When Zhu Kexin learned how much a colleague born in 1998 was earning after moving to a major technology company, she felt dizzy.
The colleague, an algorithm researcher, had tripled his compensation.
At a group dinner, Zhu and her coworkers tried to guess the size of his raise. Their boldest estimate was 50%. The answer was far higher: his annual pay had gone from RMB 1 million (USD 149,000) to RMB 3 million (USD 447,000).
By the end of dinner, everyone at the table had recalibrated what they thought they were worth.
Salary benchmarks were changing quickly. Just a year after that move, Zhu said, an annual package of RMB 3 million for a researcher working directly on models no longer raises eyebrows.
And the people receiving such pay packages are getting younger.
One graduate student born after 2000 was earning RMB 5,000 (USD 745) a day as an intern, putting his monthly income far above that of most white-collar workers in the city. His manager had told him the company could offer him RMB 3 million a year as a full-time employee after graduation.
The young man told 36Kr he wanted to graduate immediately.
From Silicon Valley to China, the scramble for elite artificial intelligence talent has pushed some annual compensation packages above RMB 100 million (USD 14.9 million) and drawn founders into recruiting.
ByteDance founder Zhang Yiming spent eight months securing Wu Yonghui to lead its AI operation. Meta has reportedly offered more than extraordinary pay: Mark Zuckerberg has personally courted OpenAI researchers, including by bringing them homemade soup.
Pay is rising throughout the AI talent hierarchy, from leading researchers to interns.
At the top are researchers whose annual compensation can exceed RMB 100 million. There are often only a few on each company’s model team.
One tier below are researchers at major technology companies earning annual packages above RMB 10 million (USD 1.5 million). A recruiter told 36Kr that a large company might have roughly 10–20 people in this group, each an industry figure whose job move would attract widespread attention.
Next are elite new graduates in programs such as ByteDance’s TopSeed, Alibaba’s AliStar, and Tencent’s high-potential talent programs. Their annual packages can reach RMB 5 million (USD 745,000) or more, with perhaps dozens of such hires at each company.
A doctoral candidate who joined ByteDance this year explained to 36Kr how an RMB 6 million (USD 894,000) package could be structured.
“It isn’t all cash. It also includes ByteDance options, Doubao equity, and bonuses,” he said.
Some new graduates receive even more than RMB 6 million, he added, but those cases are “extremely rare.”
The next tier consists of algorithm researchers earning roughly RMB 2–3 million (USD 298,000–447,000) a year. Jokingly described as ordinary salaried workers, they make up the largest group. Their market value depends heavily on notable academic work and experience on major projects.
“This year, RMB 2 million (USD 298,000) and over has become the baseline that an ordinary algorithm researcher can accept. Last year, it was still RMB 1 million and above,” an algorithm researcher at a major tech company said.
“A friend of mine previously got an offer from Seed worth more than RMB 2 million a year, but ultimately turned it down because another AI company offered more than RMB 3 million and a higher job level.”
At the bottom of the pyramid are algorithm interns.
Doctoral interns working in key areas at ByteDance, Tencent, and Alibaba can earn RMB 5,000–6,000 (USD 745–894) a day. Two years ago, a doctoral intern’s average monthly pay was about RMB 10,000 (USD 1,490), roughly equivalent to two days at today’s rates.
With some interns making more than RMB 100,000 (USD 14,900) a month, competition for AI talent, especially foundation model researchers, extends well beyond senior hires.
The rising pay reflects a shifting contest among China’s AI model developers.
ByteDance was once the most generous employer in the AI talent market. Over the past year, Tencent has stepped up recruitment and become another source of lucrative offers.
Meanwhile, DeepSeek and Moonshot AI have raised large sums in private markets, while Z.ai and MiniMax have gone public, giving leading foundation model startups more money to spend on talent.
Talent has always shaped competition between companies. But the current AI hiring race is more intense than the contest during the internet era, the previous period of heavy investment in China’s technology sector.
A closer historical comparison may be the semiconductor industry between the 1950s and 1970s, when companies competed aggressively for a small pool of technical specialists.
The two industries share a defining characteristic: a company may appear to own the technology, but much of the knowledge behind it resides in the minds of a few dozen technical staff.
