Jeremy Nixon didn't set out to threaten a $3 trillion company. He set out to port some software. It took about 10 hours.
Nixon is the founder of Infinity, a one-year-old startup with $15 million in backing from Touring Capital and a clutch of researchers from OpenAI and Anthropic. His team used AI coding agents to build CUDA-like software for D-Matrix — a Microsoft-backed AI inference chip company — in roughly 10 hours. Ten hours. For a job that, until very recently, required a specialist team grinding for months or years.
That's a story worth sitting with, especially if you've spent any time thinking seriously about what actually keeps Nvidia dominant in AI compute. It isn't the hardware. It's the software. And that software just met AI.
What CUDA Actually Is
Most people who haven't worked in HPC or AI infrastructure think of CUDA as a graphics API that Nvidia repurposed for machine learning. That undersells it by an order of magnitude. CUDA — Compute Unified Device Architecture — is a two-decade accumulation of low-level primitives, optimized kernel libraries, debugging tools, profilers, and critically, an enormous installed base of production code that no one can simply swap out.
Ian Buck, CUDA's creator, now runs Hyperscale and High-Performance Computing at Nvidia. He built a platform so deeply embedded in the ML workflow that when a research team writes a training loop, the CUDA dependency is often invisible — buried in PyTorch internals, in cuDNN, in the way a model checkpoint gets serialized. You don't choose CUDA; you inherit it.
When I was building out hosting infrastructure in the early GPU compute era, we used to joke that switching GPU vendors was like switching CPU architectures — technically possible, practically unthinkable. The software surface area was just too large. And unlike a CPU architecture switch, there was no compatibility layer waiting on the other side.
This is the moat. It's not the chip. It's the everything else.
Why Inference Changes the Calculus
Here's the thing: CUDA's stickiest grip has always been on training. Model training involves custom CUDA kernels, fused operators, gradient checkpointing logic, and distributed training frameworks that have been painstakingly tuned against Nvidia hardware. Reproducing that on AMD ROCm or Google TPUs is months of work per model family.
But the industry has moved. The era of training frontier models from scratch on proprietary clusters is maturing into an era of inference at scale. And inference is a different beast.
Marshall Choy at Rebellions put it bluntly: inference-focused CUDA adoption is "no longer a factor." What he means is that when you're running a production serving stack — fielding millions of inference requests, optimizing for time-to-first-token and decode throughput — you're not calling the same deep CUDA primitives you need for training. The inference workload is narrower, more standardized, and increasingly abstracted behind runtimes like vLLM, TensorRT-LLM, or SGLang. Those runtimes, in turn, can be retargeted.
This is the gap that companies like Infinity are pouring into. And the question they're answering is: what happens when AI coding agents can write the software that closes that gap?
The Omega Algorithm and the 10-Hour Build
Infinity's core technology is something they call the Omega algorithm — a machine learning system that generates and evaluates new algorithms through automated feedback loops. Think of it as an AI that writes low-level hardware code, tests it, scores its own output, and iterates. Nixon's team pointed this at D-Matrix's Corsair inference accelerator and produced working CUDA-equivalent software in about 10 hours.
D-Matrix's Corsair is worth understanding here. It's not a GPU in the traditional sense. It's a purpose-built AI inference accelerator built on TSMC's N6 process, using SRAM-based in-memory compute with LP-DDR5 memory — a design optimized specifically for the decode phase of autoregressive inference. When independent tests paired Corsair cards with GPUs, they saw a 24-second baseline response drop to under two seconds: roughly a 10x speedup. The platform entered full production in June 2026, shipping to priority hyperscalers through the summer.
The problem Corsair had before Infinity's intervention is the same problem every alternative AI accelerator has: if you can't easily run existing ML software on it, operators won't touch it. The software integration cost has historically been prohibitive. What Infinity demonstrated is that AI coding agents can dramatically compress that cost — from a year-long engineering effort to a morning's work.
Infinity's universal inference library is designed to work across different AI chips, dynamically generating and optimizing the low-level code that makes models run efficiently on any target hardware. If it scales the way Nixon believes it will, the software barrier to deploying non-Nvidia chips collapses.
The Skeptics Have a Point
I don't want to be glib about this, because the counterargument is real and it's coming from someone whose opinion I respect.
