Business Insider’s report that AI coding agents are beginning to rewrite the software behind Nvidia’s CUDA platform points to a real change in the AI-chip market: getting a new accelerator to run an AI model is becoming faster. It does not show that CUDA has been recreated, replaced, or made irrelevant. The evidence published so far shows a narrower but still consequential development—agents can accelerate the first, expensive stage of software bring-up for unfamiliar chips.
The distinction is important for Windows developers, enterprise AI teams, and anyone choosing between Nvidia GPUs and alternative hardware. CUDA is not a single program that can be reproduced in a long coding session. It is a mature platform spanning compilers, drivers, libraries, debugging and profiling tools, framework integrations, distributed-computing software, documentation, samples, deployment support, and years of hardware-specific performance work. An agent that produces working matrix-multiplication code has crossed a technical threshold; it has not yet delivered the operational stack an organization needs to move production workloads.
Business Insider frames the story around Infinity founder Jeremy Nixon’s claim that his company used AI coding agents to create CUDA-like software for d-Matrix hardware in 10 hours. Infinity’s own April account of that work gives the necessary qualification: the 10-hour result was for matrix multiplication at scale, a central AI operation, rather than a complete replacement for CUDA. Infinity explicitly says a usable large-language-model inference stack also needs normalization, attention mechanisms, positional encodings, activation functions, vocabulary projection, and orchestration across transformer layers.
That gap is the story. Coding agents are making the first demo easier. The harder work remains making a model fast, reliable, debuggable, secure, repeatable across software versions, and economical across a cluster.
Infinity’s work with d-Matrix is meaningful because matrix multiplication is where much of an AI accelerator’s theoretical performance is either realized or wasted. A chip company that cannot generate high-performance kernels for such operations has little chance of competing in inference, regardless of its hardware specifications. If an AI agent can inspect a new processor, generate code, benchmark it, and iteratively improve it, a startup may reach a credible model demonstration in days rather than spending months building basic compiler and kernel infrastructure.
But the company’s public description also undercuts the broadest reading of the claim. The 10-hour experiment was one stage in bringing Qwen3 inference to d-Matrix’s Corsair hardware. Its own write-up describes full-model inference as a larger job involving many operators and the coordination of work across a model. It does not publish an independent audit of the agent’s code, comprehensive production benchmarks, error rates, maintenance burden, or the engineering time needed after generation.
That is not a minor omission. In AI infrastructure, a kernel that works on a controlled benchmark and a platform that an enterprise can operate are different products. The latter has to survive changing models, quantization methods, framework updates, customer code, networking failures, new chip revisions, memory pressure, security review, and the awkward workloads that are not in a vendor’s demo.
Nvidia has accumulated those layers over roughly two decades. CUDA’s value is partly its programming model, but its practical hold on the market is the surrounding software that has already been tested against millions of developer decisions. A code-generation agent can write an implementation; it cannot instantly create years of production history.
Amazon, Google, AMD, and smaller accelerator makers have not waited for coding agents to begin attacking CUDA’s advantage. AWS operates Trainium and Inferentia through its Neuron software development kit, which integrates with frameworks including PyTorch and TensorFlow. Google’s OpenXLA project provides compiler infrastructure intended to work across multiple machine-learning front ends and hardware back ends. Modular, led by LLVM and Swift creator Chris Lattner, markets its MAX platform for inference across Nvidia and AMD GPUs.
Those projects demonstrate the actual competitive target: portability at the framework and runtime level, while preserving enough hardware-aware optimization that customers do not lose the savings promised by alternative accelerators.
They also reveal an inconvenient reality for buyers hoping to escape vendor lock-in. Moving from CUDA to AWS Neuron, an XLA-based stack, ROCm, or a startup-specific runtime does not eliminate dependence on specialized software. It usually trades one stack, tooling set, and support model for another. AWS customers deploying Trainium and Inferentia still use AWS’s Neuron SDK. A runtime that can target more than one processor can reduce rewrite work, but it cannot erase differences in memory architecture, interconnects, numerical formats, compiler maturity, or collective-communication behavior.
The market may become less Nvidia-dependent without becoming software-agnostic.
Nvidia’s use of coding assistants is more than a defensive talking point. Its GTC material has described a broad internal rollout of AI coding tools, and the company’s engineering sessions acknowledge the same problem competitors face: GPU code needs specialized context, correct measurement, and validation against hardware. Nvidia is unusually well placed to supply that context because it owns the hardware documentation, the compiler stack, the profiling tools, the libraries, and the test environments.
This is where the apparent threat can become an advantage. A coding agent is only as useful as the feedback loop available to it. An agent that can generate CUDA code and automatically profile it with Nvidia’s tooling can search for better implementations quickly. An agent targeting a new accelerator needs equivalent compilers, diagnostics, performance counters, test suites, and hardware access before it can do the same work reliably.
