Vol.161 The AI Talent War: Talking with 肖玛峰 About Where the World's Top AI Talent Goes—and Whether It Stays

Summary

• Demand for AI talent is not surging all at once; it is moving through three stages: foundation models, application startups, and traditional industries. In 2023, hiring was concentrated mainly at foundation-model companies; in 2024, application companies that raised money the year before began expanding. 李翔 predicts that 2025 will be “the year when large numbers of traditional companies need AI talent,” a view 肖玛峰 shares. TTC has mapped more than 1,000 companies, with a focus on 300–400 of them; in 2024, AI-related revenue was about RMB50M, less than half of total revenue.
• The supply shortage has pushed top AI talent into an obvious bubble, with more of the bargaining power now in individuals’ hands. According to 肖玛峰, Alibaba’s P9 周畅 saw compensation triple after moving to ByteDance; Xiaomi was rumored to have offered DeepSeek’s 罗福莉 an annual package in the tens of millions of yuan, with startups often adding founder shares. The most extreme logic is: “Even if this person only stays for six months and helps raise one round, that’s already enough for the boss.”
• The economics of bringing talent home from overseas once made no sense at all, but the calculus began to shift at the margin in late 2024. A Bay Area “dual-coder” household can earn $700K–$800K a year, work from home, and benefit from rising tech stocks. Returning to China could mean lower pay, more work, and the risk of being laid off at 35, so there was “no reason at all for them to come back.” Now that major Chinese companies are genuinely willing to pay for AI leaders, alongside workplace exclusion and immigration-policy changes in the US, 肖玛峰 believes a new wave of returnees may emerge.
• The pool of AI buyers has expanded from major internet companies to smartphones, autos, robotics, and traditional industry leaders, while traditional companies truly need talent with cross-functional, scenario-specific expertise. Companies including Taikang, ENN, Haier, Midea, and Wanhua Chemical have already put in requests. In AI for Science, 肖玛峰 says R&D cycles could be cut from three years to three to six months. Application startups are hiring most urgently because “without these people, they die”; by revenue, however, large companies remain better customers.
• DeepSeek has both broken the superstition that only big names can deliver and demonstrated that leaders still provide organizational leverage. It has forced companies to reconsider the ceiling of young talent, while 贝壳’s hiring of 叶杰平 showed that even if a marquee figure later leaves, they may leave behind an entire team and a long-term technical foundation. “Find a leader like that, and sometimes the moment he arrives, he can bring a whole group with him.”
• Talent defense cannot be solved with a non-compete agreement; the core question is whether a company can keep offering an advanced direction while creating value with the individual. 肖玛峰 argues that what companies should truly protect is intellectual property and the confidentiality boundary, not “tie people up” and leave them no way forward. A non-compete can at most change the economic cost of leaving; it cannot stop a competitor willing to pay damages while strengthening itself and weakening its rival. The long-term relationship between companies and talent is “mutual enablement, mutual淘汰.”
• TTC is betting not on a single headhunting boom, but on AI turning executive recruitment into a scalable talent-allocation network. Its agent “小麦” has cut some interview coordination from three hours to ten minutes. Based on the 2024 figures 肖玛峰 gave earlier, roughly 100,000 people were recommended, 10,000–20,000 were interviewed by clients, and more than 1,000 ultimately joined. The long-term target is 100,000 placements a year, which at RMB50K per person would imply RMB5B in revenue; he later answered “100,000” as successful hires, however, creating an internal inconsistency in the figures.

Deep dive

1. The Overall Hiring Market Is Contracting, With AI Among the Few Sources of New Demand

• 李翔’s industry baseline is that about 70% of mid- and senior-level recruiting replaces existing roles, while 30% comes from new business. As real estate, finance, and internet companies contracted simultaneously, businesses cut back on second and third growth curves, with new demand hit especially hard.

• 李翔 contrasted this with Wanda’s expansion era: projects once targeted completion in 180 days, with discussions of opening more than 100 urban complexes across as many cities in three months. Each project needed a project head, chief cost officer, finance head, and engineering head. That era of hiring driven by scale expansion is over.

• The contrast depends on which circle one inhabits. 肖玛峰 once told 李翔 that talking with people from real estate made the present feel like “a historical garbage moment,” while being surrounded by AI entrepreneurs revealed “so many opportunities to rebuild so many industries from scratch.”

2. 2023 Was About Mapping the Talent Pool; Hiring Took Off in 2024

• Demand began in late 2022 and early 2023. After GPT was released, Source Code Capital asked TTC to map NLP and foundation-model talent at China’s major internet companies. The immediate question was not who to hire, but whether China had teams capable of replicating what was happening in the US.

