The Kernel Trap: Is Infinity Building a Moat or Just a High-Tech Service Bureau?
AI-generated image · US National Wire
Opinion: Jeremy Nixon’s bid to break Nvidia’s CUDA stranglehold looks impressive on a pitch deck, but the business model suggests a precarious reliance on the very researchers funding the venture.
In the current gold rush of AI infrastructure, the most valuable real estate isn't the silicon itself—it's the software that tells the silicon what to do. For years, Nvidia has owned this territory via CUDA (Compute Unified Device Architecture), a software layer that transformed GPUs from graphics engines into general-purpose processing power. As TechCrunch first reported, CUDA is the foundation for the industry's dominant frameworks, PyTorch and TensorFlow, creating a gravity well that forces developers to stay within the Nvidia ecosystem because they lack the resources to write their own low-level kernels.
Enter Infinity. Founded last year by former Google Brain researcher and AGI House creator Jeremy Nixon, Infinity is positioning itself as the universal solvent for this hardware lock-in. The pitch is seductive: a universal inference library that allows AI models to run on any chip—be it SRAM, GPUs, phone chips, or Systolic Arrays—effectively automating the replication of state-of-the-art research results.
On paper, the technology is a feat of 'automated invention.' Nixon told TechCrunch that his obsession with meta-technology led him to create Omega, a machine learning algorithm capable of inventing and evaluating other algorithms. He has now applied this logic to hardware via 'Ignition,' an AI research agent designed to write, test, debug, and optimize the low-level code required for inference on non-Nvidia chips. According to the company, Ignition can compress months of human engineering into mere hours or days.
But as a skeptic of the 'platform' play, I find the P&L logic here concerning. Infinity recently raised $15 million at a $100 million valuation from Touring Capital, Principal VC, and a group of researchers from OpenAI and Anthropic. While the funding is a strong signal of technical validity, the composition of the cap table raises a critical question: Is Infinity building a scalable software product, or is it acting as a glorified outsourced compute play for the very researchers funding it?
Consider the business model. TechCrunch reports that Infinity eschews upfront license fees. Instead, the company takes a cut of performance gains and cost savings, measured in tokens per second. In any other sector, this 'gain-share' model is a red flag for a startup trying to scale. It transforms the company from a software vendor into a performance consultancy. Rather than selling a tool that customers use independently, Infinity is essentially selling its own labor—automated via Ignition—and taking a commission on the efficiency it creates.
If the goal is to dismantle Nvidia's moat, you don't do it by becoming a high-end service provider for a handful of chipmakers and labs. You do it by creating a standard. While Nixon claims the result is a 'CUDA-level software stack,' the current customer list—which includes chipmaker D-Matrix—suggests a narrow application. If Infinity is merely the 'easy button' for D-Matrix and other Nvidia challengers to get their chips to actually work, they aren't building a moat; they are building a bridge for someone else's moat.
Furthermore, the human element remains a bottleneck. Despite the promise of the Ignition agent, the company admits that humans are still in the loop to provide high-level direction. With a lean team of 26 employees across engineering, operations, and design, the scaling potential is tied directly to how much 'high-level direction' these 26 people can provide. If the 'automated invention' requires constant human steering to be viable across diverse chip architectures, the business is a linear service play, not an exponential software play.
Nixon’s vision of a self-optimizing system that adapts to any proprietary design is the dream. But in the cold light of the balance sheet, a company that doesn't charge for its software and relies on 'cuts' of savings is a company that hasn't yet figured out how to decouple its revenue from its effort.
Infinity is betting that by automating the 'tedious grunt work' of kernel writing, they can make non-Nvidia hardware viable. That is a noble technical goal. But until they move away from a performance-cut model and toward a scalable licensing or platform fee, they aren't a CUDA killer. They are a highly efficient, AI-powered consultancy for the researchers who signed their checks.