The Startup That Rebuilt Nvidia’s CUDA Software in 10 Hours, Using AI

Nvidia holds roughly 80% of the data centre AI accelerator market. Most people assume that’s because its chips are faster than the competition. The real reason is more interesting — and more vulnerable than it looks. The real reason Nvidia wins is CUDA.

CUDA — Compute Unified Device Architecture — is the software layer Nvidia spent nearly two decades building. It bundles ready-made code and debugging tools. It also lets thousands of chips train a model together. The big AI frameworks — PyTorch and TensorFlow — are built on top of it. Write your application in Python using those frameworks and it runs on Nvidia hardware by default, automatically, without any extra work.

Rival chipmakers — AMD, Qualcomm, AWS — often have hardware that matches Nvidia on raw specs. What they lack is the software. Porting AI models to non-Nvidia chips means writing kernels — the low-level code that operates chips — from scratch. That work takes months or years. Most companies don’t have the engineers to do it. So they stay on Nvidia.

That’s the moat. And a one-year-old startup just used AI to cross it in 10 hours.

What Infinity Did

AI infrastructure company Infinity announced a $15 million raise at a $100 million valuation from investors including Touring Capital, Principal VC, and researchers from companies such as OpenAI and Anthropic.

The company’s product is an AI agent called Ignition. Ignition is itself an AI research agent. It writes the low-level code, tests and debugs it on the hardware, measures how fast the chip actually runs, and rewrites itself when it finds a slower path.

The headline result: when Infinity’s agents got their first access to D-Matrix’s Corsair chip, they hit 92% of its theoretical peak performance in 10 hours and had three frontier models running end-to-end in 10 days. That is not a lab benchmark. That is a commercial deployment on a chip no developer had targeted before.

To put that in context: historically, getting a new chip to that level of performance takes a team of specialised kernel engineers working for months — sometimes over a year. Infinity’s agent did it in 10 hours. Then refined it to production-ready in 10 days.

Why This Matters for the AI Industry

CUDA’s dominance rests on two pillars. CUDA’s first advantage is the software. Its second is everything built on top of it. Millions of lines of company code and workflows make switching to a rival chip slow and costly. Amazon’s own documents once flagged CUDA as a major roadblock to adopting its in-house AI chips. Agents chip away at the first advantage.

If Ignition can rebuild CUDA-equivalent software for any chip in days rather than years, the cost of switching away from Nvidia hardware drops dramatically. AMD, Qualcomm, D-Matrix, and every other would-be Nvidia challenger suddenly has a path to being competitive not just on hardware specs but on the software story that developers actually care about.

With inference workloads projected to account for roughly two-thirds of AI compute spending in 2026, the question Infinity is racing to answer is whether an AI agent can genuinely substitute for the years of manual kernel engineering that have kept Nvidia’s software moat intact — and whether chip makers outside the CUDA ecosystem are willing to bet their production stacks on it.

On the inference side — where AI models are deployed and run, rather than trained — CUDA “is no longer a factor,” said Marshall Choy of Korean chip startup Rebellions. The open-source inference ecosystem is already more competitive than the training side. Infinity is targeting exactly that opening.

The Business Model Is Different Too

Most software companies charge licence fees. Infinity doesn’t. Its business model takes a cut of the speed and cost gains it delivers, rather than a licence fee. If Ignition makes a chip run AI models 30% faster at 20% lower cost, Infinity takes a percentage of that improvement in value.

That’s an alignment-of-interests model that most enterprise software vendors avoid because it requires confidence in your product. Infinity is betting its revenue on actually delivering the performance gains it claims. That’s a credibility signal worth noting in a market full of benchmarks that don’t survive contact with production workloads.

The company is already generating millions of dollars in annual recurring revenue from a single commercial partner, D-Matrix, within its first year. It’s not just a demo. It’s already a commercial deployment.

The Caveats

The story deserves some caution alongside the excitement.

First, the results are currently self-reported. The caveats are real. This is a seed-stage firm with one public chip partner, and the headline benchmarks are self-reported. Independent verification from third-party researchers or a major cloud provider would substantially increase confidence in the claims.

Second, CUDA’s second moat — the millions of lines of existing company code already optimised for Nvidia hardware — is much harder to address. Ignition chips away at the cost of new kernel development. It doesn’t instantly port the last decade of production code that companies have already written for CUDA. That migration work is still significant.

Third, Nvidia is not standing still. The company is actively working to make CUDA more portable and is investing in software tools that make it harder for developers to leave. An 80% market share gives you a lot of resources to defend your position.

What It Means for Developers

If you write software that runs on AI hardware, this is worth watching closely. The practical implication isn’t that you should immediately migrate off Nvidia. It’s that the cost-benefit calculation is changing.

A year ago, running a serious AI workload on non-Nvidia hardware meant committing a team of engineers to months of kernel work before you could even benchmark performance. If Infinity’s claims hold up at scale, that calculus shifts. The question becomes less “can we afford to try AMD?” and more “why wouldn’t we try it?”

Competition in the chip software layer benefits everyone who pays for AI compute — which at this point is nearly every technology company with a product. “The next era of AI will be defined not just by who makes the best chip, but by who can make any chip run state-of-the-art models at blazing speeds,” Infinity founder Jeremy Nixon said. For more on the AI chip wars reshaping the industry, see our coverage of AMD’s $5 billion investment in Anthropic and our breakdown of Qualcomm’s $10 billion Tenstorrent acquisition — both part of the same race to break Nvidia’s hold on AI compute.

Read more tech related articles here.