• After Source Code invested in 王慧文’s Light Years Beyond, he called directly: “Max, we’re going all in on AI. Get to work for us.” TTC then went to Seattle and Silicon Valley to connect with Chinese talent at institutions including OpenAI, but 王慧文’s reputation in China did not automatically translate into recognition among overseas talent.

• In 2023, the main employers were the “Six Little Dragons” of AI, whose founders mostly came from search, Google, Microsoft, and related backgrounds and already had teams of their own. Hiring volumes were limited. TTC therefore shifted to charging for connections: arranging online or offline meetings between companies and top US-based talent, and charging a process fee once the meeting took place.

• At the end of 2023, investors began betting that Perplexity- or Character.AI-style applications would emerge in China. The relevant companies raised money and expanded in 2024, even though fewer companies raised money that year overall. TTC drew a battlefield map spanning chips, compute, algorithms, foundation models, and applications, then selected 300–400 priority clients from more than 1,000 companies.

3. Overseas Talent Lost Its Incentive to Return, but the Door Is Starting to Open

• Around 2019, the market capitalization and total compensation offered by Chinese internet companies could still approach those of leading US companies. Returning to China also meant a higher title and bigger opportunities, and a genuine wave of returnees did emerge. By 2024, those conditions had reversed.

• The typical “dual-coder” households 肖玛峰 encountered in the Bay Area earned $700K–$800K a year, worked from home after the pandemic, and benefited from the rise in US tech stocks—a combination he described as “more money, less work, and close to home.” Returning to China could mean lower pay, more internal competition, and the risk of being laid off at 35. Chinese companies had “no competitiveness at all,” and these people had “no reason at all to come back.”

• The outreach in 2023 therefore brought back few people. In 2024, Microsoft moved many employees from its China headquarters to Canada, New Zealand, and other locations to relocate R&D centers. By year-end, however, Alibaba’s 徐主红, 谷歌 recruit 吴永辉, and former DAMO Academy chief scientist 斯罗, who returned from Salesforce, began to offer a different signal.

• Another force 肖玛峰 heard about was the stronger cohesion of the Indian executive network in Silicon Valley, which could leave Chinese professionals squeezed in senior-level competition, alongside changes in US immigration policy. At the same time, Chinese entrepreneurs had “seen big money and were willing to spend,” allowing supply and demand to combine in driving a potential return wave.

4. The Buyers of AI Talent Now Fall Into Four Tiers

• The first tier remains major internet companies such as ByteDance, Kuaishou, Alibaba, and Tencent, which continue investing in generative AI for text, images, and video. The second consists of smartphone, auto, and other hardware companies extending foundation models into embodied intelligence and robotics.

• The third is AI application startups. Investors are once again watching major internet companies for people willing to start businesses, hoping to identify the next generation of AI-era CEOs. These companies need products, demos, and core talent to secure their next funding round, making recruitment a matter of survival.

• The fourth is traditional industry leaders. Companies with tens or hundreds of thousands of employees are particularly anxious in 肖玛峰’s conversations: organizations that were once an advantage may now be a liability. Management teams worry about whether employees are using AI tools effectively and whether a massive workforce can still be an advantage.

• Outside AI, active demand also includes senior talent for overseas expansion. Further down the AI value chain are embodied intelligence and robotics. 肖玛峰 believes startups that have raised several hundred million yuan have not locked in victory: Ideal has proposed a RMB10B investment, while hardware giants including BYD, Xiaomi, and Huawei could also enter.

5. Traditional Industry’s Moat Is Its Scenarios, Not Its Foundation Models

• 肖玛峰’s clients include Taikang, ENN, Haier, Midea, and Wanhua Chemical. Their biggest hiring challenge is finding people who understand both industries such as chemicals and AI. The practical solution is usually to have industry specialists and AI teams work together and learn from one another.

• He has seen candidates who studied chemistry as undergraduates and switched to AI for their PhDs. These hybrid backgrounds are rare but better suited to industrial demand. Traditional companies do not need the very top foundation-model researchers, nor the math-competition champions sought by quantitative firms.

• Traditional companies’ first advantage is access to rich, real-world scenarios. AI startups still face a long path through R&D, engineering, and commercialization; joining a large industrial company provides direct exposure to real business and the deployment chain. The thinking and knowledge of hybrid talent are already enough to create a clear advantage inside traditional companies.

• AI for Science is one of the productivity unlocks 肖玛峰 cares most about. Using protein-structure simulation and chemical R&D as examples, he believes product-development cycles could shrink from three years to three to six months. If achieved, competitors would not face a modest productivity gain but something that would “simply crush them.”

6. DeepSeek Repriced Both Young Talent and Leaders

• DeepSeek prompted a major-company HR vice president to ask directly: if it relies mainly on young people, is it still necessary to spend heavily on marquee hires? 肖玛峰 accepts the challenge—fame is no longer a prerequisite for breakthroughs, but leaders retain another form of value.