Access to those people can determine whether a company keeps up.
Their movement between employers is reshaping organizations and the wider AI industry.
Buying the recipe
Major technology companies and foundation model developers are competing for the technical “recipes” held by key researchers.
In model training, a recipe covers nearly every important design choice: the model’s size, the mix of data, the training approach, and the technical methods used.
A successful model is usually the result of extensive trial and error, and the people who lived through that process retain much of what was learned.
The early semiconductor industry worked in a similar way. Process parameters, yields, materials knowledge, and manufacturing know-how could not be fully captured in patents.
Hiring the right person can give a company access to knowledge that is difficult to document, helping it close a generational gap in model performance much faster.
There is a reason these recipes are expensive. Restarting or correcting model training can be enormously expensive. Training a base model, or fine-tuning one with hundreds of billions of parameters, can cost an eight-figure RMB sum or even more than RMB 100 million in computing resources for a full experiment.
If the training direction or methodology is wrong, the resulting loss can far exceed the annual salaries of several researchers. But few people possess the most valuable recipes.
Unlike the entrepreneurial experience accumulated during the internet era, experience training models with 10,000 GPUs is available only at a handful of leading technology companies and laboratories with enough computing power and data.
Even within those organizations, very few people lead training at that scale.
“There are only about 200 people in all of China who can lead large-scale pretraining,” a recruiter told 36Kr.
A ByteDance HR employee said that when the company began systematically targeting key researchers in 2024, its list contained only a few hundred names.
Once people with those recipes start moving, knowledge previously concentrated inside a few companies begins to spread across the industry.
The process resembles a recipe initially known to only a few chefs. Their restaurants hold an advantage until a chef changes kitchens and takes that knowledge along. Individual restaurants may lose their lead, but the spread of expertise raises standards across the industry.
In the second half of 2024, Zhou Chang, who had previously worked on multimodal pretraining at Alibaba, joined ByteDance.
ByteDance paid an eight-figure RMB annual salary to hire Zhou. His arrival quickly helped the company improve its multimodal capabilities, laying part of the foundation for Seedance, which later gained widespread adoption.
Another industry source told 36Kr that former Z.ai researcher Cheng Ye’an, a key contributor to GLM-4.5 and GLM-4.5V, joined Moonshot AI this year and appeared among the authors of its K3 technical report.
His principal contribution was in post-training for agentic coding, an area in which Z.ai has been particularly strong.
When a key researcher leaves, the company loses more than one person’s labor. The researcher takes accumulated team knowledge with them: experimental approaches, records of failed attempts, and debugging experience. That knowledge may effectively be delivered to a competitor.
That transfer can begin during job interviews, as candidates describe how their teams work. Interviews have become one of the most frequent channels for leaks of technical details.
Researchers with this knowledge also attract investors when they start companies.
Liu Yu, formerly a research director at SenseTime, had already experienced the entire multimodal development cycle by 2024, when the technology was still at an early stage. His work spanned data processing, training infrastructure, model development, and product deployment, and he had coordinated more than 4,000 GPUs for model training.
At the time, even researchers with experience training on clusters of more than 1,000 GPUs were rare.
When Liu left to start a company, investors rushed to meet him. He turned most of them away, earning a reputation as a founder who was difficult for investors to reach.
Such moves have become an important part of competition among China’s foundation model companies.
“The core of [AI] model competition is mainly compute, data, and people,” recruiter Sam told 36Kr. “Only if you find the right people can the compute and data be put to use. Otherwise, spending more money doesn’t help.”
The value of people extends beyond the researchers who lead training.
Companies also need large numbers of smart, reliable young researchers to handle the supporting work. Demand for them has spread the hiring contest across the market.
Unorthodox tactics
Two years ago, Baidu and Alibaba, both early entrants in AI, had strong research teams that rivals frequently recruited from.
More recently, ByteDance, itself rich in AI talent, has become another recruiting ground.
An employee at a major model developer told 36Kr that the company was benchmarking ByteDance’s data organization and recruiting people role by role.
“If you move over, your compensation can double,” the person said.
