AI Talent War Triggers Salary Inversion: Fresh Graduates Start at 2 Million Yuan While Veteran Employees Stay Calm

Big tech HC (headcount) contraction and the AI talent war are happening in parallel, and an internal contradiction is surfacing: salaries for new hires in foundation model roles have already inverted those of higher-ranked veteran employees. This compensation system shock, ignited by AI, is reshaping the hiring logic of major internet companies.

1. Salary Inversion: From Isolated Cases to a Widespread Phenomenon

Headhunter Mia told Leiphone that all major companies are currently in a state of HC contraction, but when they encounter someone they truly want, they "usually don't hold back on budget." "If someone from a strong competitor (like OpenAI) comes, they immediately free up a slot." This spare-no-expense recruiting logic directly impacts companies' existing compensation systems.

One internet industry professional described how employees working on foundation models internally, even at lower job levels, already out-earn colleagues at higher levels. Although rumors of "abolishing job levels" remain unconfirmed, the divergence between salary and job level makes the awkwardness of the compensation system increasingly obvious.

This is not the first time salary inversion has occurred. One big tech engineer recalled that when he left his company around 2017, his salary was still lower than that of new campus recruits. At the time, sitting next to him was a returnee fresh graduate who had joined a year later, one job level lower, yet earning 10,000 yuan more per month. During those two years, his team of over thirty people "lost 10% per month, all leaving in groups."

The degree of salary inversion is intensifying. Data from headhunter Mia shows: this year's fresh graduates working on foundation models command annual packages "starting at 2 million yuan," with top tiers at Xiaohongshu and ByteDance Seed reaching 4 to 5 million yuan. Baidu offered foundation model talent annual packages of 2 to 3 million yuan last year, and this year "only more." Compared to a decade ago, when Baidu offered top campus recruits 200,000 to 250,000 yuan in annual salary, the price of top talent has risen nearly tenfold in ten years.

2. The Other Side of High-Pay Recruiting: Internal Conflicts and Divergent Attitudes

The internal impact of poaching with high salaries is not without precedent. A participant in Huawei's "Genius Youth" program once recalled that after compensation was made public on the internal network, many veteran employees felt their academic achievements and contributions were no less significant, asking "why should their salaries be lower than these newcomers'" — some even sabotaged the genius youths at work. At the time, some employees posted under their real names on the internal network, questioning the genius youths' credentials one by one.

A decade later, facing the same issue, many veteran employees at big tech companies are quite calm. A senior algorithm engineer at a major company said: "A new person sitting next to me earning more than me — whatever." In his view, this is simply the normal market compensation mechanism. Veterans who feel they should have gotten a raise but didn't, or that resources weren't allocated sufficiently, will naturally leave when the time comes — "this also fits the current market conditions."

But a calm attitude doesn't mean the problem disappears. Headhunter Mia shared Meta's case: Meta's recommendation system AI department and MSL (Superintelligence Lab) are parallel organizations, both working on generative AI post-training, yet the former department's salaries are relatively low. Some employees have worked five or six years with almost no change in level or compensation, while their counterparts are early senior members who participated in Meta's foundation model research. Salary inversion is not unique to China.

PwC's "Global AI Jobs Barometer" for two consecutive years shows that the wage premium for AI skills in the global labor market has risen from 25% to 62%, with recruitment markets across six continents repricing the same scarcity. The most extreme case comes from Meta: offering a four-year compensation package of approximately $250 million just to secure a 24-year-old researcher.

3. Big Tech's Money-Spending Logic: Buying Only "Certainty"

Multiple headhunters told Leiphone that big tech companies' poaching criteria are strikingly consistent: first, hands-on experience at a top-tier company; second, experience at a well-known startup. The explanation: training a model with 10-20 billion parameters versus training one with hundreds of billions or trillions of parameters is "not even on the same scale" — the latter is extremely capital-intensive and requires experienced people to avoid pitfalls.

This means the situation for ordinary AI job seekers is not optimistic. HC at major companies is basically frozen, with only DeepSeek^Ad and Qwen hiring relatively more. But the freeze is uneven: foundation model teams take in any good talent without exception; what's really contracting is the application side.

