Last week the AI and silicon worlds shared one detonation: Infinity, a software company barely a year old, used AI coding agents to build a CUDA-class stack for chip startup d-Matrix in ten hours. The conversation instantly split into two camps — one chanting that Nvidia is finished, the other dismissing the whole thing as a demo.
Both camps are wrong, and the truth is considerably more interesting than either slogan. Here is what actually happened, what it threatens, and what it leaves untouched.
What Ten Hours Actually Bought
Infinity's product is an AI research agent called Ignition, and it works the way a human kernel engineer works — just absurdly faster. The loop has four beats: generate GPU kernel code tuned to the target chip's architecture, run tests to confirm correctness, measure performance against the hardware's ceiling, then rewrite based on the feedback. Write, compile, run, test, optimize — repeat without coffee breaks.
On first contact with d-Matrix's Corsair chip, Ignition reached 92 percent of the chip's theoretical peak within ten hours, then got three frontier models — Qwen3, Qwen3.5, and Gemma4 — running end to end within ten days. For calibration: writing low-level inference software for a new chip has historically taken months to years. Investors responded to the compression accordingly: a $15 million seed round at a $100 million valuation.
What CUDA's Moat Actually Looks Like
People treat CUDA as a programming language. It is really an entire software platform wrapped around Nvidia GPUs, and it comes apart into three layers:
- The bottom layer: kernel compilers and runtimes — the code that makes silicon actually compute.
- The middle layer: heavily optimized libraries like cuBLAS and cuDNN, plus the debugging and verification toolchains around them.
- The top layer: frameworks like PyTorch and TensorFlow, oceans of accumulated enterprise code and workflow, and a developer ecosystem numbering in the millions.
A useful analogy: an Nvidia GPU is not a machine tool, it is a factory. The bottom layer tells every machine what to do, the middle layer supplies finished tooling, and the top layer is a production system that has been running for years. Customers are not buying a card — they are buying a factory that works the day it arrives. That is why Nvidia's hardware can be chased, while its customers stay put: they contracted for a system, not a chip.
So which layer did Ignition replicate in its ten hours? Essentially the bottom one. Ignition's strength is generating and optimizing inference kernels for arbitrary chips and assembling low-level software capability around them, and the company is developing a universal inference library that targets many chips at once. That is genuinely valuable — and it is not a reproduction of CUDA's twenty-year ecosystem.
You do not need to replicate all of CUDA to raise $15 million. That sentence is the real headline, and it says more about the market than either camp noticed.
The Real Battlefield Is Inference, Not Training
Here is the subtlest part of the story. AI compute splits into training and inference, and on the training side CUDA remains close to unassailable — months-long runs on tens of thousands of cards punish any instability, and enterprises buying training platforms are maximally conservative.
Inference flips the table. Marshall Choy, chief commercial officer at Korea's Rebellions, states it without decoration: "On the inference side, CUDA is no longer the deciding factor. This will be a competition of open-source software." The reason is economic. Training optimizes for absolute performance — huge clusters, months of lockstep, zero tolerance for failure. Inference optimizes for cost per token — small clusters, even single cards, whatever runs cheapest wins.
And inference software can run across chips. A framework that supports many accelerators lets users swap hardware without rewriting code — and at that moment, CUDA's core locking mechanism stops working.
The stakes are enormous. In 2026, inference is projected to consume roughly two-thirds of all AI compute spending, while Nvidia still controls about 80 percent of the datacenter AI accelerator market — a dominance built far more on software lock-in than on silicon superiority. AMD, Qualcomm, and AWS all ship chips that look strong on paper, and d-Matrix's Corsair does real work today. The industry has never lacked good chips; it has lacked good software to wield them. Coding agents are rewriting that equation.
Three Fronts Closing on CUDA
Infinity is one column of a larger assault. At least three independent lines now converge on the same fortress.
Front one: agents that generate the low-level software themselves. This is Infinity's path — AI agents that write, test, and optimize GPU kernels, compressing a new chip's software enablement from years to days. The essence is solving the cold-start problem: if you cannot hire a mature software ecosystem into existence, generate one.
Front two: high-level languages that replace hand-written CUDA. DeepSeek recently open-sourced TileKernels, a high-performance GPU operator library built on TileLang and containing 44 production-grade operators. TileLang — developed by a Peking University team — is a Python-embedded GPU DSL positioned between Triton and raw CUDA: more control than Triton, dramatically more productive than hand-writing kernels. The sentence that matters: TileKernels contains not one line of CUDA C++. TileLang's backends already cover NVIDIA, AMD's MI300X, and Huawei Ascend, so code written today migrates to non-Nvidia hardware without a rewrite. DeepSeek founder Liang Wenfeng has said publicly that coding agents plus TileLang make building AI software dramatically simpler.
Front three: the giants building their own stacks. Google's TPU, Amazon's Trainium and Inferentia, Microsoft's Maia — each comes wrapped in a software ecosystem the company had to build itself, at enormous cost. Amazon's internal documentation has reportedly listed CUDA as the primary obstacle to adopting its own AI chips. The giants' motive is not ideological; it is escape velocity.
The Moat Is Migrating, Not Vanishing
Declaring CUDA cracked is still premature, and the soberest voices come from people who worked inside the fortress. Bing Xu, founder of INT21 and a former Nvidia engineer, puts the constraint precisely: "An agent can generate a mountain of code in an hour, but verification is the real bottleneck." Producing ten thousand lines is fast; proving those lines are correct and stable across every boundary condition is slow. CUDA's deepest asset is precisely its verification toolchain — and Xu predicts that in the agent era, that verification ecosystem becomes CUDA's next moat.
Chris Lattner, Modular's cofounder, adds three buckets of cold water: writing code is a small fraction of software engineering, and production-grade optimization is what actually determines chip performance; chip software is a small, elite field whose public code is thin, so the training data available to AI is limited; and migrating millions of lines of existing enterprise code is an organizational decision, not a technical one.
There is also a counterattack already underway: Nvidia is using AI too. Ankit Patel, the company's VP of developer ecosystem, says Nvidia deploys coding agents to build CUDA faster and validate it at larger scale — more than 30,000 engineers inside the company now use dedicated AI coding tools, and code output has tripled. Xu's summary is the cleanest framing available: this fight turns on whether the challengers' catch-up speed can exceed Nvidia's self-iteration speed.
What Practitioners Should Take From This
- Inference cost is the new front line. As training converges and model capabilities cluster, whoever pushes cost per token lowest holds the ticket to the next application wave. Agents accelerating chip software ecosystems is, at bottom, inference-cost engineering.
- "AI building AI" stopped being a slogan. Infinity and DeepSeek both show AI participating in constructing AI's own infrastructure. That self-reinforcing meta-capability may upend the industry faster than any single technical breakthrough.
- Never treat a moat as static. CUDA's defense is migrating from accumulated code to a verification and optimization ecosystem. Whoever occupies the new ground first wins next — a strategic signal for anyone building AI infrastructure or tooling.
- The cost-decline chain still runs. DeepSeek V4 Flash's official release pushed inference pricing to rock bottom; agent-built agents now cost as little as a few cents per build on frameworks like PenguinHarness; and now even chip-enablement software is being compressed by AI. Each layer of falling cost paves the road for the next explosion of applications.
Nvidia's CUDA moat will not disappear this year or next. But its shape is changing — from a barrier built on writing code to a barrier built on verifying it.
And the signal that matters most to everyone else in the industry may be this: AI is becoming the builder of its own infrastructure. That is the real butterfly effect.