ByteDance responded quickly. At the end of 2025, the company sent an email to all employees announcing that spending on bonuses would rise 35%, while its salary adjustment budget would increase by 1.5 times.
It also introduced a new job-level system and raised the maximum pay at each level. The change was widely seen as a way to accommodate highly paid AI researchers who no longer fit its old grading framework.
Its retention efforts have intensified further this year. “Doubao equity,” available only to employees in AI-related businesses, now has an explicit valuation and a repurchase mechanism. Employees who hold it can also choose to convert more of their annual bonus or total cash compensation into Doubao equity.
Internally, it is being treated much like ByteDance options were in the company’s early years.
These measures may share a common trigger: nearly 70 key technical employees left its Seed unit last year.
Tencent has offered authority alongside money.
High-profile Silicon Valley researchers such as Yao Shunyu and Tian Yonglong can arrive at Tencent and immediately lead teams, or even take charge of an entire model organization.
But some candidates want more than money and titles.
Some startups have developed their own unconventional recruiting tools.
According to 36Kr, two unnamed AI startups in Shanghai frequently target the same candidates. Both pursued one doctoral student this year, who was prepared to accept an offer from one of them but ultimately chose the other, which had close ties to a university.
The deciding factor was not money. The second company reportedly said it could help provide him with a faculty position at a Shanghai university.
Z.ai may have pioneered this university-linked recruiting model. Its founder and chief scientist, Tang Jie, and much of its senior management came from Tsinghua University. Some Tsinghua students hoping to pursue master’s or doctoral degrees under Tang have viewed joining Z.ai as a possible route into that academic circle.
When companies cannot offer similarly powerful resources, HR teams and recruiters may try a more psychological approach.
Li Qing works at one of China’s six “AI tigers,” a group of startups once widely regarded as the country’s six leading contenders to compete in AI globally. Recently, he has been receiving several calls from recruiters every week, often during working hours.
Rather than opening with positions or salaries, recruiters frequently raise concerns about his employer’s share price, his working hours, or his health, trying to weaken his confidence in the company. Li remembers one recruiter who was especially fond of discussing his company’s stock:
“Your company’s share price has fallen. Wouldn’t you consider opportunities elsewhere? Once XX goes public, its upside will definitely be greater.”
Large technology companies also take a more systematic approach: hiring extra people to protect against departures.
One ByteDance employee told 36Kr that in some strategically important research areas, Seed intentionally maintains two teams with comparable capabilities even when that exceeds the resources strictly required.
“That way, even if a competitor takes an entire team, the other one can definitely step in,” the employee said. “Even if you leave one of the most critical roles, it won’t make much of a splash at ByteDance.”
Recruiters see the same pattern. One told 36Kr that Tencent tends to hire aggressively when assembling particular teams, while ByteDance recruits at a steady pace.
“Whether it needs people at that moment or not, ByteDance keeps recruiting.”
Other tactics aim more directly at weakening a competitor’s research team. An employee at a major technology company told 36Kr that his company was “systematically cultivating the model talent inside a rival.”
“We identified the competitor’s ten most important people and approached every one of them. If they’re willing to join us, great. If they aren’t, we encourage them to start their own companies, and we can help find investors.”
He described a process managed through target lists and quotas.
The company maps key people inside a rival’s model team and finds ways to contact each one.
Whether they join the company or leave to start a venture with investor support, their departure weakens the rival.
Some investors begin courting ByteDance employees while they are still working there, encouraging them to leave, helping assemble founding teams, and arranging initial financing.
Others spend time around the gates of Tsinghua University and Peking University, seeking out professors and students, and even backing student-founded startups.
HR departments are relatively neutral when employees leave to start companies rather than joining direct competitors.
“Options can retain people who might otherwise be poached by rivals, but they can’t stop someone who genuinely wants to start a business,” one investor said. “If someone really wants to start a company, they will eventually leave. At least they aren’t going off to strengthen a competitor.”
Poaching and retaining talent have become essential skills for AI companies’ HR teams.
A researcher at a major technology company told 36Kr that some HR teams coach priority candidates line by line on what to say when dealing with rival recruiters.