Business Insider quotes former Nvidia-linked software entrepreneur Bing Xu describing verification as the real bottleneck. That assessment fits the available record. Generating code is increasingly cheap; establishing that the code is correct and that it keeps delivering expected performance after the next compiler, framework, or model release remains expensive.
That discrepancy does not prove CUDA adoption is shrinking. It does show why raw ecosystem counts should not be used as a precise measure of lock-in or momentum. The more durable evidence is the software that depends on CUDA in practice: proprietary model pipelines, performance-tuned internal libraries, trained engineering teams, and operational procedures built around Nvidia hardware.
Those are costly to migrate, even if a coding agent can translate portions of the code. Translation is also only the beginning. A converted program must still be profiled, regression-tested, benchmarked against the expected workload, and supported when an incident occurs at 2 a.m.
It does not justify assuming that an alternative chip will run an existing CUDA workload without material migration work. Buyers should ask vendors for end-to-end results on their own model, their serving framework, their batch sizes, their quantization scheme, their latency target, and their failure-recovery requirements. A peak throughput number, or an agent-generated kernel demo, is not enough.
The defensible conclusion from Business Insider’s reporting is that CUDA’s moat is becoming easier to approach at the edges. The moat has not been crossed. AI agents can help challengers turn new silicon into a credible inference platform faster, but they also give Nvidia new ways to extend and optimize the software estate that made its hardware dominant in the first place.
For now, the practical consequence is more competitive hardware trials—not a mass migration away from CUDA.
Business Insider frames the story around Infinity founder Jeremy Nixon’s claim that his company used AI coding agents to create CUDA-like software for d-Matrix hardware in 10 hours. Infinity’s own April account of that work gives the necessary qualification: the 10-hour result was for matrix multiplication at scale, a central AI operation, rather than a complete replacement for CUDA. Infinity explicitly says a usable large-language-model inference stack also needs normalization, attention mechanisms, positional encodings, activation functions, vocabulary projection, and orchestration across transformer layers.
That gap is the story. Coding agents are making the first demo easier. The harder work remains making a model fast, reliable, debuggable, secure, repeatable across software versions, and economical across a cluster.
Infinity’s 10-hour claim is a kernel milestone, not a CUDA replacement
Infinity’s work with d-Matrix is meaningful because matrix multiplication is where much of an AI accelerator’s theoretical performance is either realized or wasted. A chip company that cannot generate high-performance kernels for such operations has little chance of competing in inference, regardless of its hardware specifications. If an AI agent can inspect a new processor, generate code, benchmark it, and iteratively improve it, a startup may reach a credible model demonstration in days rather than spending months building basic compiler and kernel infrastructure.But the company’s public description also undercuts the broadest reading of the claim. The 10-hour experiment was one stage in bringing Qwen3 inference to d-Matrix’s Corsair hardware. Its own write-up describes full-model inference as a larger job involving many operators and the coordination of work across a model. It does not publish an independent audit of the agent’s code, comprehensive production benchmarks, error rates, maintenance burden, or the engineering time needed after generation.
That is not a minor omission. In AI infrastructure, a kernel that works on a controlled benchmark and a platform that an enterprise can operate are different products. The latter has to survive changing models, quantization methods, framework updates, customer code, networking failures, new chip revisions, memory pressure, security review, and the awkward workloads that are not in a vendor’s demo.
Nvidia has accumulated those layers over roughly two decades. CUDA’s value is partly its programming model, but its practical hold on the market is the surrounding software that has already been tested against millions of developer decisions. A code-generation agent can write an implementation; it cannot instantly create years of production history.
The challengers are building their own software moats
Business Insider correctly identifies inference as the area where Nvidia’s position faces the most pressure. Training giant frontier models still favors enormous clusters and tightly coordinated hardware, networking, compilers, and libraries. Inference is more fragmented: organizations can optimize for latency, cost per token, power use, memory capacity, or a particular model family. That gives specialized silicon vendors more openings.Amazon, Google, AMD, and smaller accelerator makers have not waited for coding agents to begin attacking CUDA’s advantage. AWS operates Trainium and Inferentia through its Neuron software development kit, which integrates with frameworks including PyTorch and TensorFlow. Google’s OpenXLA project provides compiler infrastructure intended to work across multiple machine-learning front ends and hardware back ends. Modular, led by LLVM and Swift creator Chris Lattner, markets its MAX platform for inference across Nvidia and AMD GPUs.
Those projects demonstrate the actual competitive target: portability at the framework and runtime level, while preserving enough hardware-aware optimization that customers do not lose the savings promised by alternative accelerators.
They also reveal an inconvenient reality for buyers hoping to escape vendor lock-in. Moving from CUDA to AWS Neuron, an XLA-based stack, ROCm, or a startup-specific runtime does not eliminate dependence on specialized software. It usually trades one stack, tooling set, and support model for another. AWS customers deploying Trainium and Inferentia still use AWS’s Neuron SDK. A runtime that can target more than one processor can reduce rewrite work, but it cannot erase differences in memory architecture, interconnects, numerical formats, compiler maturity, or collective-communication behavior.
