Nvidia spent twenty years building the software empire known as CUDA. It took a one-year-old startup and its AI research agent about ten hours to put the first visible crack in it.
This was not a concept demo or a press-friendly mockup. Infinity’s AI research agent, Ignition, generated a complete low-level software stack for the Corsair accelerator from AI chip company d-Matrix — and pushed it to 92 percent of the chip’s theoretical peak performance. Within ten days, three frontier models, Qwen3, Qwen3.5, and Gemma4, were running end to end on hardware they had never touched before. Work that has historically consumed engineering teams for months, sometimes years, just compressed into hours.
A Crash Course in CUDA, Written by an Agent
Infinity is an AI infrastructure startup founded by a former Google Brain researcher. Its product, Ignition, is an AI research agent built specifically for chip-level software — the unglamorous code that sits between silicon and the models everyone actually runs.
What Ignition does is remarkably concrete. It writes GPU kernels for a target chip, compiles them, runs them, tests them for correctness, measures their performance, and then rewrites the code based on what it learned. Generate, compile, run, test, feed back, revise — a closed loop with no human hands in the middle. Engineers set the high-level direction; the agent grinds through everything else.
In the d-Matrix engagement, Ignition got its first look at the Corsair hardware and, within ten hours, had distributed tensor-parallel matrix multiplication across all 32 of the chip’s compute units at 92 percent of theoretical peak. Ten days later, every layer of three frontier models had been generated from scratch and validated in full. The standard timeline for that work, across the industry, has been months to years.
Infinity has applied the same approach beyond its own customers. The company demonstrated it on DeepSeek’s TileKernels, an operator library written in the TileLang language that similarly cuts down the amount of hand-written low-level kernel code a team needs to produce.
AI Is Commoditizing the Chip Software Stack
To understand why this matters, you have to be clear about what Nvidia’s moat actually is. It is not the silicon. Hardware generations leapfrog each other in a linear race, and rivals routinely ship competitive specs on paper. What keeps customers locked in is the software ecosystem wrapped around CUDA: dozens of math libraries, debuggers, compilers, and deployment frameworks, plus millions of developers who have spent years learning the CUDA programming model.
For a new chip company, that ecosystem is the trap. Designing competitive hardware is hard but achievable. The software is the pit you cannot walk around. Every new chip needs low-level kernels written from scratch, mainstream models adapted, performance tested and tuned through iteration after iteration. And the engineers who can do that work — people who understand both the hardware and the AI stacks on top of it — are scarce everywhere on earth.
Ignition attacks exactly that scarcity. Kernel generation turns out to be a near-perfect task for an AI agent: the objective is unambiguous (correct plus fast), the feedback loop is crisp (compiler errors, performance counters), and the iteration space is enormous (parameter tuning, scheduling choices, algorithmic alternatives). Humans were the bottleneck precisely because the work is tedious, deep, and rare. Agents do not get tired of it, and they do not need to be hired.
The direct consequence is that the cost of building a chip software stack is collapsing. Infinity already books several million dollars in annualized revenue, and its customers are exactly the companies you would expect: new chipmakers desperate for a software story. Instead of hiring a forty-person low-level software team and spending a year, they hand their hardware to Ignition and receive a usable stack days later.
The Three Layers of the CUDA Moat
Break the moat apart and it has three distinct layers.
Layer one is the kernel layer — generating and optimizing compute kernels for a specific chip. It is the lowest, most tedious, most labor-hungry stratum of the stack, and it is precisely where Ignition struck.
Layer two is the libraries and toolchain — cuBLAS, cuDNN, TensorRT, and the debuggers and profilers around them. This layer sits higher, but once kernels can be generated automatically, adapting the upper libraries accelerates dramatically too.
Layer three is the developer ecosystem — millions of CUDA programmers, mountains of documentation, community, training pipelines. This remains the hardest layer to replicate.
What fell in ten hours was layer one. The significance of layer one, though, is that its collapse puts the other two under real pressure for the first time. Once a new chip’s kernels can be produced quickly, the chip can run mainstream models. Once it runs the models, developers have a reason to try it. Once developers try it, a community starts to grow on its own.
The question is no longer whether CUDA can be replaced. It is how cheap replacement is about to become.
Who Wins, Who Worries
The most immediate winners are AI inference chip companies. d-Matrix, Rebellions, Cerebras, and Groq all attack the inference market with genuinely differentiated hardware, and all share the same weakness: no software ecosystem. Tools like Ignition let them leap straight over the build-the-software-first chicken-and-egg problem that has killed or delayed so many challengers.
Inference also happens to be Nvidia’s softer flank. In training, the company is close to a monopoly. Inference is fragmented, brutally cost-sensitive, and full of customers actively looking for a way off Nvidia pricing. Fast software enablement hands those customers a practical exit.
The worrier is Nvidia. Nobody should claim CUDA gets replaced next quarter. But the signal is unambiguous: twenty years of accumulated software barriers are now depreciating on an accelerated schedule, eroded by AI itself. Nvidia’s countermove is its Dynamo inference framework, an attempt to recreate ecosystem gravity on the inference side. The difference this time is that the pace of AI-driven improvement belongs to its opponents.
And the bigger variable is what this says about software engineering itself. Infinity is one data point in a fast-accumulating series: AI optimizations sped up 7-Zip compression by 97 percent without touching its core code; DeepSeek’s TileKernels slashed hand-written CUDA. The direction is consistent — AI agents are commoditizing low-level software work that the industry treated as irreplaceable. When agents can generate, test, and optimize the bottom of the stack, chip competition migrates back from software ecosystems to hardware price-performance. That is precisely the matchup Nvidia has spent two decades avoiding.
Three Things Worth Watching
- Where Ignition goes next. Infinity raised a $15 million seed round and already produces millions in revenue. If the agent keeps compounding, it could become something like a LLaMA Factory for chip software — shared infrastructure that lets any hardware team spin up a working stack in days.
- Accelerated de-Nvidiafication of inference. Through the second half of 2026, watch the inference market’s structure shift. Once software enablement stops being the bottleneck, price, performance, and power decide purchases — good news for cloud providers with custom silicon and for anyone buying tokens at scale.
- Recursive self-improvement, cashing out. From Yuan Dong’s Recursive work to XYZ AI Lab to Ignition, a clear path has emerged: AI helped humans write code, then helped itself rewrite code, and now writes code for chips. Each turn of that loop accelerates the next, probably faster than most people expect.
The CUDA moat is not drained. But for the first time in twenty years, someone else controls the water — and they filled it in an afternoon.