That coaching does not always work. Some candidates immediately use one offer to negotiate a better package elsewhere. When two employers compete fiercely, a candidate can sometimes skip parts of the interview process and secure a higher offer simply by presenting the first one.
Some AI companies use deadlines to accelerate decisions. A candidate may receive a signing bonus equivalent to three months’ pay for joining within a week, two months’ pay for joining within two weeks, and nothing after three weeks.
Recruiters are also changing how they find people.
Much of the AI talent pool is effectively invisible on conventional hiring platforms.
“You can search for an ordinary programmer and find thousands of people on recruitment platforms,” recruiter Pei Xiaoke said. “For foundation model talent, you might not find ten. Sometimes you can’t even find five.”
Researchers are easier to find in author lists for papers at leading academic conferences, contributor lists on GitHub, and technical posts on X. Many publish email addresses on their personal websites.
Recruiters now need to read papers and send emails as well as screen resumes and make calls. Understanding those papers is often the difficult part.
Some recruiters began using AI to do it for them, asking models to summarize papers and generate customized outreach emails.
The recipient might get a message such as: “Professor Zhang, I read your paper, and I found your understanding of multimodal systems particularly insightful.”
The technique spread quickly. Researchers’ inboxes filled with similar messages bearing an unmistakable AI-generated tone, exposing the tactic.
Researchers stopped responding, pushing recruiters back toward personal networks and in-person meetings.
They wait around university grounds, tap professors’ and students’ networks, and attend conferences, invitation-only events, closed-door meetings, and after-parties.
Wherever AI talent gathers, recruiters are rarely far away.
Choices and fiefdoms
For all the tactics recruiters deploy, AI researchers have their own priorities.
“I wouldn’t go,” one ByteDance researcher often tells recruiters from rival technology companies.
His response can be blunt: “Your organizational structure isn’t designed properly. Come back to me after you’ve fixed it.”
On one occasion, he turned down a senior role because the job description sounded to him like a series of fragmented algorithm tasks. It required one person to work simultaneously on image understanding and a portion of a vision model.
“Just applying vision models to gaming is enough for one person to study for five years,” he said. “If you have one person doing scattered work like this, they won’t even develop a complete understanding of the data. How could the project possibly work?”
“Only people who don’t understand the field would design an organization that way.”
For many AI researchers, joining a large company or maximizing immediate pay matters less than the work itself. That differs from the later years of the mobile internet era, when many employees at major technology companies focused on promotion and pay.
“To put it plainly, the growth dividend was gone,” a recruiter told 36Kr. “People had to secure a good position and then hold on to it.”
“But AI is still in a period of enormous growth. Capable people would rather go somewhere they can produce results quickly.”
Many AI workers, particularly young high performers, prioritize the ability to shape research and produce results over immediate pay. That preference has drawn some toward startups such as Moonshot AI, DeepSeek, and Z.ai.
People in key roles at those companies also tend to be younger and more likely to have built their careers directly in the foundation model era.
For these researchers, organizational structure is often the biggest difference between a large technology company and a startup.
How a company organizes its teams, and whom it puts in charge, reveals how it intends to build its business.
Startups are generally leaner, flatter, more flexible, and more AI-native.
A large company that gives its foundation model division those same characteristics can also become more attractive to researchers.
In model training, an “AI-native” organization is one whose people have built foundation models, worked in companies devoted to them, and understand how to organize that research.
ByteDance, eager to make progress in foundation models in 2024, brought in Wu Yonghui the following year to lead Seed.
His predecessor, Zhu Wenjia, was an experienced algorithm specialist and systems architect, but his background was largely in search, recommendation, and advertising systems. When he took over Seed, he had little direct experience with foundation models.
The subsequent developments reinforced a growing industry view that foundation model training needs leaders with firsthand experience.
Wu, a former Google DeepMind vice president of research who had been deeply involved in Gemini development, was widely seen as fitting that profile.
Tencent made an even bolder choice in 2025.
Long regarded as cautious and steady, Tencent had fallen behind in foundation models. At the end of 2025, it hired former OpenAI researcher Yao Shunyu, then just 28, to lead its Hunyuan foundation model operation.
