Chinese AI Talent Pay Inverts as Enterprises Face Model Overprovisioning Costs

Leiphone said headhunter Mia described a hiring market in which major companies are in “HC contraction” and selecting only the strongest candidates, but still compete for people they want and usually do not cap budgets. A candidate from a strong competitor such as OpenAI can prompt a company to free a slot immediately, she said. The result is internal tension: one internet employee told Leiphone that staff working on foundation models can earn more than colleagues at higher job levels, and that the company has circulated unconfirmed talk of canceling job levels. Companies face a trade-off between paying enough to attract scarce talent and disrupting existing pay structures.

The inversion is not new. A big-company engineer recalled that around 2017, when he left, his salary was lower than that of campus hires; a returnee fresh graduate one job level below him earned 10,000 yuan more per month. In his 30-person team, about 10% left each month, he said. A senior headhunter said Alibaba and Tencent were already aggressive a decade ago; Tencent gave campus offers to students with Baidu internships, while Baidu paid top campus hires about 200,000 to 250,000 yuan a year and Alibaba Star capped at 600,000 yuan.

Current packages dwarf those figures. Mia said foundation-model fresh graduates this year start at 2 million yuan, with top offers at Xiaohongshu and ByteDance Seed at 4 million to 5 million yuan; Baidu paid 2 million to 3 million yuan last year and will pay “only more, not less” this year, she said. Leiphone reported that PwC’s Global AI Jobs Barometer found the wage premium for AI skills rose from 25% to 62%, and that Meta offered a four-year package of about $250 million for a 24-year-old researcher.

Not all AI workers benefit. Headhunters told Leiphone that major companies’ hiring is largely frozen, with DeepSeek and Qwen recruiting relatively more, and that base-model teams still accept strong candidates while application-side roles shrink. A senior recruiter, Zhang Yu, said the market is stratifying: scarce top talent retains bargaining power, while middle-tier candidates face harder competition. Hot areas include Agent, especially co-worker Agent, and Coding Agent, as well as model evaluation and pre-training data; AI search and reasoning have cooled as technology matures, and AI for Math remains difficult to fund independently, according to the report.

Young model companies are using different recruiting logic. DeepSeek, after completing its first external funding round this year, is expanding engineering roles for server and back-end development and Agent infrastructure, including an “Agent Harness” for orchestration and runtime, Leiphone reported. Kimi has about 400 employees and a short decision chain, while Zhipu relies heavily on interns—about 10 formal employees to 20 interns per team, according to Zhang Yu. Young AI workers are drawn by narrower big-company divisions and more autonomy, and “small geniuses” under 30 with model experience also weigh compute, resources and a company’s record of training models, not just pay, Zhang Yu said.

The cost of overprovisioning is the other side of the AI boom. In a TechRadar column, a Core42 senior vice president wrote that businesses often fall into the trap of wanting the most powerful model, the most sought-after accelerators and the fastest route from pilot to implementation. A proof of concept used by a handful of people can look cheap, but running the same application thousands of times a day, or having agents call multiple models for one task, changes the economics.

The column argued that the mistake is not owning a “Ferrari” but using the same model, hardware or deployment path for every project. A customer service copilot needs fast responses; a large document summarization job can wait longer for more careful “thinking”; government applications may prioritize sovereignty and where workloads run. Smaller models can handle routine classification, extraction and routing at lower cost, mid-sized models support everyday reasoning, and the largest models should be used selectively for complex reasoning, strategic planning and high-stakes tasks, according to the column.

Enterprises should start from the desired outcome and match workloads to models and accelerators, the Core42 executive wrote. Low-latency applications need responsive hardware, while high-volume workloads benefit from throughput and cost efficiency; running both on the same infrastructure can mean paying for performance one workload never uses. The column recommended drawing on heterogeneous accelerators from providers including Nvidia, AMD, Qualcomm, Microsoft and Cerebras, and noted Cerebras ran OpenAI’s gpt-oss-120b at about 3,000 tokens per second.