BLOG
Three Waves of AI Talent Migration in Three Years: Which River Should You Stand in If You Enter the Field Now?
Three and a half years into the surging large model wave, domestic AI talent has gone through three major waves of migration.
This judgment comes from an in-depth survey by Leiphone in early September — they interviewed multiple headhunters specializing in AI algorithm talent, and sorted the talent flow during this period into three waves: first, Baidu’s Wenxin team was frantically poached; then talent from the “Six Tigers” was divided up; now, inference, Agent, and Infra talent have become the new favorites.
One figure in the report is striking: the average tenure of AI talent has dropped from 2-3 years to 1-1.5 years. That means people in this industry change jobs on average every little over a year.
As someone who has led teams for over a decade and personally experienced two technology cycles from mobile internet to AI, I want to talk about the logic behind these three waves of migration, and a more practical question: for people who want to enter the AI industry now, which river should they stand in?
Three Waves of Migration, Riding the Rhythm of the Industry
First wave, 2023: scrambling for pre-training talent, with Baidu as the bellwether.
Baidu reacted the fastest after ChatGPT launched. In March 2023, Wenxin Yiyan fired the first shot in China, carrying out a “wartime reorganization” of its years of accumulated NLP elites. The result was — anyone with “Wenxin” on their resume saw their market value skyrocket directly. According to headhunters interviewed by Leiphone, for the same person, one company offered 2 million, another 3 million, and yet another 5 million, with bids escalating.
Two types of companies were poaching: the newly founded “Six Tigers” (Moonshot AI, Baichuan, StepFun, 01.AI, etc.), and Alibaba, ByteDance, and Tencent which were urgently catching up. Headhunters estimate that 70-80% of pre-training talent was divided up by ByteDance, Xiaohongshu, Pinduoduo, and the Six Tigers.
Second wave, early 2025: the Six Tigers diverged, with ByteDance and Tencent becoming the inflow destinations.
Hit by DeepSeek’s impact and crowded competition, Baichuan shifted to vertical healthcare, and 01.AI merged its training team into Alibaba Cloud’s joint laboratory — the startup camp contracted, and a large number of pre-training talent re-entered the market. During this period, ByteDance poached talent in three consecutive phases, and Tencent also completed two batches of talent introductions around the time Yao Shunyu joined.
Third wave, from after the 2025 Spring Festival to now: scrambling for post-training, Agent, and Infra talent.
DeepSeek R1 made the entire industry see the cost-effectiveness of inference clearly. In 2026, Agent exploded comprehensively — after Manus went viral and Claude Coding Agent succeeded, talent flow around Coding Agent teams became extremely intense. Meanwhile, recruitment demand for pre-training has dropped by 30-40% from its peak.
Do you see the pattern? The crest of every wave of talent flow precisely hits the rhythm of industry competition. When base models can no longer be rolled competitively, pre-training talent flows; when the cost-effectiveness of inference is verified, post-training talent prices rise; when Agent takes off, people who can do engineering implementation are frantically poached.
Whatever the industry lacks most at each stage, talent runs there — salary is just the result, not the cause.
Why Can’t Big Tech Companies Retain Talent? Three Lethal Factors
Headhunters summarized three forces, and I’ll rank them by the lethality I’ve observed.
First is stock price and market capitalization. More than half of the compensation for core technical personnel in big tech companies is stock. When the stock price is lukewarm, talent retention becomes difficult. The survey specifically named Kuaishou — under stock price pressure, core members of its text model team left early; Meituan is similar. When money isn’t in place, idealism can’t last past two paydays.
Second is strategic orientation. The most typical example is Huawei — its core business is selling Ascend chips and computing hardware, and large models are more of a supporting positioning. This directly affects the investment expectations of the model team, making it a talent outflow side. Baichuan’s shift to vertical healthcare is the same logic: once a company’s strategy changes direction, the entire team loses its sense of direction and leaves faster than anyone.
Third is organizational structure adaptation, which is the most hidden and also the most fatal. The hierarchical structure of traditional big tech companies can’t adapt to the R&D rhythm of large models. The headhunters’ original words were very penetrating: “Faced with new business and new strategies, organizational structure adjustment often needs to be done first, but not many companies can make up their minds quickly.”
I’ll add a version I’ve seen with my own eyes: old employees occupying positions, layers of reporting lines, a decision that can’t get out of the conference room for three months — real frontline AI talent is hard currency in the market, and they won’t waste time with you.
Interestingly, Alibaba is a counterexample: the vast majority of its Tongyi base model talent relies on campus recruitment and internal training, with very little social recruitment, but instead it has the most stable talent structure. Poaching can solve the problem of “having or not having”, but it can’t solve the problem of “retaining or not retaining”.
Age Anxiety: Half Real, Half False
Two details in the survey are very touching.
One is that pre-training talent rarely transitions to post-training — headhunters say the reason is “division of labor barriers” plus “technical obsession”: people who do base models are inherently unwilling to turn around and do fine-tuning.
