China's AI pay pyramid: interns at ¥5,000 a day, stars at ¥100M
Two numbers out of China this morning show what the AI buildout is actually costing — one in salaries, one in steel and lithium. Neither looks like a market that's slowing down.
A frontline algorithm researcher in China now costs ¥3 million a year, and the interns above him cost ¥5,000 a day. That's the base of a pay pyramid 36Kr mapped out this week: roughly 2–3 million yuan for ordinary researchers — the baseline was just over 1 million last year — 5–6 million for star fresh graduates at ByteDance's TopSeed, Alibaba's AliStar and Tencent's equivalent, 10 million-plus for the 10–20 industry names at each big lab, and nine-figure packages for the handful of people at the very top. PhD interns on core directions at ByteDance, Tencent and Alibaba now pull 5,000–6,000 yuan a day; two years ago the average PhD intern monthly salary was 10,000 yuan, which is two days of today's pay.
The reason is that the scarce skill is expensive to get wrong. Researchers call the full set of training decisions the "recipe" — model size, data mix, training path, the tricks — and a single wrong run on a trillion-parameter model burns tens of millions of yuan in compute, far more than the annual salary of the people who could have avoided it. The number of people who have actually commanded a ten-thousand-GPU training run is tiny, and they only exist inside a handful of companies. That's why ByteDance now runs redundant "saturated staffing" — two parallel teams per core research area, so losing one whole team to a poacher doesn't leave a hole — and why headhunters describe rivals mapping a competitor's ten most important people and working them one by one, with funding their startup as Plan B if they won't switch.
Our read: this is the clearest signal yet that compute isn't the binding constraint on frontier AI — people are. You can raise $13 billion for a data center in six weeks; you cannot manufacture a training commander on that timeline. The tell is how companies are defending: ByteDance created "Doubao shares" issuable only to AI staff, with set pricing and buyback terms, after nearly 70 key technical people left Seed last year. When compensation starts being structured like equity in the model itself, the labour market has already repriced.
US battery installations hit a record 20.2 gigawatt-hours in Q2 2026 — and data centers are now the biggest customer behind the meter. Utility-scale projects accounted for 18 of those 20.2 GWh, with Arizona (6.2 GWh), Texas (3.8) and California (3.6) leading, and the US is on pace for 71 GWh this year, up 20% from 2025. In the commercial segment, roughly three-quarters of new behind-the-meter batteries went to data centers — developers are increasingly pairing storage with on-site or co-located generation because the grid can't deliver capacity fast enough.
The residential side went the other way: home installations fell 27% year over year to 657 MWh, mostly because a tax credit that subsidized home batteries expired in 2025, and MIT Technology Review reports a projected 16% drop for the full year. So the storage boom is now almost entirely an industrial story — and increasingly an AI story — while the household market waits for a policy that ended.
Stanford's Chris Piech is recruiting 1,000 volunteers to teach probability for AI, one teacher per ten students. The free course, Probability for AI, starts October 9 with applications due at the end of September; more than a thousand people have already applied to teach, and Piech says an alum's funding covers the tools and servers. It's an unusually direct answer to the pipeline problem above: the industry is bidding salaries up because the talent pool is thin, and this is one of the few efforts trying to widen the bottom of the funnel rather than just raid it.
Would a pay ceiling on AI researchers slow the race, or just push it underground? Tell us in the comments.
Sources: 36Kr DeepKr · Wall Street CN / Sina Finance · TaiyangNews · MIT Technology Review · SEIA · Stanford Probability for AI
