The revelation of a 1998-born colleague's new annual salary after jumping to a major tech firm left Zhu Kexin feeling dizzy. The raise was a staggering threefold increase — when Zhu and his colleagues guessed the figure during a group dinner, their "boldest estimate" was a 50% bump. The actual answer: a leap from 1 million to 3 million yuan, far exceeding their expectations. Within a single meal, everyone at the table re-evaluated their own market value. The old salary benchmarks were rapidly rewritten, and "dizziness" became a recurring theme.

Just one year after that job change, Zhu Kexin noted that no one is surprised anymore by first-tier researchers earning 3 million yuan annually. The age bracket for high salaries is also dropping. A post-00s graduate student, still on campus, commands a daily intern rate of 5,000 yuan — a monthly income surpassing most white-collar workers in the big city. His supervisor hinted, "Upon graduation and full conversion, you can expect an annual package of 3 million yuan." The young man is candid: he wants to graduate this year.

From Silicon Valley to China, as top AI talent is fought over with nine-figure salaries and "first-hand warmth," the price for every tier of the AI talent pyramid — from the apex to the base — is being pushed ever higher. At the very top, individuals can secure packages exceeding 100 million yuan, but this group is minuscule, with only a few members per model team. The next tier consists of those earning over 10 million yuan annually at major firms. One headhunter noted that each large company has roughly 10-20 such individuals, all industry stars whose job moves draw sector-wide attention.

The next rung down includes top fresh graduates in programs like TopSeed, Ali Star, and Tencent "Daka," with packages of "5 million yuan and above," numbering dozens per company. A PhD candidate joining ByteDance this year explained the composition of a 6 million yuan package: "It's not all cash; it includes ByteDance options, Doubao shares, and bonuses." He added that some fresh grads receive even more than 6 million, "but those are rare cases."

Below them are the first-tier AI algorithm engineers earning 2-3 million yuan, dubbed "the working class," the largest group, priced on academic publications and core project experience. "This year, a 2 million yuan annual salary is the baseline for a typical algorithm researcher; last year it was 1 million," said one major firm algorithm engineer. "A friend of mine secured a 2 million+ package at Seed but didn't take it because another AI company offered him 3 million+ and a higher title."

At the very base are AI algorithm interns. Doctoral interns in core directions at ByteDance, Tencent, and Alibaba can earn 5,000-6,000 yuan per day, whereas two years ago, the average monthly salary for a doctoral intern was 10,000 yuan — equivalent to just two days' pay now. With intern monthly incomes surpassing 100,000 yuan, the battle for AI talent, especially for large model researchers, has entered a white-hot phase across every level.

Behind this lies the evolving competition among Chinese model companies. ByteDance was once the most generous buyer of AI talent, but in the past year, Tencent has surged forward as a top-tier talent financier. Meanwhile, DeepSeek and Moonshot AI have secured massive funding in the primary market, while Zhipu and MiniMax have gone public, arming leading AI startups with fresh capital. Historically, talent has been a critical variable in industry competition, but the intensity of the current AI talent war far exceeds that of the last hot-money surge — the internet boom. The only comparable era might be the semiconductor industry's early days between 1950 and 1970, where the common thread was: companies appear to possess technology, but in reality, the true holders are a few dozen core technical staff. Thus, talent became the "bottleneck" variable in corporate rivalry, transforming competition between AI companies into a contest for people. The flow of talent dramatically reshapes every company involved and profoundly influences the entire AI landscape.

Buying the Recipe

What major firms and model companies are truly vying for is the technical "recipe" held by core talent. When training models, researchers use "recipe" to describe every aspect of model training: model size, data ratios, training approaches, and key tricks. A model's success emerges from extensive trial-and-error, and the individuals who navigate this process possess the resulting knowledge. This mirrors the early semiconductor industry, where process parameters, yield rates, materials, and manufacturing know-how were difficult to fully patent. Recruiting the right person grants access to these "invisible cognitions," allowing a company to rapidly close model gaps and shrink competitive disadvantages.

The "recipe" is expensive for good reason. The reset cost of model training is immense: a single full experiment on a base model or fine-tuning a 100-billion-parameter model can incur computing costs reaching tens of millions or even 100 million yuan. If the training direction or methodology is flawed, the loss far exceeds a few researchers' salaries. More crucially, the number of people who hold the core recipe is very small. Unlike much of the internet-era entrepreneurial experience, the "million-card training experience" prized in the model community is only supported by the computing power and data of a few top firms and elite labs; among those, the individuals qualified to act as "training commanders" are exceedingly rare. "Across China, there are only about 200 people who can lead large-scale pre-training," said headhunter York. A ByteDance HR official also revealed that when they began intensively recruiting core researchers in 2024, the list they drew up contained only a few hundred names.

