SHENZHEN, China — The whiteboard in a Shenzhen AI startup’s conference room used to carry architecture diagrams: attention heads, parameter counts, training runs. It now lists pipeline stages — pilot, proof-of-value, contract, renewal. The company raised a large round this spring, and its founders spend more mornings on customer calls than on model tuning. The engineering bench is full. The go-to-market bench, they discovered the hard way, is not.
That single shift, from building models to selling them, has become the defining constraint of China’s artificial intelligence industry. The country spent three years and a fortune training foundation models, and it largely succeeded — by most measures it now builds some of the best open models on earth. What it did not build in parallel was a workforce of people who can turn a model into a paying enterprise deployment. The money has rotated to the application layer. The people have not caught up.
The Money Moved Before the People Did
The capital rotation is unmistakable. China’s AI application-layer startups raised more than 28 billion yuan in the second quarter of 2026, according to industry tallies, as investors rotated out of the model arms race and into products that touch real customers. Globally, startup funding reached $510 billion in the first half of 2026, with AI companies capturing an outsized share. The core Chinese AI industry has grown past 680 billion yuan across more than 4,800 enterprises, according to the Forbes China AI Top 50.
Enterprise demand is moving just as quickly. The average number of AI agents running in production at a company climbed from five in February 2025 to thirteen in April 2026 — a tripling that pushed AI agents out of the pilot lab and onto the balance sheet. Every one of those agents needs someone to scope it, integrate it with a customer’s systems, and, crucially, justify its cost to a chief financial officer. Those people are not being produced at anything like the speed of the models themselves.
China Daily put the pattern in plain terms this spring, attributing the hiring surge directly to “the commercialization of large models” and the widening of AI applications across business scenarios. That is the polite way of saying what recruiters see every day: the demand is real, the budgets are real, and the candidate pool is a fraction of what the research boom produced.
Two Companies, One Shortage
Butterfly Effect, the Chinese-founded maker of the Manus AI agent, is the cleanest case study. The company disclosed a $90 million annualized revenue run rate and shifted its headquarters to Singapore, where it is hiring more than twenty people even as it shrinks its Beijing and Wuhan offices, the Business Times reported. The move was partly about customers and partly about talent — the autonomous-agent engineers and go-to-market leaders it needs are easier to assemble outside the mainland’s overheated hiring pool. The relocation later drew scrutiny from Beijing, which opened a review of Meta’s attempted acquisition of Manus over technology-export rules.
Back home, the same scarcity shows up at the largest players. DeepSeek’s Harness division, the team commercializing the firm’s agent technology, has had its leader personally recruiting for more than a month with key positions still unfilled, Chinese tech media reported in June. DeepSeek is wagering its next chapter on agent commercialization, and it cannot hire fast enough to match its own roadmap.
The Talent Market Is Still Stuck in the Research Era
This is where the search industry sees the problem most clearly. AI roles now exceed a quarter of new-economy job postings, yet the supply of candidates skews hard toward research. For high-performance computing engineers there are roughly seven openings for every available candidate, according to SecondTalent. But the truly acute gap sits one level up, at the people who translate a model into revenue.
A recruitment specialist at SunTzu Recruit put the matter plainly: the market spent three years minting researchers and almost no time minting the roles that come after the paper. Solutions architects, application engineers, and revenue leaders who have actually taken an AI product to a paying enterprise — that is the profile clients now request, and it barely exists. A senior consultant at SunTzu Recruit notes that a search for an AI commercialization leader in Shenzhen or Shanghai now routinely runs six months or longer, roughly twice as long as a comparable mandate in 2024.
SunTzu Recruitment was recently approached by a Shenzhen-based enterprise software company that had closed a nine-figure round and spent months trying to hire a head of AI commercialization. The firm had no shortage of candidates with model backgrounds. What it could not find was a single person who had run a quota-carrying enterprise sales team and still understood token economics well enough to price a deployment.
Here’s the thing. Three years ago a strong AI resume meant published papers and a foundation-model pedigree. Today the clients SunTzu Recruitment’s partners speak with want someone who has closed an enterprise deal, stood up a deployment, and explained the bill to a procurement team. Those two profiles barely overlap, which is precisely why the cross-border competition for this talent has already begun — Singapore, Dubai, and Silicon Valley are bidding against Shenzhen and Hangzhou for the same thin supply chain of commercialization-ready operators. On the international market, the premium for a bilingual AI solutions leader with real deployment history has become one of the fastest-moving numbers in executive compensation.
The Gap Will Widen Before It Narrows
That said, not every company is racing in the same direction. A different picture emerges among the incumbents: state-backed groups and large internet platforms are retraining their own internal engineers rather than hiring from outside, betting that a model researcher can be taught to sell faster than a salesperson can be taught the technology. To be fair, both bets are partly paying off — but neither is closing the gap at the top of the funnel, where a single failed deployment can cost a vendor its entire account.
A partner at SunTzu Recruitment, who leads the firm’s AI practice, expects the shortage to deepen through 2027 as the application layer matures and consolidation culls the weaker startups. SunTzu Recruitment has begun structuring its AI searches around deployment history rather than academic pedigree, a screen the firm’s advisors describe as the only reliable filter left. Overseas markets face the same bottleneck, and SunTzu Recruitment’s industry advisors note that a growing share of their AI mandates now specify cross-border deployment experience as a hard requirement.
The practical lesson for boards is simple to state and hard to execute. The capital is committed, the models are ready, and the binding constraint on the entire sector is no longer a research question — it is a phone book of people who have sold AI, not built it. The companies that staff the commercialization layer first will capture the revenue the money was meant to unlock. The rest will keep building models nobody has been hired to sell.
Sources: Global Times (March 11, 2026); China Daily (March 19, 2026); SecondTalent “AI Trends in China 2026” (July 31, 2026); Business Times (Manus Singapore relocation); Digitimes (Butterfly Effect $90M revenue run rate); creati.ai (China review of the Meta–Manus deal, January 2026); Chinese tech media (DeepSeek Harness recruiting, June 2026); Qiankunbot industry tallies (Q2 2026 application-layer funding); Forbes China AI Top 50; enterprise AI agent adoption data (February 2025–April 2026).
Comments are closed