The AI landscape is changing, not because of new models or bigger labs, but because of how companies are financing and deploying their AI workloads. In a move that could reshape the economics of AI, General Compute has secured a $400 million loan from Upper90, with the unusual collateral: inference chips. These aren’t training chips. They’re designed to run already-trained AI models efficiently, and that’s exactly what’s becoming more valuable as AI costs balloon.
The chips in question are SambaNova’s SN50, which promise 16x faster inference than what you’d get from standard GPU-based cloud providers. That’s not just speed, it’s efficiency. These chips are power-efficient and don’t require the expensive cooling systems that GPUs demand. That means faster deployment, lower energy bills, and less infrastructure overhead. For startups and enterprises alike, this is a compelling alternative to the high-cost, high-compute models that dominate the headlines.
This deal isn’t just about financing, it’s about a market shift. As AI adoption grows, so does the cost of running it. Training new models, especially large language models, is expensive, resource-intensive, and often requires access to frontier lab infrastructure. But inference, the act of running those models to generate output, is where most businesses actually need to spend their compute. And inference is where chips like the SN50 shine.
The SN50’s performance advantage is real. SambaNova’s chips are built for precision and speed in inference tasks, not for training. That’s why they’re so much more efficient. They’re not trying to outperform GPUs at training, they’re trying to outperform them at inference. And in that domain, they’re winning.
This shift is also about open-source infrastructure. As AI becomes more democratized, companies are looking for ways to deploy models without relying on proprietary, expensive platforms. Inference chips like the SN50 are part of that movement, they’re designed to work with open-source frameworks and models, making them more accessible to a wider range of developers and businesses.
It’s worth noting that this isn’t the first time we’ve seen a shift toward inference. In fact, many companies have quietly moved away from training-centric AI stacks toward inference-first architectures. But this $400 million deal is a clear signal that the financial world is starting to take inference seriously, and that’s a big deal.
As AI costs rise, businesses are looking for cheaper, more scalable ways to deploy models. Inference chips offer that. They’re faster, more efficient, and more cost-effective than traditional GPU-based infrastructure. That’s why companies like General Compute are betting on them, and why financiers like Upper90 are willing to back them with real money.
This isn’t just about hardware. It’s about economics. It’s about making AI deployment more accessible, more scalable, and more affordable. And as more companies follow this path, we’ll see a new kind of AI infrastructure emerge, one that’s not defined by training power, but by inference efficiency.
As first reported by TechCrunch, this deal could be the tipping point for a new era of AI deployment, one that’s more practical, more scalable, and more cost-effective than ever before.
For those building workflow automation without losing control, this is a critical development. Inference chips offer a way to deploy AI models without the overhead of training, and that’s exactly what many automation platforms need to scale efficiently. As we’ve discussed in our post on Building Workflow Automation Without Losing Control, the key to scalable automation is not just speed, it’s cost efficiency. And inference chips are helping make that possible.