The CUDA Problem Nobody Wants to Admit
Here's something most AI infrastructure coverage glosses over: Nvidia didn't win because of chips alone. The H100s and Blackwells are impressive, sure. But the real lock-in is CUDA — Compute Unified Device Architecture.
CUDA lets GPUs, originally designed for graphics rendering, act as general-purpose processors. PyTorch and TensorFlow — the two frameworks that power most of modern AI development — are built on top of CUDA. Developers write in Python, use those frameworks, and their apps run on Nvidia chips by default. That's not just convenience. That's a moat.
Most AI chip startups can't write their own kernels — the low-level software that actually operates hardware. They don't have the talent, the time, or the patience for what's essentially a multi-year engineering grind. So they stay on Nvidia. And Nvidia stays on top.
Infinity is betting that this whole problem can be automated away.
What Infinity Actually Built
Infinity announced Monday it raised $15 million at a $100 million valuation from Touring Capital, Principal VC, and researchers from companies such as OpenAI and Anthropic. The company builds software to make it easier for AI chips to run AI models — specifically, it writes the low-level kernel code that makes non-Nvidia hardware actually usable for inference workloads.
Their product is called Ignition. It's an AI research agent that writes, tests, debugs, and self-optimizes kernel code for chips that aren't Nvidia. Here's what makes it different from the usual "AI writes code" pitch: the system is self-optimizing. It doesn't just generate code once and call it done — it continuously measures hardware performance, then rewrites the code if needed to boost throughput.
It adapts to different chip architectures regardless of proprietary designs. SRAM, GPUs, phone chips, Systolic Arrays — Infinity is building what they call a universal inference library that works across all of them. The result, founder Jeremy Nixon says, is a CUDA-level software stack.
Humans stay in the loop for high-level direction. But the agent handles the grunt work. In one case study Infinity shared, the system compressed what would have taken months or years into hours or days. That's the kind of speedup that makes investors lean forward — not because it's guaranteed at scale, but because if it works even partially, the value proposition is enormous.
The Revenue Model: Skin in the Game
Infinity doesn't charge an upfront license fee. Instead, they take a cut of performance gains and cost savings — measured in tokens per second.
This is smart, honestly. Chip companies don't want to pay for software that might not work. But if Infinity can actually deliver faster inference, they're aligned with their customers' success in the most direct way possible. It's also risky for Infinity — if the agent doesn't deliver measurable improvements, they get nothing. That puts real skin in the game on the product side.
D-Matrix, the AI chip maker positioning itself as an Nvidia challenger, is already a customer. Infinity says it's in talks with other major chip and cloud companies.
Jeremy Nixon and the AGI House Connection
Nixon launched Infinity last year. He's a former Google Brain researcher and the creator of AGI House — though that attribution needs some nuance.
AGI House is a community, VC fund, and applied AI lab in Hillsborough, California. TechCrunch credits Nixon as its "creator," but the site itself lists Rocky Yu as the official founder who took over the lease in early 2023. Nixon was clearly involved in the ecosystem — he's associated with the community and its ventures arm, which offers investments up to $1 million and is backed by Eric Schmidt, Marc Andreessen, Rich Miner, and Adam D'Angelo.
Nixon's background is relevant here. He invented an ML algorithm called Omega that essentially created new machine learning algorithms in a feedback loop — generating and automatically evaluating them. That success got him thinking about applying the same approach to hardware.
His thesis: AI as a "meta technology" enabling "automated invention." He believes automated systems can generate the low-level code needed to run chips more effectively.
Infinity now has 26 employees across design, operations, and engineering. Founded in 2025, they're early but moving fast.
What Principal VC Sees in the Deal
Principal Venture Partners is an early-stage AI-native fund based in Palo Alto. Their thesis, stated plainly on their site: "AI-native companies will win the next era as digital-native companies won the internet era."
They invest in foundation models, dev tools, middleware, applications, consumer tech, and games. Infinity fits squarely into their dev tools category.
The Principal VC angle is worth watching. They're not just writing a check — they're betting that the companies built around AI from day one will outperform those retrofitting legacy systems. Infinity is about as AI-native as it gets: an AI agent writing code for AI hardware.
Touring Capital and the individual researchers from OpenAI and Anthropic round out the round. The researcher involvement is notable — it suggests domain experts who understand the kernel optimization problem are backing this personally, not just through institutional channels.
AI Developer Tools Startups India Investments: The HCL Factor
This raise lands in a market where India's AI developer tools startups are increasingly visible. Companies like Sarvam, which became India's newest AI unicorn with a $234 million round led by HCLTech, show that the ecosystem extends far beyond Silicon Valley. (Read about Sarvam's $234M funding round)
HCL's move into AI datacenter business signals that India's tech services giants are getting serious about infrastructure. Their ₹3,500 crore ($36.5M) full-stack play — owning design, DevOps, cloud ops, and software rather than just renting compute — is the kind of infrastructure bet that creates demand for the developer tools Infinity is building. (See how HCL is shifting to AI datacenters)
India has a massive talent pool for low-level systems programming. The kernel optimization work that Infinity is automating? That's exactly the kind of deep technical work where Indian engineers have historically excelled. If Ignition can compress years of kernel development into days, it doesn't just help chip companies — it changes the economics of who can participate in AI infrastructure.
The implications for India's developer tools sector are real. As chip companies race to build non-Nvidia alternatives, they need software that works across architectures. That's a developer tools problem at its core. And India has been building precisely this kind of infrastructure-focused software talent for decades.
The Bigger Picture
The CUDA problem isn't going away. Nvidia's moat is real, and it's built on years of developer mindshare and ecosystem lock-in. But if companies like Infinity can demonstrate that AI can compress years of kernel optimization into days, they're not just building a product — they're changing the economics of chip development.
Whether that $100 million valuation holds up will depend on whether Ignition can deliver consistent, measurable performance gains across multiple chip architectures. The technology is promising. The timing is right. But the proof will be in production deployments, not case studies.
For now, this is one of the more grounded AI infrastructure raises I've seen. The problem is real, the solution is plausible, and the business model aligns incentives. That's more than you can say for half the seed rounds in this space.
And as India continues to build out its AI developer tools ecosystem — with HCLTech leading the charge on infrastructure and companies like Sarvam pushing the boundaries of indigenous AI models — rounds like Infinity's remind us that the next wave of AI infrastructure innovation could come from anywhere. The software layer is where the real opportunity lies.