When those holding the core recipe move, the model training knowledge once confined to a few companies gets "open-sourced," transferring across firms and ultimately raising the industry-wide ceiling. It's akin to a recipe held by a few chefs that once made their restaurants seem unbeatable; but once a top chef switches restaurants, the recipe spreads, preventing any single establishment from maintaining a lasting edge — yet the proliferation of better restaurants elevates the entire industry. In late 2024, Zhou Chang, who previously led multimodal pre-training at Alibaba, joined ByteDance. Though ByteDance paid a multi-million salary, Zhou's arrival rapidly advanced ByteDance's multimodal capabilities, laying the foundation for their highly successful Seedance model. Industry insiders also told 36Kr that Cheng Ye'an, a key contributor to GLM-4.5 and GLM-4.5V at Zhipu, has joined Moonshot AI and appears on the K3 technical report author list. His primary contribution lies in Agentic Coding post-training, an area where Zhipu excelled.

For a company, losing core talent means more than losing a workforce; it hands over the team's collectively accumulated knowledge — experimental paths, failure records, debugging experience — directly to competitors. This transfer can even occur before a new hire starts. Some individuals begin leaking internal methodologies to rival firms during the interview process itself. Interviews have become a high-frequency channel for technical detail leakage. If a recipe-holder chooses to found a startup, they inevitably become a magnet for capital. Liu Yu, a former research director at SenseTime, completed a full cycle from data processing and training infrastructure to model R&D and product launch in 2024, overseeing over 4,000 GPUs for model training — a time when even those with thousand-card training experience were scarce. When he left to start his own company, investors swarmed immediately. So inundated was he with funding offers that he had to shut most doors, earning him the moniker "the investment community's unseeable man" in the industry.

Talent mobility has become a significant part of the competitive capabilities race among domestic large model companies. "The core of model competition is compute, data, and people," said headhunter Sam. "Without the right people, compute and data are wasted. Spending more money is futile otherwise." This is why "people" have become the most critical asset in AI industry competition. Beyond finding the few who hold the recipe, model companies also need to recruit a large number of smart, reliable young people to handle the grunt work. The talent war has thus been fully ignited.

Unconventional Tactics

Two years ago, the high ground and source of attrition for AI talent were Baidu and Alibaba, pioneers in technology. Recently, ByteDance, with its abundant AI talent, has also become a new "mountain" and target for poaching. An insider at a major model company revealed they are benchmarking ByteDance's data systems, targeting its roles for poaching: "The moment they switch, their salary can double." ByteDance reacted swiftly. In late 2025, ByteDance sent an all-hands email announcing a 35% increase in bonus investment and a 1.5x boost in salary adjustment, alongside a new job grade system that raised salary caps across the board. The grade adjustments are seen as a direct response to AI talent needs: high-paid researchers could no longer fit into the old framework. Retention measures have continued to escalate this year: ByteDance's "Doubao shares," granted only to AI-related employees, now have a clear valuation and buyback opportunity, and holders can convert more of their year-end bonuses or cash compensation into these shares, making them as valuable as the options from ByteDance's early days. These moves may all trace back to a common trigger — last year, Seed lost nearly 70 technical backbone members.

Tencent, meanwhile, offers not just money but also positions. Star Silicon Valley researchers like Yao Shunyu and Tian Yonglong can immediately lead a team upon arriving at Tencent, even becoming the head of the entire model team. With such precedents, headhunters now have a pitch: "If you stay at OpenAI, you're just one of hundreds of top researchers, but if you come to a Chinese big tech firm, you can be the business's number one." However, money and positions aren't everything. Many startups have their own unconventional recruiting tricks. Two Shanghai AI startups often target the same person. This year, both pursued a doctoral student who was ready to accept an offer from one, but he was pulled away by the other, which boasted backing from a university. The deciding factor wasn't money but the promise of "a faculty position at a Shanghai university." Zhipu may be the pioneer in this "university track." Founder and Chief Scientist Tang Jie, along with the core management, are almost all from Tsinghua University, and in this strong "academic atmosphere," many Tsinghua students choose to join Zhipu as a stepping stone to study under Tang for their master's or doctorate.