Mia explains salary trends through supply and demand: in the past two years, AI talent was in short supply, candidates held bargaining power, and high campus recruitment salaries plus companies outbidding each other created today's inversion. This year, market supply is already sufficient (big companies also have laid-off personnel flowing out), supply and demand are reaching equilibrium, and the premium space is shrinking. But salaries won't drop immediately — if the high salaries given to campus recruits in previous years are suddenly undercut by lower offers for new hires, new dissatisfaction will emerge internally. Compensation adjustment can only be a dynamic process.

Stratification in the talent market is also deepening. "Excellent people never worry about jobs; mid-tier people will find it very hard to find work." Another senior recruiter at a major company, Zhang Yu, stated that truly scarce people still have bargaining room, while candidates whose abilities and experience aren't as outstanding face fiercer competition. For top talent like Yao Shunyu and Tian Yonglong, companies are willing to invest heavily to secure them; others receive far less attention.

The AI market changes too fast, and talent trends turnover as quickly as model updates. Headhunter Mia says the current hot area is still Agents, especially Co-worker Agents for collaborative scenarios. Coding Agents represent a persistent talent gap, and talent related to Evals (model evaluation) and pre-training data is also beginning to receive attention. AI search is far less hot than the previous year; reasoning, with the successive emergence of OpenAI o1 and DeepSeek-R1, has matured technically, leaving less room for talent to make their mark. AI for Math is in an even more awkward position — the capability gaps between models are not yet sufficient to support large-scale independent investment by domestic companies, unless one has ample compute, funding, and research resources.

The Order-of-Magnitude Difference Between Parameter Scale and Training Cost

The training difference between "10-20 billion parameters" and "hundreds of billions or trillions of parameters" mentioned in the article fundamentally lies in the non-linear growth of computational cost and engineering complexity. Each order-of-magnitude increase in model parameters typically requires two to three orders of magnitude more compute. Taking a hundred-billion-parameter model as an example, a single training run may require thousands of GPUs running for weeks, costing millions to tens of millions of dollars; while trillion-parameter models (such as GPT-4, DeepSeek-V3, etc.) involve complex engineering challenges like cluster scheduling of tens of thousands of GPUs, Mixture-of-Experts (MoE) architecture, communication optimization, and fault recovery. Those lacking large-scale training experience can easily stumble in data mixing ratios, parallel strategies, stability tuning, and other areas, leading to millions of dollars in wasted compute. This is precisely why big companies only recognize experience of "having hands-on trained large models."

Conceptual Clarification of Agent, Coding Agent, and Evals

Agent (intelligent agent) refers to AI systems that can autonomously plan, invoke tools, and execute multi-step tasks. Co-worker Agent specifically refers to agents designed for collaborative scenarios that work alongside humans. Coding Agent is an agent specialized in code generation, debugging, and refactoring, with representative products including Claude Code and GitHub Copilot; its technical challenges lie in long-context understanding, code execution environment isolation, and multi-turn error correction. Evals (model evaluation) is the methodology and toolchain for systematically measuring model capabilities, involving benchmark design, human evaluation protocols, and automated scoring — a key component of model iteration and safety alignment. As model capabilities improve, the scarcity of Evals talent becomes prominent because evaluation quality directly determines model optimization direction.

4. Young Companies' Different Logic: Dismantling the Old Job-Level System

Big tech spends money to buy "certainty," while a group of young companies is dismantling the old job-level system with a different recruiting logic.

DeepSeek, which completed its first round of external funding this year, is aggressively expanding, with a clear bias toward engineering: server-side/backend development, and Agent elastic computing/Agent Infra for building the Agent runtime foundation. In Agent productization, they call the orchestration and runtime beyond the model "Agent Harness," benchmarking against code agents like Claude Code/Codex. Building Agents requires not just algorithms but also developers, and infra is equally important.