“Senior management wanted him to come in and break the existing order,” one Tencent employee said.
His arrival produced a “catfish effect,” coming across as a newcomer disrupting an established organization. Tencent began changing rapidly, starting with an influx of researchers with direct foundation model experience.
According to 36Kr’s review, at least eight prominent external hires have joined Hunyuan in the nearly 12 months since Yao arrived, all reporting directly to him.
The list includes Zhang Chi and Huang Qi from ByteDance Seed, as well as well-known overseas researchers Pang Tianyu, Tian Yonglong, and Lin Xudong.
All had firsthand research experience training foundation models.
“Yao Shunyu has a kind of star effect,” one Hunyuan researcher said. “He attracts more people who hold the recipes and makes them willing to come to Tencent.”
He again used a kitchen analogy.
“If the person with the recipe looks at your kitchen and thinks the spatula isn’t the one they want and the cookware isn’t the cookware they need, how are they supposed to make a good dish? Now they believe Yao Shunyu is at least a competent head chef.”
Longtime leaders, meanwhile, have left or moved to other roles.
Key figures from Hunyuan’s previous structure, including former Tencent AI Lab deputy director Yu Dong, former Hunyuan head Jiang Jie, former multimodal understanding lead Hu Han, and post-training specialist Xu Can, have either left or changed roles since Yao took over.
The organization is rapidly becoming younger.
“Those so-called veterans don’t actually understand more than we do,” one Hunyuan researcher told 36Kr. “In some aspects of [AI] model training, our understanding is far ahead of theirs.”
“Not all seniority adds value. A lot of people have never actually won a battle in foundation models.”
An employee at another leading foundation model company made a similar observation. Almost everyone on the team, the person said, was born after 1995, and some were born as late as 2005.
For startups, the person argued, there is little need for managers whose main function is to issue instructions.
He has little interest in large technology companies because their complicated organizations and many reporting layers can leave managers focused on supervision rather than research.
“At a startup, because everyone’s goals and working style are highly aligned, it’s actually easier to produce results.”
Foundation model teams are developing flatter structures that judge people by training results rather than seniority.
Experience building virtual humans, recommendation systems, advertising systems, search products, or cloud infrastructure in the previous technology era no longer automatically qualifies someone to lead a foundation model team.
The effect of organizational change can show up quickly in the product.
Before 2026, industry assessments of Tencent’s Hunyuan model were overwhelmingly negative. After Yao and his team arrived, Hy3 was widely seen as bringing Tencent back into contention. One more measured assessment was that Hunyuan had evolved from a model built largely for benchmarks into a foundation for productivity products such as WorkBuddy.
Two months later, Hy4 Preview was released and was described widely as placing Tencent among China’s leading open-source model developers.
The sequence shows how new people can reshape an organization and, through its products, change its competitive position.
“It isn’t that Yao Shunyu changed Hunyuan by himself,” a person close to Tencent said. “But at that point in 2025, if there hadn’t been someone with real authority, someone who could cross internal fiefdoms and take responsibility, Hunyuan definitely wouldn’t be where it is today. It would only have fallen further behind.”
Alibaba has followed a different path. Its early start in foundation models and deep talent pool initially left it less exposed to poaching than many rivals.
Tongyi Lab traces its roots to Damo Academy. Early Chinese AI researchers including Zhou Chang, Lin Junyang, and Luo Fuli developed their careers at Alibaba. Its open-source culture also attracted engineers motivated by the technology, making it a fertile recruiting ground for rivals in 2024.
Over the past two years, however, its ability to attract and retain talent has weakened.
ByteDance and Tencent have reworked their job-level systems to reduce the weight of seniority, while a handful of startups have launched state-of-the-art models and gone public in 2026.
Alibaba’s relatively rigid job-level and compensation systems have consequently become recruiting obstacles.
Former Qwen head Lin Junyang held a P10 rank, according to people familiar with the organization. Apart from his direct reports, most researchers could reach only around P7. The same researchers could often secure positions one or two grades higher at ByteDance or Tencent.
“It’s difficult for Alibaba to put a young outsider directly into a core position,” one Alibaba employee said.