If resources are lacking, HR or headhunters resort to "psychological warfare." Li Qing works at one of the "Big Six Tigers" of large models. Lately, he receives several headhunter calls weekly, often timed during work hours. The callers employ a tactic of attrition — they don't rush to discuss roles or salaries but circle around negative topics like his company's stock price, work intensity, and his personal health, aiming to subtly shake his confidence in his current employer. Li Qing recalls one headhunter who loved discussing his company's stock. "Your company's stock is falling. Have you considered external opportunities? After [competitor] goes public, its potential will be much greater."

Major tech firms have a more systematic approach, which we call "saturation staffing." A ByteDance employee told 36Kr that Seed, in some core research areas, will always retain two teams with similar capabilities beyond the required resource allocation. "So even if one team is entirely poached by a competitor, the other can absolutely fill the gap. Even if you leave from the most critical position at ByteDance, it won't make a splash." Headhunters also sense this saturation staffing. One headhunter said that among the companies he serves, Tencent concentrates its talent mining during team-building phases, but ByteDance's recruiting rhythm shows little fluctuation — "whether or not they're short-staffed, ByteDance keeps recruiting continuously."

There's also the "cutting off at the root" tactic. A source from a major firm revealed they are "specifically incubating competitors' model talent." "We've identified the 10 most important people at a competitor and are systematically working to flip them all. If they're willing to jump ship, great; if not, we encourage them to start their own ventures and help them find investors." In his description, this operation is list-based and quota-driven: for example, ByteDance maps out Seed's 10 most important individuals, contacts them one by one. Jumping ship is Plan A; investing in their startup is Plan B. Both plans serve to weaken the opponent. VCs call this approach "investing in the negative round" — funding entrepreneurs before they even start. Some investors, while ByteDance employees are still employed, instigate them to leave, help them build teams, and secure their first funding. Others hang around Tsinghua and Peking University gates, snagging professors and chatting with students, even funding student startups. Compared to losing people to competitors, HRs are relatively neutral about employees being lured away by investors. "Options can bind those who might be poached by rivals, but they can't bind those who truly want to start a company. Those people will leave sooner or later. At least they're not strengthening a competitor," one investor said.

Poaching and counter-poaching are now a required course for every AI company's HR. A researcher at a major firm told 36Kr that some companies' HRs, when targeting a candidate, will coach them verbatim on a script to fend off other companies' HRs. But this is often wishful thinking — some new hires use the offer as "leverage to negotiate a higher salary with the next HR." Between two fiercely competitive companies, a newcomer can even skip interviews entirely and inflate their package based on one offer. Some AI companies have turned recruiting into an "auction": new hires who sign within a week get a 3-month signing bonus; within two weeks, a 2-month bonus; and none after three weeks — all to shorten the candidate's hesitation window.

Headhunters' methods are also constantly evolving. One of the biggest traits of AI talent is their "invisibility" — those who truly fit enterprise needs often don't list themselves on recruitment platforms. "A regular programmer can yield thousands of results on a recruiting site, but for large model talent, you won't find 10, sometimes not even 5," said headhunter Pei Xiaoke. Yet these talents aren't entirely untraceable. Their names appear in author lists of top conference papers, in GitHub contributor lists, and in technical posts on X, with many leaving personal email addresses on their homepages. This means headhunters must change their approach: no longer sifting resumes and making calls, but reading papers and sending emails. However, reading and understanding these papers is extremely challenging, so some headhunters have turned to AI to do the work — using AI to read papers and send personalized emails: "Professor Zhang, I've read your paper, and your insights on multimodality are particularly compelling." This tactic quickly went viral, and all headhunters adopted it. Researchers' inboxes were flooded with AI-scented headhunter emails, and the scheme was rapidly exposed. "They don't even read my papers!" Many such complaints flooded Xiaohongshu. Once the channel was public, it failed instantly. Researchers stopped responding, forcing headhunters to return to basics — relying on personal networks and physical presence. They camp out at Tsinghua and Peking University gates, mingle in professor and student circles, attend conferences, targeted invitation-only events, closed-door meetings, and after-parties. Wherever talent congregates, headhunters are sure to follow.

