For many years, engineers specialized in optimizing CUDA for Nvidia chips have been among the most sought-after professionals in the tech industry. Nowadays, they are increasingly managing AI that undertake this work on their behalf.
These engineers specifically use a unified computing device architecture ( CUDA ) – Nvidia to program their AI chips, which is the software for GPU – in order to write code that can run AI as efficiently as possible.
The reason it is so favored is that the computational power cost is extremely high; engineers who are able to extract more value from a single chip can save a company millions of dollars.
In the past, engineers would spend time writing kernel, which are those small segments of code that instruct GPU on how to complete a single task as quickly as possible, and then they would repeatedly test them to find the fastest version. Nowadays, AI is taking over a large portion of this tedious work: it generates hundreds of kernel, conducts tests, and selects the best solution.
As a result, more and more engineers CUDA are spending their time supervising coding agents. This is also part of a broader transformation within the entire software industry: developers are shifting from writing code to managing AI.
AI, co-founder of infrastructure startup Standard Kernel, Anne Ouyang stated that this means engineers need to set goals, check results, and intervene when there are bottlenecks in AI.
She said that since AI may introduce "strange" vulnerabilities that humans would not be able to write, the work of reviewing the code of AI has become "more intense."
Despite this change, recruitment data still shows that the demand for CUDA engineers remains strong. Lightcast, a labor market analysis company, and its global research director, Elena Magrini, stated that although the overall demand for software engineers is lower than in 2023, "the demand for certain specialized skills, such as CUDA, has grown."
AI Optimizing startups, INT21 The founder Bing Xu stated that even though AI has significantly increased productivity, scarcity remains a challenge. This is because the most profound CUDA experience is accumulated over more than 20 years, long before AI led to a surge in demand.
"In the past, we couldn't recruit enough high-quality CUDA engineers, but now AI is filling this gap," said Xu.
CUDA The role of engineer is evolving.
Engineer positions with the code CUDA exist both internally and externally at Nvidia, including in AI laboratories and among cloud computing giants. According to data from Lightcast, in the first 8 months of 2026, the number of job postings requiring CUDA skills in the United States has already exceeded that of the entire year of 2025.
The institution stated that Nvidia remains the largest employer for recruiting such positions, and as of September, there were still over 300 active job vacancies in the United States requiring CUDA skills. Nvidia publicly announced CUDA related engineering positions, with base salaries reaching up to $431,250, in addition to stock and benefits.
Nixon indicates that AI has lowered the entry barrier for engineers with less experience, while also pushing senior engineers to higher-level positions, such as supervising AI.
Ouyang indicates that AI may have the greatest impact on junior CUDA engineers, while those experts who can surpass AI and validate their work results will become more valuable.
Over time, AI may take on an even greater role.
Xu said that research conducted by his company has shown that, on certain benchmarks developed by kernel, the performance of AI can surpass that of human engineers.
Nixon indicates that, in some cases, AI is already capable of writing CUDA code that engineers cannot fully understand, although engineers can still verify its correctness.
He said that this ability, which surpasses that of human engineers, provides a preliminary glimpse of the “superhumans” in the real world AI.
Have a tip? Contact this reporter via email at gweiss @ businessinsider.com or Signal at @ geoffweiss.25 . Use a personal email address and a nonwork device ; here 's our guide to sharing information securely .