Since Lin left in March, Alibaba has repeatedly reorganized its foundation model operations.
In March, it established the Alibaba Token Hub, or ATH, business group.
In April, it created a group technology committee and upgraded the Tongyi business unit.
In June, it merged the Tongyi unit with Future Life Lab, its two most important model teams, to establish the Token Foundry business unit.
All of them report directly to CEO Eddie Wu.
The changes have concentrated control over Alibaba’s AI strategy in the CEO’s hands. One Alibaba executive questioned whether that was the best approach.
“Eddie Wu, as group CEO, has too many things to manage. His attention is easily divided, and it’s hard for him to stay deeply focused on AI alone,” an Alibaba executive told 36Kr.
Several industry figures believe Alibaba still needs a leader with direct foundation model experience who can focus on that work.
Tides rise and fall
Researchers move between AI companies as their employers gain and lose momentum.
Recruiters identify four phases in that movement across China’s model industry.
In 2023, GPT-3.5 became the benchmark, and employees from OpenAI, Google, Meta, and other overseas technology companies became prized recruits for Chinese employers. China’s model industry and talent pool were still developing. Early entrants Alibaba and Z.ai had relatively mature teams.
In 2024, six “AI tigers” emerged from China’s crowded field of model developers: Z.ai, Moonshot AI, MiniMax, Baichuan AI, StepFun, and 01.AI. They became the next major training ground for talent.
By 2025, employees from DeepSeek, ByteDance, and Alibaba were nearly as sought after as Silicon Valley researchers. Companies looking for people with pretraining experience concentrated heavily on those organizations.
In 2026, people from DeepSeek, ByteDance, and Alibaba remain in demand, while Z.ai and Moonshot AI have again become recruiting targets.
Moonshot AI, for example, was among the first Chinese companies to release an extremely large native multimodal model.
Immediately after the model’s release, recruiters began waiting downstairs from Moonshot AI’s offices in hopes of poaching multimodal researchers.
The mix of hiring demand is also changing. Some roles are stabilizing or declining in demand; others are gaining importance.
Cursor chief people officer Adam Ward said in an interview that, beyond AI researchers, Silicon Valley is seeing particularly strong demand for forward-deployed engineers, or FDEs.
AI model developers or cloud providers send these engineers to customers to deploy models and applications in their workplaces.
“AI researchers are a very clearly defined group. Their definition, numbers, and distribution are all relatively clear,” Ward said.
“What is harder to recruit now are the ambiguous or entirely new roles.”
That reflects a broader shift in Silicon Valley. Companies are still investing people, GPUs, and money in building the best models, but turning those models into revenue-generating applications is becoming more important.
A similar change is taking place in China. A June report jointly released by Tsinghua University’s School of Economics and Management and recruitment platform Liepin found that roles in foundational AI algorithms and models accounted for roughly half of AI labor demand in 2022. By the first quarter of 2026, that share had fallen to 20%.
In the second quarter, demand for AI agent product development engineers exceeded demand for algorithm engineers for the first time, making it the single largest job category in China’s AI sector.
Over those four years, demand for talent working on agent-related products recorded quarter-on-quarter growth of as much as 40%, according to the report.
Hiring priorities are shifting from training models toward putting them to work.
Recent organizational and product changes at major technology companies point in the same direction.
Alibaba integrated QoderWork, Wukong, and MuleRun into QwenWork and began prioritizing AI-driven office software.
ByteDance, meanwhile, combined the Feishu product team with the Doubao product team and integrated Trae and Coze into Doubao, marking its biggest restructuring of business-facing products in five years.
Both companies’ moves have been widely interpreted as responses to Tencent’s AI office assistant, WorkBuddy.
Competition among the three technology giants for this market could quickly increase demand for the relevant talent.
The key roles are not conventional product managers, but product engineers who build harnesses: systems that steer models and turn their capabilities into reliable task execution.
Demand for AI algorithm specialists at major technology companies has nevertheless begun to cool compared with the previous two years.
Hiring continues, but after two or three years of aggressive recruitment, major technology companies and leading model developers have filled many of their research roles. ByteDance Seed alone has more than 1,000 algorithm specialists.
