AMD VP: CUDA's Edge Over Nvidia Is Fading

During the press conference ahead of the annual Advancing AI event in San Francisco, AMD vice president Andrew Dieckman said bluntly that Nvidia's software advantage is losing significance. According to him, the company's customers have almost stopped asking about CUDA, the closed programming platform long regarded as the main foundation of Nvidia's dominance in the AI chip market.

What Dieckman Said

Dieckman admitted that until recently, conversations with AMD customers regularly turned to CUDA, Nvidia's closed programming environment that has tied developers to its hardware for more than fifteen years. Today, he says, the topic has practically disappeared from sales conversations, because companies building AI systems have moved to higher layers of abstraction, where they no longer have direct contact with low-level hardware code.

In practice, this means that tools such as OpenAI's Triton compiler, now built in as the default kernel-generation path in the torch.compile pipeline of the PyTorch framework, can automatically translate code into instructions for both Nvidia and AMD GPUs. A developer writing a machine learning model no longer has to consciously choose between CUDA and ROCm, AMD's open equivalent, because the middleware layer itself decides which hardware the code runs on.

Not long ago I was talking to our customers about CUDA quite a lot - Andrew Dieckman, AMD vice president and general manager of the data center GPU business

ROCm.ai Enters the Picture

Alongside Advancing AI 2026, AMD unveiled ROCm.ai, a new developer platform described as native to agentic artificial intelligence. The tool integrates coding assistants, including Claude, Codex and Cursor, with automatic optimization and deployment of models on AMD hardware. The company says coding agents will now understand the ROCm architecture natively, without needing to manually adapt code originally written for CUDA.

According to AMD's earlier announcements, the new version of ROCm is expected to deliver up to a 3.3x increase in inference performance compared to previous iterations. This continues a strategy the company has pursued for several years, gradually closing the performance and developer-experience gap with CUDA rather than trying to copy it outright.

Why CUDA Still Holds Its Ground

Dieckman's claim doesn't mean Nvidia's advantage has disappeared. Analysts estimate the company still controls the vast majority of the AI accelerator market, depending on methodology somewhere between roughly 75 and over 85 percent, while AMD still holds only single-digit percentages. CUDA benefits from a developer base numbering in the millions and thousands of ready-made, optimized applications, and deeper integration of tools such as cuDNN, TensorRT-LLM and NCCL still gives Nvidia an edge in large-scale model training.

That's why the AMD vice president's remarks are best read as part of the company's communication strategy rather than a neutral market diagnosis. AMD has a direct interest in convincing customers that the software layer is becoming less important, since that is precisely the ground on which Nvidia built its most durable advantage. Full parity, as industry analysts themselves admit, remains years away.

What It Means for AI Hardware Buyers

For companies planning investments in AI infrastructure, including in Poland, the signal is practical: a growing number of development tools now work across the CUDA-ROCm divide, which genuinely lowers the cost of switching to cheaper AMD hardware in places where the lack of compatible software used to be a barrier. That could matter when choosing a cloud computing provider or building in-house GPU clusters, where the price difference between Nvidia and AMD chips can be significant.

At the same time, caution remains warranted. Large production deployments, especially training the biggest language models, still rely in practice almost exclusively on the Nvidia ecosystem, and vendor claims of tool parity don't always translate into real performance results under production conditions.

Advancing AI 2026 also confirms a broader trend: the rivalry between Nvidia and AMD is shifting from chip specifications themselves to the software and developer-tools layer, an area where vendor choice is increasingly being decided by coding agents rather than engineers writing code by hand.