Before expansion, DeepSeek's hiring profile was extremely distinctive: initially preferring Peking University students, with typical traits of geek spirit, highly technical, strong belief in AGI, and Silicon Valley-style researcher culture. Zhang Yu summarized it as "young, not yet 'polluted' by big tech."

Kimi is another answer: a team of about 400+, maintaining good talent density, with extremely short decision chains — even hiring an intern requires lengthy consideration. Zhipu goes to the other extreme — "heavy use of interns, with more interns than full-time employees. A team of 10 full-time employees with 20 interns is normal for a model company."

Young people are willing to go to these companies because big tech's division of labor is too fine and room for contribution is limited; those pursuing technical exploration and greater autonomy tend to prefer the latter. Zhang Yu explains that the AI circle habitually calls talent under 30 with successful model experience "little geniuses." Beyond salary, they also look at whether the company has sufficient compute and resources, whether it can support them in actually training models, and whether it has successful model-training experience. In summary: "empty promises don't work."

Campus recruitment and experienced hire recruitment use two different pitches: campus recruits care more about whether compensation and job level match; experienced hires care more about scope — the range of responsibilities and what specifically they'll do. And for people they truly want, neither pitches nor rules exist anymore. Mia reveals that major large-model companies handle compensation and scope for excellent AI talent entirely case by case, very flexibly.

One noteworthy shift: overseas companies reverse-poaching Chinese large-model talent became particularly evident in 2025. In the past, it was more often domestic companies going to the US to poach talent; now that Chinese model capabilities are recognized overseas, the direction of this talent channel is beginning to reverse.

Agent Harness and Agent Infra

Agent Harness refers to the orchestration and runtime framework built around large models, responsible for task decomposition, tool invocation, memory management, and multi-turn interaction control — essentially the agent's "operating system." Agent Infra (infrastructure) covers the backend services supporting Agent operation, including elastic compute scheduling, sandbox environments, state persistence, and security isolation. DeepSeek's recruitment of engineering talent to build the Agent runtime foundation is precisely because Agent productization depends not only on model capabilities but also on highly reliable, low-latency engineering systems. Taking Claude Code as an example, behind it lies the need to manage code execution sandboxes, file system access, multi-step reasoning chains, and error recovery — all belonging to the Infra category. Talent in this field needs to combine distributed systems knowledge with AI application understanding, and current market supply is extremely scarce.

5. After the Inversion: The Market Will Eventually Return to Rationality

In every technological wave, companies break compensation order to grab talent, with newcomers earning more than veterans; afterward, industry supply and demand rebalance, and the premium recedes. The salary inversion script is nothing new — it's just that this round, AI has pressed the fast-forward button.

For veteran employees, the frustration brought by inversion is real, but the market mechanism also provides an exit — those who should have gotten raises but didn't, or weren't given enough resources, will vote with their feet. For young people entering the industry, sky-high offers come with extremely high capability thresholds; "empty promises don't work" forces companies to put real resources on the table. For all participants, this talent war is far from over, and all the sky-high offers will ultimately prove the same thing: no matter how much you pay, nothing beats hiring the right person.

One-line Summary

Amid big tech HC contraction, AI talent salaries invert veteran employees, and sky-high poaching reshapes hiring logic.

Paragraph Summary

The article reveals that against the backdrop of overall HC contraction at major internet companies, fresh graduates in foundation model roles command annual salaries starting at 2 million yuan, with top tiers reaching 4-5 million, creating severe salary inversion relative to higher-ranked veteran employees. Big tech poaching criteria are highly consistent: candidates must have large-model experience at top-tier companies or well-known startups, and it's application-side positions that are really contracting. Meanwhile, young companies like DeepSeek, Kimi, and Zhipu are competing for talent with different logics — geek culture, extremely short decision chains, and heavy use of interns — attracting talent with technical autonomy. The global wage premium for AI skills has risen from 25% to 62%, and the phenomenon of overseas companies reverse-poaching Chinese talent is beginning to emerge. The article argues that salary inversion is a阶段性 phenomenon in technological waves, and the market will eventually return to rationality, though top talent will retain bargaining power.