Choices and Allegiances

Despite the myriad tactics and temptations from HRs and headhunters, AI talent has their own ideas. "I won't go," a ByteDance researcher often tells the rival firm representatives who call him. He's blunt: "Your organizational structure is unreasonable. Come back when you've fixed it." For instance, he once declined an offer because the role was a senior position, but the job description read like an "algorithm gig worker" — simultaneously handling image understanding and a slice of the visual model. "Just applying the visual model to gaming would keep one person busy for five years. If you make someone work odd jobs like that, their data understanding will be incomplete — how can they succeed?" His conclusion is direct: "Only those who don't understand would design an organization that way."

For many AI talents, big tech firms and money aren't the most important metrics. In the late mobile internet era, most people strived for job titles and salaries at large firms, essentially because "there are no more dividends, so everyone has to secure a good position and then hunker down," said headhunter Henry. "But the AI industry is still in a clear dividend period, and capable people are more willing to go where they can quickly deliver results." "Pursuing leverage, not immediate cash," has become a prevailing mindset among AI talent, especially the young "whiz kids." "If I can participate in the core training of a SOTA model this year, that's my biggest leverage — future opportunities will only multiply." This thinking has tipped many young people's scales toward startups like Moonshot AI, DeepSeek, and Zhipu. The people in key positions at these companies are indeed trending younger, more "native AI."

In their view, what's the biggest difference between big tech and startups? The answer might be "organization." How the organization is designed and who is placed in core positions often directly reflects how a company intends to operate. Typically, startups have leaner, flatter, more agile, and more "AI Native" organizations. Alternatively, if a major firm can imbue its large model division with these qualities, it also becomes more attractive to talent. In the context of large model training organizations, "AI Native" essentially means: Does the organization have people who have actually built large models themselves? Do they have experience in a true large-model organization? Do they understand what an organization that fits the AI era should look like? ByteDance, eager to break through in large models in 2024, brought in Wu Yonghui as the head of Seed the following year. His predecessor, Zhu Wenjia, was a seasoned algorithm and architecture expert, but his technical experience was largely concentrated in "search, recommendation, and advertising" — he hadn't truly worked on large models when he took over Seed. Subsequent events proved that large model training still requires someone who has done it directly; Wu Yonghui, a former research VP at Google DeepMind who was deeply involved in Gemini's development, was deemed the right fit.

Tencent made an even bolder move in 2025. The company, known for its "maturity and stability," recruited 28-year-old Yao Shunyu, a former OpenAI researcher, to lead its Hunyuan large model team amid lagging performance in late 2025. "The top office brought him in precisely to break the status quo," said a Tencent insider. His arrival was like dropping a catfish into a still pond. Tencent's large model organization began to churn dramatically. First came a wave of native AI talent. According to 36Kr's analysis, within nearly a year of Yao's arrival, the Hunyuan system introduced at least 8 key external talents reporting directly to him. Reviewing this list — whether it's Zhang Chi and Huang Qi from ByteDance's Seed, or the widely recognized overseas star researchers Pang Tianyu, Tian Yonglong, and Lin Xudong — all are research talents with genuine hands-on experience training large models on the front lines.

"Yao Shunyu has a star effect; he attracts more recipe-holders willing to come to Tencent," said a Hunyuan researcher. He continued the kitchen metaphor: "If the recipe-holders feel the spatula isn't the spatula they want, or the pots aren't the pots they need, how can they cook a great dish? Now, they trust that Yao Shunyu is at least a qualified head chef." Old-timers are also leaving. Key leaders from various large model positions, including former Tencent AI Lab deputy director Yu Dong, former Hunyuan head Jiang Jie, former multimodal understanding head Hu Han, and former post-training mainstay Xu Can, have all seen changes — departures or transfers — since Yao took over. Hunyuan's organization is rapidly becoming younger. "Those so-called veterans don't actually know more than we do; in some ways, our understanding of model training even surpasses theirs," one Hunyuan researcher told 36Kr. "Not all seniority brings benefits; many haven't actually won battles in large models." An employee at a top model company also lamented that the team is almost entirely post-95s, with even some born in 2005.

