Nvidia's CUDA elite now manage AI that writes code they can't fully grasp

For Ian Buck, the Nvidia vice president who invented CUDA, the job he created is being reshaped in real time. CUDA engineers, the elite coders who squeeze speed out of Nvidia's chips, have long been among tech's most sought-after specialists. Their work matters because compute costs are massive: an engineer who can squeeze the most value from a single GPU can save a company millions of dollars. Now, AI is taking over much of that painstaking work, generating hundreds of kernels, testing them, and picking the fastest version. The engineers who once wrote those kernels by hand are increasingly managing the AI that writes them for them.

The shift is visible in what CUDA engineers actually do. Traditionally, they spent their days writing kernels, the small pieces of code that tell a GPU how to complete a single job as quickly as possible, then testing each variant to find the fastest one. Increasingly, engineers are supervising coding agents that do this work automatically. In some cases, AI already writes CUDA code that humans can verify but cannot fully understand, said Jeremy Nixon, founder of the startup Infinity, which builds AI that optimizes chip software. Nixon called the ability to outperform human engineers an early glimpse of "superhuman" AI in the real world.

This is part of a broader shift across the software industry, where developers are moving from writing code to managing AI. The job now involves setting goals, checking results, and stepping in when the AI gets stuck, Nixon said. But the work has not become easier, because AI can introduce "bizarre" bugs that a human would never have written, said Anne Ouyang, cofounder of the AI infrastructure startup Standard Kernel. Reviewing AI-generated code, she said, has become "more intense" than reviewing code written by people.

Despite the automation, demand for CUDA engineers remains strong. Overall demand for software engineers is lower than in 2023, but "demand for some specialized skills, like CUDA, has grown," said Elena Magrini, head of global research at the labor market analytics firm Lightcast. There have already been more US job postings requiring CUDA skills through the first eight months of 2026 than during all of 2025, per Lightcast. Nvidia remains the largest employer hiring for these positions, with over 300 active US job postings requiring CUDA skills as of September. Public Nvidia postings advertise base salaries as high as $431,250 before stock and benefits.

Scarcity is still a challenge, even as AI has supercharged productivity. The deepest CUDA expertise was built over more than 20 years, long before AI sent demand soaring, said Bing Xu, founder of the AI optimization startup INT21. "In the past, we couldn't hire enough good-quality CUDA engineers, and now AI is filling the gap," Xu said. His company's research has already shown that AI can outperform human engineers on certain benchmarks for writing kernels, which suggests AI will take on an even greater role over time.

The career implications are uneven. Nixon said AI lowers the barrier to entry for engineers with less experience while simultaneously pushing veterans toward higher-level work, like overseeing AI. Ouyang said AI could hit junior CUDA engineers hardest, while specialists who can outperform AI and verify its work will become more valuable. That creates a two-track future: entry-level coding is being commoditized, while deep expertise plus the ability to audit AI output commands a premium.

For companies that depend on Nvidia's hardware, the lesson is strategic. CUDA roles exist both inside and outside Nvidia, including at AI labs and cloud giants, so the shift to AI-driven kernel writing will ripple across the entire AI infrastructure stack. The painstaking work that made CUDA engineers Silicon Valley's most-coveted specialists has not disappeared; it has been automated and elevated. Executives should expect verification and oversight to become core engineering competencies, and should plan for a future where the most valuable coders are the ones who can manage the AI that codes better than they can.