Leadership1 publisher2 min readPublished
CUDA engineers now supervise the AI that writes their GPU kernels
CUDA engineers, among tech's most sought-after specialists, increasingly supervise AI that writes and tests hundreds of GPU kernels. At base salaries Nvidia advertises as high as $431,250, employers are now paying for judgement about code the engineers did not write.
The Board Room · Leadership desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Jeremy Nixon, founder of chip-software startup Infinity, said the supervising job means setting goals, checking results and stepping in when the AI gets stuck.
- Anne Ouyang, cofounder of Standard Kernel, said AI introduces "bizarre" bugs that humans would not write, making code review "more intense."
- Nixon said AI sometimes writes CUDA code that engineers cannot fully understand, though they can still verify that it is correct.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- decision Once the AI generates and ranks hundreds of candidate kernels, counting kernels written or speedups found stops measuring the engineer; reviews have to score the goals set and the errors caught.
- exposure When engineers sign off on kernels they cannot fully explain, the reviewer is the last check before production, so a weak reviewer ships code nobody on the team can account for.
- contradiction Nixon says AI lowers the barrier for less-experienced engineers while Ouyang says juniors could be hit hardest, so an entry-level hiring plan built on either founder's view is a bet.
Compute is the reason employers pay for this skill. CUDA is Nvidia's software for programming its GPUs, and an engineer who gets more work out of each chip can save a company millions [2][3]. The roles sit at Nvidia and also at AI labs and cloud providers [18].
A finance chief would ask why that pay holds once the AI writes and tests the kernels. The hiring record answers her. US postings requiring CUDA skills in the first eight months of 2026 outnumbered all of 2025, so this year's monthly pace is at least 1.5 times last year's average [1]. Nvidia, the largest employer hiring for the skill, had more than 300 active US postings as of September [9]. Elena Magrini, head of global research at Lightcast, said overall demand for software engineers is below its 2023 level, but "demand for some specialized skills, like CUDA, has grown" [7].
Bing Xu, founder of the AI optimization startup INT21, described the tools as covering for hires companies could not make. "In the past, we couldn't hire enough good-quality CUDA engineers, and now AI is filling the gap," Xu said [12]. The deepest CUDA expertise, he said, was built over more than 20 years, long before AI drove demand up [11].
That history creates a sequencing problem. Anne Ouyang of Standard Kernel expects specialists who can outperform the AI and verify its work to become more valuable [19]. I'd expect verification to draw on what senior engineers learned by writing and testing kernels themselves, which is the work the agents now do [4]. A team that hands all of it to agents this quarter gets faster output now. It also decides, by default, how many people it will have a few years out who can tell a correct kernel from a fast, wrong one. We do not know yet whether engineers who learn the trade by supervising agents build the same judgement. The reporting does not include evidence either way.
The forward-looking claims come from people with a stake in the answer. Nixon, Ouyang and Xu each founded a startup that builds AI infrastructure or optimization software [5][6][11]. Xu's finding that AI beats human engineers on certain kernel-writing benchmarks comes from his own company's research [15]. Nixon called that ability to outperform humans an early glimpse of "superhuman" AI in the real world [17]. The demand figures are Lightcast's [7][8].
What to watch
- Lightcast's full-year 2026 count of US postings requiring CUDA skills, and whether Nvidia's active listings stay above 300.
- Whether entry-level CUDA postings rise or fall as agent tooling spreads, which would test Nixon's view against Ouyang's.
- Independent benchmarks, outside INT21's own research, comparing AI-generated kernels with human-written ones.