For startups, the last thing needed is someone who just issues orders. They don't aspire to big tech because "the organization is bloated, with many layers, and managers are just 'managing teams'." "In a startup, because goals and working styles are highly aligned, it's easier to deliver results." Clearly, the large model industry is building a flat, front-line approach where people are judged by training track record rather than tenure. Those who previously worked on virtual avatars, recommendation systems, ad systems, search, or cloud in the last era can no longer trade on seniority and age to lead a large model team. The effects of this organizational change quickly surfaced on the business side. Before 2026, evaluations of the Hunyuan model were almost uniformly dismissive. But after Yao Shunyu and his new core team took over, the Hy3 model was deemed "at the table" — a fair assessment being that it transformed from a "benchmark-listing product" into a foundational model that could actually be embedded in productivity tools like WorkBuddy. Two months later, the Hy4 preview was viewed by the industry as pushing Tencent into the top tier of domestic open-source models. This is a story of "a key person's arrival changing the organization, the new organization changing the business state, and ultimately changing the company's industry ranking." "It's not that Yao Shunyu single-handedly changed Hunyuan, but if at that point in 2025, there was no one given such authority to cut through internal factions and take on this responsibility, Hunyuan would not be where it is now — it would only be further behind," a source close to Tencent commented.

Alibaba finds itself in a narrative heading a different direction than Tencent. Because Alibaba entered large models early and amassed talent, it hasn't suffered much from poaching: the Tongyi Lab's origins trace back to DAMO Academy, and the earliest domestic large model talents like Zhou Chang, Lin Junyang, and Luo Fuli all emerged from there, developed natively by Alibaba. Moreover, its open-source culture attracted many with technical passion, making Alibaba a talent high ground targeted by other big firms in 2024. However, over the past two years, Alibaba's appeal to talent has been loosening. As ByteDance and Tencent reformed their grade systems to break seniority-based constraints, and as startups rolled out SOTA models and went public in 2026, Alibaba's longstanding rigid grade and salary systems have become obstacles to recruitment. It's understood that former Qwen head Lin Junyang held a P10 grade, and aside from his direct reports, most first-line researchers could only reach around P7. If these people moved to ByteDance or Tencent, their grades would be at least one or two levels higher. "Alibaba struggles to place external young talent directly into core positions," said an Alibaba insider.

Since Lin Junyang's departure in March 2026, Alibaba has been restructuring its large model organizations: establishing the Alibaba Token Hub (ATH) business group in March; founding the group technology committee and upgrading the Tongyi large model business unit in April; and merging the Tongyi large model business unit with the Future Life Lab — its two most important model teams — into the new Token Foundry business unit in June. All of these are directly overseen by group CEO Wu Yongming. "Management is indeed anxious," an Alibaba executive told 36Kr, noting that these adjustments centralize AI leadership under the CEO. But is this the best solution? "Mama Wu [Wu Yongming], as group CEO, has too much on his plate, and his attention can easily be fragmented, making it hard to stay deeply focused on AI alone." Many industry observers believe Alibaba still needs to find a more native, more dedicated number-one for large models.

Rising and Falling Tides

The flow of talent in the AI circle is like the tide, with constant highs and lows. According to headhunter recollections, the migration of domestic model talent has gone through four phases, reflecting the rise and fall of companies: In 2023, GPT-3.5 became the benchmark, and employees from overseas giants like OpenAI, Google, and Meta were the hottest targets for domestic companies. Domestic large model talent and companies were just starting out, with Alibaba and Zhipu being more mature due to their earlier entry into AI. In 2024, aside from Alibaba, the "Big Six Tigers" of AI (Zhipu, Moonshot AI, MiniMax, Baichuan AI, StepFun, and 01.AI) emerged from the "hundred-model war," becoming the second-phase "cradle of talent." By 2025, people from DeepSeek, ByteDance, and Alibaba were nearly as sought-after as Silicon Valley researchers. For pre-training expertise, these were the companies to look at. Headhunter York summarized it succinctly: "In today's domestic talent pool, for the model layer, look at Seed; for the application layer, look at CapCut." In 2026, talent from DeepSeek, ByteDance, and Alibaba remains in high demand, but Zhipu and Moonshot AI, which have successfully raced to the front from the "Six Tigers," have again become the focus of the talent market. For instance, Moonshot AI was the first to release an ultra-large-parameter native multimodal model, and right after its release, headhunters immediately camped outside Moonshot AI's offices, ready to poach their multimodal researchers.

