running inference on open models is 87 percent lower than on closed models. For high-volume work like customer support, document review, or content drafting, the quality difference is often undetectable.
That cost gap is exactly what a startup called General Compute is positioning itself to serve. The company just secured a $400 million loan from Upper90, a finance firm that has been in this field longer than almost anyone. Upper90's CEO Billy Libby financed GPU purchases for a startup called Crusoe back in 2021, which is believed to be the first time advanced chips were ever used as loan collateral. Traditional banks would not touch such a deal at the time because no one knew how quickly chips would lose value. Libby's firm took the risk and built a playbook. That playbook was then copied at enormous scale by CoreWeave, which has since borrowed more than $21 billion using chips as backing and went public on that foundation.
What makes this new deal different is that General Compute is not buying Nvidia chips. It is buying specialized chips from SambaNova, an Intel-backed chipmaker, specifically designed for inference rather than for training models from scratch. Training is the expensive, one-time process of building an AI model. Inference is the ongoing process of using it. These are very different computing jobs, and chips built for one are not necessarily ideal for the other.
SambaNova's SN50 chips claim to deliver AI responses at 600 to 700 units of output per second, compared to around 250 for standard GPU setups. They are also air-cooled, which means they can be installed in ordinary data centers and even repurposed cryptocurrency mining facilities without expensive water-cooling infrastructure. General Compute has ordered $300 million worth of these chips and says it will be the first cloud provider to deploy them at scale.
Upper90's Libby is applying his 2021 logic to 2026 conditions. GPUs are now a well-understood asset class; CoreWeave and others have made that market crowded and expensive. Inference-specific chips from outside Nvidia's circle are still an inefficient market, which is where the opportunity sits. Another infrastructure company, TensorWave, is making a parallel move using AMD chips rather than Nvidia's.
The backdrop here is important. Open-source AI models are no longer a compromise. They have closed the performance gap to within a few months of the most capable closed models, according to analysis from Epoch AI. For most production business tasks, that residual gap is irrelevant. DeepSeek V4 was reported to be fifty times cheaper to run than Claude Opus 4.6 in a May 2026 comparison. GPT-5.5 costs $5 per million input tokens; DeepSeek V4-Flash costs $0.14.
For any business currently paying for AI tools: the question is not whether cheaper infrastructure exists, but whether the tools your team relies on are built on it. If you are paying per query to a premium API for tasks like summarizing documents, routing customer service requests, or drafting routine content, there is a high probability you are paying more than necessary. The infrastructure being built by