Key Takeaways
- Infinity raised $15M at a $100M valuation from Touring Capital and researchers from OpenAI and Anthropic
- The startup is building kernel software that lets AI models run on any chip, not just Nvidia GPUs
- Infinity's AI agent Ignition writes and optimizes low-level code, reducing months of work to hours
Infinity, an AI infrastructure startup building software to run AI models on non-Nvidia chips, announced a $15 million raise at a $100 million valuation. Touring Capital led the round, with participation from Principal VC and individual researchers from OpenAI and Anthropic. The bet: Nvidia's grip on AI computing comes less from silicon and more from CUDA, its software layer. Break the software lock-in, and competing chips become viable.
Founded by Jeremy Nixon, a former Google Brain researcher, Infinity is building what it calls a universal inference library. The company's core product is Ignition, an AI research agent that writes the low-level kernel code needed to run models on alternative chips. Think SRAM, phone processors, Systolic Arrays, and GPUs from Nvidia competitors.
Why CUDA is so hard to displace
Nvidia didn't win the AI chip market on hardware alone. CUDA, launched in 2007, lets GPUs originally designed for graphics act as general-purpose processors. PyTorch and TensorFlow, the two dominant AI frameworks, are built on CUDA. Developers write in Python, use those frameworks, and their code runs on Nvidia silicon by default.
Most startups building AI applications don't have the expertise or resources to write their own kernels. Porting an app to a non-Nvidia chip means rewriting the low-level software that actually operates the hardware. That's expensive, slow, and rarely worth it when Nvidia's ecosystem just works.
The result is a moat 15 years deep. Nvidia holds an estimated 80-95% market share in AI training chips. Even as AMD, Intel, and a wave of startups produce competitive silicon, adoption lags because the software isn't there.
How Ignition works
Infinity's approach is to automate the hardest part: writing the kernel code. Ignition generates, tests, and debugs low-level software for AI inference on alternative chips. It measures hardware performance with the generated code, then rewrites automatically to improve speed. The system is self-optimizing and adapts to different chip architectures regardless of proprietary designs.
Humans remain in the loop, providing high-level direction. The agent handles the tedious work. In one case study, Infinity claims the agent reduced what would have been a months-long or years-long process to hours or days.
Nixon's interest in this approach came from his earlier work on a machine learning algorithm called Omega, which created new ML algorithms and evaluated them in a feedback loop. He calls it "automated invention." The question he asked: could the same principle apply to hardware? Could AI systems generate the code that makes chips run better?
Business model and traction
Infinity doesn't charge upfront license fees. Instead, it takes a cut of performance gains and cost savings, measured in tokens per second. This aligns incentives: Infinity only earns if customers actually see faster inference.
The company counts D-Matrix, an AI chip maker positioning itself as a Nvidia challenger, as a customer. Nixon told TechCrunch that Infinity is in talks with other large chip and cloud companies. The startup has 26 employees across design, operations, and engineering.
Where this fits in the Nvidia-alternative wave
Infinity is part of a broader push to erode Nvidia's market dominance, product by product. Hardware startups like Cerebras, Groq, and D-Matrix are building alternative silicon. Cloud providers including Google (with TPUs) and Amazon (with Trainium and Inferentia) have their own chips. But software compatibility remains the bottleneck.
Google's recent Frozen v2 chip, which embeds Gemini in hardware for 6-10x efficiency gains, shows the industry's direction: custom silicon paired with custom software stacks. The question is whether a universal layer like Infinity's can bridge the gap for the rest of the market.
Logicity's Take
Infinity's valuation of $100M on a $15M raise signals investor confidence, but the technical challenge is enormous. CUDA's advantage isn't just code. It's 15 years of documentation, tutorials, Stack Overflow answers, and muscle memory in every ML engineer who learned on Nvidia hardware. Infinity's real competition isn't Nvidia's software team. It's inertia. The performance-based pricing model is smart because it shifts risk away from cash-strapped chip startups. But it also means Infinity needs to prove gains quickly. For founders building AI products, this is worth watching. If Infinity delivers, it could finally make AMD, Intel, or custom silicon a real option for inference workloads, potentially cutting GPU costs significantly.
Frequently Asked Questions
What is Infinity building?
Infinity is building a universal inference library and an AI agent called Ignition that writes the low-level kernel code needed to run AI models on non-Nvidia chips.
How does Infinity make money?
Infinity takes a percentage of the performance gains and cost savings its software delivers, measured in tokens per second. There's no upfront license fee.
Who invested in Infinity's $15M round?
Touring Capital led the round, with participation from Principal VC and individual researchers from OpenAI and Anthropic.
What chips does Infinity's software support?
The company says its software works with various chip types including SRAM, GPUs, phone chips, and Systolic Arrays.
Who is Infinity's founder?
Jeremy Nixon, a former Google Brain researcher and creator of the hacker community AGI House, founded Infinity.
Related coverage of custom AI chip development and the push to optimize inference hardware.
Another recent AI infrastructure funding round at a similar stage.
Need Help Implementing This?
If you're a startup evaluating AI infrastructure options or considering alternatives to Nvidia for inference workloads, reach out to Logicity's consulting team for guidance on chip selection and optimization strategies.
Source: Venture Capital News | TechCrunch / Dominic-Madori Davis
Manaal Khan
Tech & Innovation Writer
Produced with AI assistance and reviewed by the Logicity editorial team. Learn more in our Editorial Policy.