The market may become less Nvidia-dependent without becoming software-agnostic.
AI agents could help Nvidia as much as its rivals
The premise that coding automation harms Nvidia assumes challengers gain speed while the incumbent stands still. Nvidia’s public activity points the other way. The company has presented work on AI coding for GPU development, including specialized agents aimed at CUDA optimization rather than generic application programming. It has also published agent-oriented instructions for working with CUDA-X libraries and its broader platform tools.Nvidia’s use of coding assistants is more than a defensive talking point. Its GTC material has described a broad internal rollout of AI coding tools, and the company’s engineering sessions acknowledge the same problem competitors face: GPU code needs specialized context, correct measurement, and validation against hardware. Nvidia is unusually well placed to supply that context because it owns the hardware documentation, the compiler stack, the profiling tools, the libraries, and the test environments.
This is where the apparent threat can become an advantage. A coding agent is only as useful as the feedback loop available to it. An agent that can generate CUDA code and automatically profile it with Nvidia’s tooling can search for better implementations quickly. An agent targeting a new accelerator needs equivalent compilers, diagnostics, performance counters, test suites, and hardware access before it can do the same work reliably.
Business Insider quotes former Nvidia-linked software entrepreneur Bing Xu describing verification as the real bottleneck. That assessment fits the available record. Generating code is increasingly cheap; establishing that the code is correct and that it keeps delivering expected performance after the next compiler, framework, or model release remains expensive.
CUDA’s developer numbers are less clear than Nvidia’s marketing suggests
Nvidia’s software reach is unquestionably large, but even the company’s public developer counts should be treated carefully. Its 2025 annual review said the CUDA ecosystem had more than six million developers. GTC 2026 highlights later described CUDA as having more than four million developers. The materials do not explain whether the difference reflects a changed definition, a revised count, a narrower active-user measure, or simply different promotional language.That discrepancy does not prove CUDA adoption is shrinking. It does show why raw ecosystem counts should not be used as a precise measure of lock-in or momentum. The more durable evidence is the software that depends on CUDA in practice: proprietary model pipelines, performance-tuned internal libraries, trained engineering teams, and operational procedures built around Nvidia hardware.
Those are costly to migrate, even if a coding agent can translate portions of the code. Translation is also only the beginning. A converted program must still be profiled, regression-tested, benchmarked against the expected workload, and supported when an incident occurs at 2 a.m.
What changes for AI infrastructure buyers
The immediate effect of agent-generated low-level software is likely to be a faster evaluation cycle for alternatives. A company considering d-Matrix, Trainium, Inferentia, AMD Instinct, or another accelerator may be able to get a representative model running sooner. That lowers the cost of proving whether a non-Nvidia option can meet a particular latency or cost target.It does not justify assuming that an alternative chip will run an existing CUDA workload without material migration work. Buyers should ask vendors for end-to-end results on their own model, their serving framework, their batch sizes, their quantization scheme, their latency target, and their failure-recovery requirements. A peak throughput number, or an agent-generated kernel demo, is not enough.
The defensible conclusion from Business Insider’s reporting is that CUDA’s moat is becoming easier to approach at the edges. The moat has not been crossed. AI agents can help challengers turn new silicon into a credible inference platform faster, but they also give Nvidia new ways to extend and optimize the software estate that made its hardware dominant in the first place.
For now, the practical consequence is more competitive hardware trials—not a mass migration away from CUDA.
References
- Primary source: Business Insider
Published: 2026-08-03T09:00:02.461000+00:00
Loading…
www.businessinsider.com - Related coverage: developer.nvidia.com
Loading…
developer.nvidia.com - Related coverage: developer.nvidia.com
Loading…
developer.nvidia.com - Related coverage: nvidianews.nvidia.com
Loading…
nvidianews.nvidia.com - Related coverage: nvidia.com
Loading…
www.nvidia.com - Related coverage: businessinsider.es
Loading…
www.businessinsider.es - Related coverage: github.com
Loading…
github.com - Related coverage: nvidianews.nvidia.com
- Related coverage: images.nvidia.com
Loading…
images.nvidia.com - Related coverage: images.nvidia.com
- Related coverage: tomshardware.com
Loading…
www.tomshardware.com - Related coverage: windowscentral.com
Loading…
www.windowscentral.com - Related coverage: nvidia.com
Loading…
www.nvidia.com - Related coverage: techradar.com
Loading…
www.techradar.com - Related coverage: axios.com
Nvidia chips to reap $1 trillion, CEO Jensen Huang says at Nvidia GTC
Nvidia CEO Jensen Huang says he expects the company to reap "at least" $1 trillion in revenue for its newest chips through 2027.www.axios.com