Upper90 Lends $400 Million Against AI Inference Chips Instead of Nvidia GPUs - Startup Fortune

**Skyler Monroe** (0:10)
Hey everyone, welcome back to The AI Hardware Show, the podcast where we dig into the chips, silicon and data center tech that's making AI actually happen. I'm Skyler Monroe, your host and resident hardware nerd. Big thanks to today's sponsors, AI EDA, helping businesses integrate AI into their real world workflows. Ago Consulting, that's Ago, your go-to for silicon development from AI to so. And Zen Semiconductor, the AI and Silicon venture group behind some seriously cool stuff like their Sierra RISC-V CPUs and Loom AI fabric. All right, we've got a packed episode today. We're talking alternative chip financing, Nvidia's new RTX Spark Superchip, Intel teasing a brand new data center GPU, and TSMC posting absolutely mind-blowing revenue numbers. Let's get into it. Okay, first story, and this one is fascinating from a financial angle, not just a hardware angle. A firm called Upper90 has put together a $400 million lending facility. But here's the twist. Instead of using Nvidia GPS as collateral like pretty much everyone else in this space has been doing, they're lending against AI inference chips, other chips, non-NVIDIA chips. And that is a pretty significant statement about where the market is heading. So let me back up a second and explain what's going on here, because this is actually a really interesting intersection of finance and hardware. Over the last couple of years, we've seen a whole cottage industry emerge around GPU-backed lending. The basic idea is that Nvidia GPUs, especially H100s and now B200s, hold their value really well. They're in massive demand. So lenders have been treating them almost like hard assets, like real estate or equipment, and lending money against them. Cloud startups and AI companies have been using this to get capital without giving up equity. But Upper90 is doing something different. They're saying, hey, there are other inference chips out there, chips from companies like Syribris, Grok, TenStorent, or maybe Custom Silicon from Hyperscalers. And those chips also have real value, real deployment demand, and can serve as legitimate collateral. That's a meaningful vote of confidence in the broader AI chip ecosystem beyond just Nvidia.
Think about it this way.
If you're a bank and you're willing to lend against a piece of equipment that tells you something important, you believe that equipment has stable, recoverable value. If the borrower defaults, you can go take those chips and sell them or lease them.
Upper90 lending $400 million against non-NVIDIA inference chips means they believe those chips are liquid enough, valuable enough, and in demand enough to back that kind of capital. And inference is really the key word here. We talk a lot about training chips. The massive GPU clusters used to train frontier models. But inference is where a deployed model actually runs and generates outputs for users. That's the workload that scales with adoption. As more people use AI products, inference demand just keeps climbing. So, it makes sense that inference-specific silicon would start to develop the kind of market value that makes lenders comfortable. This also signals something broader about market maturity. When alternative chip vendors start getting recognized by financial institutions as legitimate collateral, not just experimental tech, that's the ecosystem growing up. It means the AI chip market is diversifying beyond one dominant player, at least in the eyes of sophisticated capital allocators. And that's a really healthy sign for competition and innovation in this space. Keep an eye on Upper90 and similar firms. They're basically placing a $400 million bet that the AI chip world has room for more than one winner. All right, moving on to story 2 And this one hits close to home for anyone who loves seeing big ideas crammed into small packages.
Nvidia's RTX Spark, their super chip for AI PCs, is getting closer to launch. And they've just released the first preview developer drivers with native Windows on ARM support. This is a big deal, and I want to explain why. So first, what is RTX Spark? It's Nvidia's take on a tightly integrated system on a chip design for the edge, think laptops and mini PCs. It combines the Grace CPU architecture, which is Nvidia's ARM-based CPU, with a Blackwell GPU, all on a single chip connected by high-speed fabric and paired with 128 gigabytes of unified memory. All of that on one chip in a laptop form factor.
Now, if you've been following Nvidia's Grace Hopper Superchip, that's the big data center chip that combines Grace CPU and Hopper GPU, RTX Spark is kind of like that concept's little sibling, optimized for edge AI rather than the data center. The key insight is the same though, when you put the CPU and GPU on the same die or package and give them shared high-bandwidth memory access, you eliminate a ton of latency and bandwidth bottlenecks that you'd normally see when data has to travel between separate chips over PCIe. The 124 if 124 AI workloads, one of the biggest limitations for running large language models locally is memory. You need enough to hold the model weights in memory while also processing context. Most current consumer laptops top out at maybe 32 or 64 gigabytes, and that memory is split between CPU and GPU.