The other is that 30-40 year old talent from the first half of AI has begun to flow back to universities and public research institutes — former Kuaishou Vice President Wang Zhongyuan became the president of Zhiyuan Research Institute in early 2024. At the same time, the market clearly favors young people in their 20s at the peak of creativity.
Is age anxiety real? My judgment is: age anxiety for “pure execution ability” is real, but there is no age anxiety for “judgment ability”.
If you look closely at the group of people flowing back to universities, they are not the ones eliminated — they are the top batch. They go to research institutes to do cutting-edge exploration, relying on accumulated judgment. The 20-something young people favored by the market are being recruited for “physical strength + learning speed” type positions: high-intensity pre-training, fast-iterating engineering work.
So the real question is not “can you still work in AI at 30”, but: after 30, does your value still stay at “I can endure hardship”, or has it accumulated into “I know what to do and what not to do”?
The former does depreciate quickly in the AI industry — because AI itself is automating the “enduring hardship” part; the latter instead becomes more and more valuable — because the industry rhythm is so fast that there’s no time for trial and error, and judgment has become a scarce commodity.
What Kind of Talent Will Be Scrambled for Next?
First, let’s look at the two most sought-after types now:
AI Infra talent — positions like GPU cluster scheduling, inference framework optimization, FPGA compilation. Why? Computing power is expensive. With the same cards, the efficiency can differ by several times depending on how well scheduling is done. These people directly determine “whether the cards can be fully utilized”.
Data talent — computing power has become a fixed objective factor (you can buy it if you have money), and model architecture has reached a phased plateau (everyone can’t拉开差距). So the industry consensus has become “high-quality data is the core barrier in the second half”. The recent centralized adjustments of data teams by big tech companies like ByteDance are a signal.
What will be scrambled for next? My inference is two types:
“AI + business” translation talent. After Agent explodes comprehensively, the bottleneck shifts from “whether the model is smart enough” to “whether it can be implemented in business”. People who understand both business processes and AI capability boundaries, and can design implementation paths, are seriously scarce right now — this is precisely the most realistic entry point for most working people, much more realistic than switching to an algorithm track.
AI quality and security talent. The release cycle of large models has been compressed from 4.5-6 months to 3.5 months, and the probability of accidents rises accordingly. Who is responsible for verifying model launches? Who takes responsibility for hallucinations? Who handles compliance? These positions haven’t formed a clear rank sequence yet, but demand is already growing — just like ten years ago no one knew that “test engineer” would become a huge occupation.
The judgment framework is just one sentence: see where the industry bottleneck shifts, and talent there will become more expensive. In the base era, the bottleneck was algorithms; in the inference era, the bottleneck was efficiency; in the application era, the bottleneck is implementation — aligning your skills with the next bottleneck is better than chasing rises in any trend.
If You Want to Enter the AI Industry Now, Three Paths
Match yourself according to the threshold from low to high.
First path: don’t switch careers, first add “AI+” to your current position. This is what I most recommend to most people. You don’t need to jump into an AI company to get involved with AI — redo the work in your position with AI, and achieve quantifiable efficiency improvements, and you will already be an “AI + business” translation talent. This is a hundred times more realistic than switching to algorithms barehanded, and large-scale deployment of their own post-training models in vertical industries is happening right now — finance and healthcare are all building their own AI teams, and what they lack most is not algorithms, but people who understand business and can use AI.
Second path: switch to AI engineering, starting from the Agent application layer. If you are a developer, don’t touch pre-training right from the start (that window has closed, and demand has dropped by 30-40%). Start with Agent application development — workflow orchestration, tool calling, RAG, evaluation. These are the directions with the strongest current demand, relatively friendly thresholds, and will not be covered by model progress in the short term. Refer to the survey: teams around Coding Agent are poaching the most fiercely now.
Third path: for fresh graduates or deep career changers, go for Infra or data. If you are a student or decide to make a deep career change, Infra (GPU scheduling / inference optimization) and data engineering are hard currency for the next three years — high barriers, few talents, and rigid demand. But be clear: this path has a steep learning curve and requires real effort, it can’t be solved by brushing up on courses for two months.
Finally, a splash of cold water: in an industry with an average tenure of 1-1.5 years, volatility is the norm. Don’t treat “entering AI” as a one-time action — entering the field just gets you a ticket. Whether you can stay at the table depends on calibrating the alignment between your direction and industry bottlenecks every six months.
That’s also why I always say: instead of chasing hot topics, first practice your “judgment ability” — it’s the only asset that doesn’t depreciate with age and doesn’t shift with trends.
Where the wave hits is not up to you, but which river you stand in is up to you.
Reference: Leiphone, 2026-09-07, Three Major Migrations of AI Talent: Baidu Takes the Lead, Six Tigers Diverge, Tencent’s “Cage Replacement” (Talent flow data and headhunter views in this article are all quoted from this survey, with pseudonyms used; industry judgments are personal views, for reference only.)
Author: Yongliang, 17 years in software industry, 7 years of AI experience, multiple internationally renowned enterprise certifications. AI technical director, telling the truth about careers and education in the AI era.