Chris Lattner, co-founder of Modular (and creator of Swift, LLVM, and MLIR), has been outspoken that the hype around AI-generated chip software is "very overblown." His point: code generation is only a fraction of what goes into production software. Verification, debugging, performance characterization, integration testing — these are where teams actually spend their time, and AI coding agents don't automate any of that away.
Bing Xu frames it as a moat shift rather than a moat collapse: "Verification is the biggest bottleneck." The team that figures out how to automate correctness verification of generated low-level hardware code will own the next competitive advantage. That's not a solved problem.
Both of them are right, in the sense that Infinity's 10-hour demo is a proof of concept, not a shipping product at scale. The question is whether the gap between proof-of-concept and production-grade closes in 18 months or five years. Based on how fast AI coding capabilities have improved since 2024, I'd bet on the shorter timeline.
Nvidia Isn't Standing Still
In December 2025, Nvidia quietly acquired SchedMD — the company that develops Slurm, the dominant job scheduler for high-performance computing. Slurm runs more than half of the top 10 and top 100 systems on the TOP500 supercomputer list. If you've ever run a large-scale training job on a cluster, you've almost certainly touched Slurm.
Nvidia called Slurm "part of the critical infrastructure needed for generative AI." Which is true. But the timing of the acquisition is interesting. Nvidia bought SchedMD right as alternative accelerators were beginning to show production viability. What they got isn't just a workload scheduler — they got a chokepoint in the infrastructure stack, and they get to decide how quickly non-Nvidia hardware gets first-class support.
Daniel Newman at Futurum Group read this clearly: "Nvidia just deepened the CUDA moat." The layer above the GPU — the job scheduler that decides which workloads run where — is now Nvidia-controlled. Even if your chip is better for a specific workload, and even if Infinity can write your software in 10 hours, if Slurm routes jobs preferentially to CUDA-native hardware, you've got a distribution problem.
Nvidia said it would continue developing Slurm as vendor-neutral open-source software. That commitment is only as durable as the strategic incentive that produced it.
What This Means If You Run Real Systems
If you're operating AI infrastructure today — whether that's a cloud provider, an enterprise ML platform, or a team running inference for a product — the CUDA moat calculus is changing faster than most procurement cycles can track.
Here's how I'd think about it practically:
- For new inference workloads, the case for evaluating non-Nvidia hardware has never been stronger. D-Matrix's Corsair is in production. AMD's MI450-series GPUs are shipping into AMD Helios at 2-gigawatt scale via the Anthropic deal announced in July. The software gap is shrinking. If you're standing up a new inference cluster, it's worth doing a real evaluation instead of defaulting to H100s or B200s on autopilot.
- For existing training workloads, the moat is still real. If you have a codebase built on CUDA-specific kernels and custom operators tuned against Nvidia hardware, you're not migrating that next quarter. You shouldn't feel bad about that — it's a rational position. Just make sure you're building new code in ways that don't add to the lock-in unnecessarily.
- Watch the Slurm situation closely. The SchedMD acquisition is the most underreported story in AI infrastructure right now. Open-source job schedulers are foundational to how multi-tenant clusters work. If you have clusters running Slurm, track how Nvidia integrates it. The integration roadmap will tell you something important about their long-term intentions.
- The AI coding agent story isn't just about chips. The same dynamic Infinity demonstrated — AI coding agents compressing months of integration work into hours — applies broadly to platform engineering. Legacy systems that seemed too costly to integrate or modernize are going to start looking like weekend projects. This changes the economics of technical debt dramatically.
The Meta Point
What's striking about the Infinity story isn't just the speed. It's that the thing making CUDA sticky — a massive software surface area that took years and thousands of engineers to build — is now the kind of thing AI coding agents can approximate in 10 hours. The moat that was supposed to be self-reinforcing (more adoption → more CUDA code → higher switching cost → more adoption) is encountering a technology that doesn't care how long it took to build the original.
I've spent a lot of time over the years in the unsexy infrastructure layer that makes everything else possible. I've watched vendor lock-in get constructed over decades and unmade in months when the switching cost finally crossed a threshold. That moment for CUDA in inference may be closer than the confident analysts think.
Nixon raised $15 million in July. His company is one year old. He just built CUDA-equivalent software in 10 hours. Whatever Nvidia's answer to that is, it had better be faster than the next funding round.