• He cited 贝壳’s hiring of chief scientist 叶杰平 from Didi. 叶杰平 found that buying a home was essentially about filtering listings by location, size, and orientation, leaving AI with limited room to contribute at the time. He joined Alibaba’s DAMO Academy about two years later, but many of his “disciples and grand-disciples” stayed at 贝壳 because he had joined.

• That team helped 贝壳 move quickly into opportunities such as spatial intelligence and image generation for home renovation. A leader’s value therefore lies not only in personal output but also in branding, recruiting, and organizational leverage: “After he left, these people eventually stayed at 贝壳.”

• 李翔 questioned whether DeepSeek’s rise involved marketing or PR orchestration. 肖玛峰 initially said there “should not have been any top-down” direction, then explained the spread as a shared need: the industry, ordinary people, and even the country all needed an uplifting story. The result was that even his brother back home, a construction contractor, began discussing “how to do AI.” He also noted that a presentation before DeepSeek showed most people were merely learning about AI through the news, not actually using it.

7. The Big-Tech Talent War Also Exposes Spending Gaps and a Valuation Bubble

• 肖玛峰 says ByteDance has invested the most money and time in talent. He heard that 张一鸣 planned to build a team of 1,000 top AI researchers from around the world in Singapore last year, and believes no other major internet company can match that commitment. Tencent and Alibaba are catching up; 李翔 cited reports that Alibaba plans to invest RMB380B in AI.

• The competition has reached directly into rivals’ core teams: ByteDance has recruited model talent from 01.AI and Alibaba’s Tongyi Qianwen, while Xiaomi was rumored to have offered DeepSeek’s 罗福莉 an annual package in the tens of millions of yuan. When candidates decline, TTC even gives advisers a line to use: “Don’t reject it too quickly. It might be the next 王兴兴.”

• But 肖玛峰 explicitly sees a bubble. 周畅 was reportedly around P9 at Alibaba and saw compensation triple after moving to ByteDance—“it just doesn’t sound normal.” Equity further magnifies individuals’ optionality; in some cases, a person’s halo only needs to support six months of operations and help the company complete one financing round for the economics to look worthwhile to the founder.

8. Non-Competes Raise Transaction Costs but Cannot Stop Strategic Hiring

• California has no non-compete agreements. In domestic transactions, the receiving company may assume the compensation liability directly, or structure the employment contract in California or elsewhere. For a well-funded competitor, paying damages is simply part of the cost of acquiring talent.

• 肖玛峰’s position is unequivocal: companies should rely on confidentiality agreements and IP rules to stop employees from taking their former employer’s work to a competitor, but should not “tie people up and leave them no way to live.” Some major companies have extended restrictions beyond the internet into consumer products, effectively banning employees from continuing in their entire career field.

• The imbalance is greater because employees often do not see the non-compete when accepting an offer and are only asked to sign it after joining. TTC now has to check the scope of restrictions in advance. Candidate concerns and clients’ unwillingness to assume liability can both cause projects that consumed substantial time to fail at the final step.

• The former employer’s real leverage is usually economic recourse. If roughly two-thirds of prior income came from long-term stock compensation, the company can retain the right to seek repayment later. But for a competitor, hiring someone both strengthens itself and weakens its rival, so it may still accept the cost after running the numbers. Faced with a so-called “cut-off-a-finger” strategy, companies have almost no reliable defense.

9. The Top End of the Talent Supply Must Be Cultivated on Campus and in Competitions

• DeepSeek did not wait until it became famous to seek young talent. 肖玛峰 says 幻方量化 had already built a team dedicated to maintaining long-term relationships with math-competition winners. Huawei, Meituan, Alibaba, and Tencent likewise have people tracking the best students nationwide, while DJI has done especially well in hardware talent.

• This is a list compared every year: how many people each company hired, and who went overseas. Huawei’s “Genius Youth” program may offer more than RMB2M, but leading overseas quantitative firms can pay more, so a large share of candidates ultimately leave the country.

• The top quantitative circle is small enough that specialist recruiters can know “exactly who is who.” Continuously managing relationships and data is itself an asset. Companies also co-build laboratories with universities and meet students early to lock in supply before graduation. Broad-based programs such as Tsinghua and Peking University delaying specialization for the first two undergraduate years, along with Zhejiang University’s Chu Kochen Honors College and Peking University’s Yuanpei College, are seen as another route.

• The conversation ended with an unresolved question: AI may currently be “a game for a small number of people,” with perhaps only several thousand to tens of thousands making the core list. Whether it is good or bad for overall employment will not be clear until large-scale deployment arrives.