Meanwhile, the demand structure of the AI talent market is undergoing subtle shifts — some roles are stabilizing or receding, while others are gaining momentum. In a recent interview, Adam Ward, Chief People Officer at Cursor, noted that beyond AI researchers, Silicon Valley now has a greater demand for "Forward Deployed Engineers (FDEs)" — people stationed at client sites by model companies or cloud vendors to deploy their models or applications into customers' real-world environments. "AI researchers are a very well-defined group; their definition, numbers, and distribution are all clear. What's harder to recruit now are those ambiguous or brand-new roles," Ward said in the interview. This reflects a change in Silicon Valley's AI circle: while continuously investing in people, chips, and money to build the best models, model application and commercialization — making models profitable — are becoming more important than ever. Similar changes are occurring domestically. The "AI Era Skills Trends Report," released in June by Tsinghua SEM and Liepin, shows that in 2022, AI foundational algorithm and model roles accounted for about half of AI workforce demand; by Q1 2026, this dropped to 20%; in Q2, demand for AI Agent product development engineers surpassed algorithm engineers for the first time, becoming the largest single job category in AI. Over the past four years, demand for Agent product talent has grown 40% quarter-over-quarter. The industry's focus is shifting from "those who train models" to those who can "put models to use."

This is also evident in recent organizational and business adjustments at major firms: Alibaba integrated QoderWork, Wukong, and MuleRun into Qianwen Office, elevating "AI office" to a strategic direction; ByteDance merged the Feishu product team with the Doubao product team and integrated TRAE and Kouzi into Doubao, marking its biggest To B adjustment in five years. Both moves are externally interpreted as a counter to Tencent's AI office assistant WorkBuddy. As all three giants vie for this territory, demand for relevant talent is rapidly expanding — not product managers, but product engineers who build the harness to drive models and execute tasks. However, demand for AI algorithm talent at major firms is indeed receding compared to two years ago. It's not that they've stopped hiring; after the frantic recruitment of the past two or three years, talent pools at major firms and leading companies are nearly saturated — ByteDance's Seed alone now has over a thousand algorithm staff. "Companies have all formed structured training teams with clear ladders, so the number of open positions has naturally decreased," said an industry insider. "Major firms' main tasks now are to track only the very top algorithm talent and retain their existing people, rather than continue large-scale hiring."

In Q2 of this year, Tencent and Alibaba's combined quarterly capital expenditure exceeded 100 billion yuan, and both giants' free cash flow turned negative simultaneously — a rare occurrence. Industry estimates suggest ByteDance's quarterly spend is at least on par with Tencent. By this calculation, the three giants' combined quarterly capex exceeds 150 billion yuan, most of which will flow toward computing power for AI. 150 billion yuan could build 1.25 Hong Kong-Zhuhai-Macao Bridges, or 60% of the Three Gorges Dam, or three Fujian-class aircraft carriers. This is also why the relentless, cost-be-damned poaching of large model researchers won't last for many more years — even the landowners are running low on grain. Headhunter Pei Xiaoke has seen it directly: major firms' recruiting standards are more discerning. He used an analogy: in 2023, domestic teams were hiring anyone who had "seen a pig run"; once their understanding matured, the standard became "tasted pork"; now that more people have tasted pork, only those at the "well-fed" level receive the olive branch. As a result, a "shrinking circle" has emerged. For first-line researchers, for example, "C9 university computer science programs are now the first tier," Pei Xiaoke explained. "After C9 comes 985 universities, while 211 and others are largely excluded from the talent pool." The reason is simple: C9 students often practice in more frontier, better-resourced research groups, and their experience is more highly valued.

An algorithm intern at a major firm also observed: AI departments are reducing headcount quotas for full conversion, but those who do receive offers have salary packages that have doubled. This isn't contradictory; it's two sides of the same coin — some basic tasks are gradually being eaten by models, so budgets are concentrated on the few at the top. "This year, some areas, like post-training, are gradually automating, and headcount demand is declining," the intern said, unfazed. "Some groups' headcounts are already frozen, and even 'big names' among your peers will have their resumes rejected." While major firms slow their hiring pace, startups are still in full swing. Having observed ByteDance's talent spillover, they're eager to poach. One headhunter agreed that "ByteDance's era of large-scale expansion is basically over," and his firm is seriously considering whether to terminate its contract with ByteDance next year. "If we sign with ByteDance, we can't poach from them — getting caught means fines and losing the partnership. Compared to serving ByteDance, the returns from poaching their talent for startups look much bigger."

The AI talent war is far from over, but its most feverish phase has passed. In 1983, physicist Emanuel Derman received a call from a Wall Street headhunter asking if he wanted a $150,000-a-year job, while his salary at Bell Labs was $50,000. He stayed at Goldman Sachs until 2002. What he and that generation of physicists and mathematicians gained on Wall Street was more than tripled or quintupled salaries — it was a professional identity that lasted two or three decades. Many of them became billionaires, tenured professors, or hedge fund founders. But will today's AI prodigies enjoy such a long window of opportunity?