A fresh PhD in AI could ask for RMB 3 million to RMB 4 million a year in 2024. Two years later, graduates with top conference papers and internships at leading labs are not getting interviews. The reversal took less time than the boom.
The bidding war, in numbers
In the 2023 and 2024 model rush, a doctoral graduate with no industry experience could command RMB 3 million to RMB 4 million annually, and engineers from the previous generation of natural language and vision work who touched large models at all could get RMB 2 million to RMB 3 million. Most of those companies did not survive.
Then Tencent arrived in 2025 and reset the ceiling. Wei Xiaokang, who ran recruiting at ByteDance and Meituan, describes the method plainly: where others paid RMB 2 million to RMB 2.5 million for the same person, Tencent paid RMB 3 million or RMB 4 million. A handful of top packages reached RMB 100 million, and pay across the industry doubled for people who had not changed job.
The turn
By 2026 the demand collapsed. Model competition consolidated into a few labs, and the supply of capable researchers, which had been climbing since 2024, met a market that needed far fewer of them. Wei’s framework divides the field into geniuses, a few dozen people who open new directions, talents, the annual top hundred, and hands, everyone who implements.
In foundation models, the talent tier is oversupplied. A few dozen good people is enough, and beyond the genius tier the choice between candidates matters much less than the recruiters believed.
Was it a bubble
Wei argues most of it was not, and his test is a good one. Tencent’s entire foundation model effort costs less than the RMB 200 billion spent on the food delivery subsidy war in a single year, and less than a large platform spends buying accelerators over a few years. If the outcome of that spending is decided by a few dozen people, paying them a great deal is rational.
The bubble was the other part, the dozens of companies that paid heavily for people who could not build a model at all. They are gone, and their departure is what the 2026 market is actually measuring.
The uncomfortable conclusion for anyone entering the field now is that the shortage was never about AI skills in general. It was about a very small number of people, and the industry spent two years discovering how small that number is.
Editor’s note: This is an adapted translation of the original Sohu Tech report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.sohu.com/a/1100734911_121948396.