10. Retention Depends on Co-Evolution; the Model-Company Race Is Far From Settled

• 肖玛峰 summarizes the relationship between companies and talent as “mutual enablement, mutual淘汰”: a weak company quickly loses its employees, while employees leave when they see the company lose its direction. The real retention mechanism is for a company to keep possessing advanced productivity and an advanced direction of development.

• He does not believe DeepSeek’s emergence means the “Six Little Dragons” should give up. “The war has only just begun.” Tencent, Alibaba, and ByteDance can follow first, then invest in or acquire competitors if necessary; major internet companies have multiple ways to re-enter the contest.

• 李翔 mentioned that Baichuan AI is moving into healthcare and that he recently heard it had acquired several hospitals and entered the service layer directly. 肖玛峰 believes that if it embeds itself in specific businesses rather than focusing only on general-purpose models, it could become more competitive than traditional-industry rivals. Model startups do not necessarily have to remain committed to general-purpose models; they can ultimately find their own ecological niche.

• 李翔 added an investment perspective: assess teams dynamically rather than judging the product at a single point in time. Pop Mart initially sold trend-market trinkets such as glasses and socks in physical stores, then evolved into an IP-operations company. What matters is whether the team can learn, adjust, and continue attracting new people.

11. Top Talent Is Becoming Partners Rather Than Employees, and Their Moves Are More Private

• 肖玛峰 has spent 18 years in recruiting. Early on, he would repeatedly question a candidate who had stayed in a job for only one year; in the mobile-internet era, a tenure of more than three years already counted as stable. AI talent is now moving even faster.

• For top talent, employment relationships are giving way to partnerships. Companies need to share equity, while getting someone to collaborate for a year may already be a good outcome. Incentives designed around long-term subordination cannot match the optionality these people possess.

• The smaller the circle, the more sensitive a move becomes. A former colleague making an introduction may cause word to travel rapidly back to the original employer. 肖玛峰 imagines that everyone will eventually have a career AI, with the candidate’s AI and the company’s AI shaking hands privately. Salary negotiations, however, will still require an intermediary to prevent both sides from offending each other with their first offer.

12. AI Is Rewriting Recruiting Itself

• 肖玛峰 first corrected the positioning: TTC is not a traditional headhunting firm but an AI recruiting platform. His industry view is that the PC era had ChinaHR.com, Zhaopin, and 51Job; the mobile-internet era had Liepin and BOSS Zhipin; the AI era will also produce new large-scale players.

• TTC integrated GPT at the end of 2023, then added DeepSeek and built its own agent, “小麦,” covering internal advisers, client groups, and candidates. The goal is not to add tools to isolated steps, but to become a platform driven by agents that every participant can interact with.

• A typical interview-coordination process once involved three rounds of waiting among the client, adviser, and candidate, consumed three hours, and could still miss the available slot. Now a bot can collect information from the group and, in some cases, contact the candidate directly, compressing the entire process to ten minutes.

• 肖玛峰 wants the company to “operate like AI”: respond immediately after a client submits a requirement, and even if the first answer is not accurate enough, continue refining it through two more prompts. Borrowing 张一鸣’s phrase “develop company as product,” he is turning delivery speed itself into a product differentiator.

13. The Commercial Ceiling of AI Recruiting Depends on Placement Volume, Not Per-Deal Pricing

• TTC’s smallest clients pay about RMB100K a year, while some startups pay several million yuan annually. In the most extreme case, a client had no money in the account but promised to raise financing if it found a key hire. The adviser found a CTO, the company completed its financing, and paid a RMB550K service fee.

• When investment firms ask TTC to map talent at major internet companies, they pay several thousand yuan in “tea money” for each meeting. If an investment is ultimately completed, they pay another intermediary fee similar to a light FA fee. 肖玛峰’s view of partnership is: “Make a little more sacrifice; suffering a loss is a blessing.” Long-term introductions and trust are often worth more than a one-off fee.

• 肖玛峰 mentioned that Mercor, an AI recruiting company in the US founded roughly two years ago, is already valued at $2B. The gap is not only technical but also behavioral: the SaaS model works in the US, while Chinese customers are often unwilling to pay for tools and demand that someone deliver the service directly.

• TTC’s current candidate packages can reach RMB10M–20M, and 肖玛峰 has previously placed core executives whose equity value exceeded RMB100M. But cash fees can sometimes be calculated only on base salary, so volume remains essential. Based on the 2024 funnel he gave earlier, roughly 100,000 people were recommended, 10,000–20,000 were interviewed by clients, and more than 1,000 ultimately joined. If AI expands placements to 100,000, at RMB50K per person that would imply RMB5B in revenue. He later answered “100,000” as successful hires, however, so the figures are inconsistent.