“Most companies have already built reasonably complete training teams, so naturally there are fewer openings,” one industry source said.
“The main task for large companies now is to focus only on the very best algorithm talent and retain the people they already have, rather than continue hiring at scale.”
In the second quarter of this year, Tencent and Alibaba together recorded more than RMB 100 billion (USD 14.9 billion) in capital expenditure. Both companies’ free cash flow turned negative, an unusual coincidence.
Industry estimates cited by 36Kr suggest ByteDance’s quarterly investment was at least comparable to Tencent’s.
On that basis, the three companies may have spent more than RMB 150 billion (USD 22.3 billion) in capital expenditure in a single quarter, much of it on AI computing infrastructure. By 36Kr’s estimates, that is roughly 1.25 times the cost of the Hong Kong-Zhuhai-Macao Bridge, 60% of the cost of the Three Gorges Dam, or enough to build three Fujian aircraft carriers.
Those infrastructure bills are one reason the willingness to spend almost without limit on researchers is unlikely to last indefinitely.
Recruiter Pei Xiaoke has felt the shift as large technology companies become more selective.
He uses an analogy.
In 2023, Chinese teams hiring AI talent were satisfied with people who had merely “seen a pig run,” a Chinese expression meaning they had observed how the work was done without necessarily doing it themselves.
As companies developed a clearer understanding of the technology, their standard rose to people who had actually “eaten pork,” meaning they had direct experience.
Now that many people have such experience, companies increasingly want only those who have “eaten the good stuff,” the candidates with exposure to the strongest models, teams, and resources.
The pool of candidates employers will consider is therefore narrowing.
For AI researchers, Pei said, computer science students from the C9 League, a group of nine of China’s most prestigious universities, now form the first tier of candidates.
Next come graduates of universities included in Project 985, a government initiative that funded a broader group of elite institutions.
Graduates of Project 211 universities, another national higher education initiative covering a still wider group of institutions, and other schools are increasingly excluded from top companies’ core AI recruiting pools.
Students at the most elite institutions are more likely to join research groups with access to frontier projects and greater resources. Employers increasingly value that experience.
An algorithm intern at a major technology company has noticed another change: fewer full-time positions are available to AI interns, even as pay for those who receive offers has doubled.
The two developments reflect the same shift in hiring priorities.
As AI models absorb more basic work, companies are concentrating budgets on a smaller number of top performers.
“Since the beginning of this year, some areas, including post-training, have been becoming increasingly automated, so headcount demand is also falling,” the intern said.
He is no longer surprised.
“Some teams have completely frozen their headcount. Even people who are considered standouts among their peers can submit a resume and get rejected at the screening stage.”
While major technology companies slow recruitment, startups remain active. They see an opportunity to hire researchers leaving ByteDance and are preparing to recruit aggressively.
One recruiter agrees that “the era of broad-based expansion at ByteDance is basically over.”
His recruiting firm is considering ending its ByteDance contract next year.
“If you’re contracted with ByteDance, you can’t poach ByteDance employees. If you’re caught, you can be fined and lose the partnership,” he said.
“Compared with serving ByteDance, the returns from moving ByteDance talent into startups now look higher.”
The contest for AI talent continues, but its most feverish stage may have passed.
In 1983, physicist Emanuel Derman received a call from a Wall Street recruiter asking whether he wanted a job paying USD 150,000 a year.
At Bell Labs, he had been making USD 50,000. Derman later joined Goldman Sachs and stayed until 2002.
Derman and a generation of physicists and mathematicians found more than salaries three to five times higher on Wall Street. They found a professional identity that could last 20 or 30 years.
Many members of that generation later became billionaires, tenured professors, or hedge fund founders.
Today’s young AI researchers have no assurance that their opportunity will last as long.
Zhu Kexin and Li Qing are pseudonyms used at the interviewees’ request to preserve their anonymity.
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Wang Yuchan and Wen Lihong for 36Kr.
Note: RMB figures are converted to USD at rates of RMB 6.71 = USD 1 based on estimates as of September 23, 2026, unless otherwise stated. USD conversions are approximate and, where appropriate, rounded for ease of reference. They may not fully match prevailing exchange rates.