3 million vying for PhDs, post-95s are already "old": AI recruitment is "burying" middle management

Written by: Ada, Deep Tide TechFlow

“An internet giant has offered over 60 recent PhDs with AI backgrounds annual salaries of 3 million this year.” When TTC founder Xiao Mafeng, a headhunter who has served over 1,500 AI companies, mentioned this number, his tone was very flat, as if reporting the day’s weather.

In the same month, Maimai data showed AI job postings skyrocketed 29 times, and Zhaopin reported a 200% surge in job seekers. 29 times more jobs, 200% more applicants—these numbers look as beautiful as a bull market’s candlestick chart.

But these figures hide a secret: a massive influx of capital and attention is pouring into a narrow funnel. The few dozen at the top of the funnel are driving up the entire market’s salary expectations, while the hundreds of thousands at the bottom are bearing all the anxiety.

Meanwhile, those in the middle of the funnel—people with five or ten years of experience—are quietly being drained.

The prosperity of the talent market is an illusion; the illusion of liquidity is real.

One star is hard to find, thousands compete fiercely

According to Liepin’s report, 47% of AI positions require master’s or doctoral degrees, and nearly half of companies only recognize 985/211 universities.

Headhunter Eva is more direct: “Big companies are hiring, 211 is barely acceptable, at least 985. Resumes without vertical project experience are basically not considered.”

What does the top look like?

On the day Alibaba’s Qianwen announced its departure, “people from all major companies came to us, asking if we could help contact Lin Junyang,” Xiao Mafeng recalled.

There are probably only dozens of such talents nationwide. To find them, headhunters have long stopped browsing resumes. They scour GitHub for code commits, track paper authors on Google Scholar, infiltrate podcast listener groups and AI startup communities. Eva even joined a Tsinghua AI startup competition group, full of 21-22-year-olds. “Now we chat early, two or three years ahead, when they might need a job, so we get a head start.”

Another headhunter, Steve, who started recruiting in AI in 2022, said something profound: “I highly doubt there will be resumes in the future.”

He gave an example: In January this year, a company wanted to hire someone familiar with OpenClaw. The field is so new that no one would list it on a resume. His approach was to break down the requirements—at its core, it’s a multi-agent framework problem. Has anyone built a similar framework? Is it open source? Who are the contributors in the open source community?

Resumes are losing value, traditional recruitment channels are failing.

Someone has spotted an opportunity in this crack.

Sam, co-founder of DINQ, started his venture from a similar observation: The top AI paper authors are often not from prestigious schools, some even dropped out, young, without titles—it’s hard to tell how talented they are just by their resumes. LinkedIn’s logic of looking at education and experience doesn’t work for AI talent.

So, Sam created DINQ, a “LinkedIn for AI scientists and developers,” which focuses on achievements—top conference citation counts, GitHub contributions, and whether collaborators are top-tier technologists. HR inputs “Sora 2” into DINQ, and the platform extends to related paper authors, not just those with experience related to Sora 2, uncovering hidden talent.

Xiao Mafeng’s alternative approach is build in public: directly showcasing your product as the best proof of capability.

Although 621 universities now offer undergraduate AI programs, McKinsey predicts a shortage of 4 million AI talents in China by 2030. But the word “shortage” is misleading; what’s needed are experimental scientists who have trained on hundreds of thousands of data cards, who understand the limits of large models and can find commercial applications. People who say “I’m very interested in AI” after listening to two podcasts are never in short supply.

Ye Xiangyu, founder of Niuke, summarized accurately: The top “star” is hard to find, while at the bottom, “thousands compete fiercely.” Maimai’s statement that “one suitable candidate can match every two AI positions” refers to the top tier. As for the bottom? No one tracks it because those resumes never make it into the system.

Leverage pricing: the closer to the model, the more valuable

So, where is the money flowing?

Eva provided some numbers. For big companies at P7 level, non-technical roles have a ceiling of about 1 million yuan. For AI technical roles at the same level, 1.5 to 2 million. The salary increase when switching jobs is even more significant—50% is common, doubling is possible; non-technical roles see only a 10-20% rise, rarely exceeding 30%.

Steve used a word to explain this pricing logic: leverage.

Imagine the model as the sun. The closer you are to the core, the greater your leverage and value. A core researcher’s improvements in model capabilities can impact a company’s market value by billions. Running tens of thousands of data cards costs far more than their salary. From this perspective, paying them a billion isn’t expensive.

What about those farther from the sun? Product managers, operations, sales—they don’t have such direct leverage, so their salaries are naturally limited. Steve estimates that at the application layer, the salary gap between technical and non-technical roles can be two to three times.

Xiao Mafeng added a key variable: he believes this “hierarchy” is fundamentally driven by supply and demand, with two layers. On a macro level, only a few people have trained on hundreds of thousands of data cards, so their salaries are sky-high. But micro-level, it depends on the founding team’s DNA. If the founders are Tsinghua professors with many technical talents in the lab, those who can commercialize are more valuable.

The scarcity of a few dozen people defines the entire industry’s salary narrative. Others take this narrative as a benchmark, but what’s measured is only the gap.

A purge of “Old Deng”

“AI era rejects Old Deng,” Xiao Mafeng offered a sharp critique.

The previous AI wave involved companies like Megvii and SenseTime, now mostly in their 40s. Their experience has become a burden.

Steve’s words are more tactful but consistent: “We don’t believe you can find new continents with old maps. People who have worked in the industry too long have too much momentum and inertia. Their brains’ immediate reactions are results of reinforcement training, but times have changed, and the right response might be completely opposite.”

Age anxiety has penetrated every level. Some investment firms are seeking post-00 entrepreneurs, and “post-95s are already old,” such remarks are emerging.

It sounds absurd, but the signals from the job market are real: when resources are limited, the scale tilts decisively toward the young.

“Now, the competition is about speed of execution and implementation. Everyone is training special forces, not a big army,” Steve said. Special forces don’t need many commanders.

But there’s an unspoken contradiction nobody wants to face.

Those who truly implement AI products and turn technology into business value rely precisely on industry experience, tacit knowledge, and learned lessons. Steve admits these tacit skills are found in more experienced people—they may not know exactly which road to take, but they know which ones to avoid.

The industry needs the drive of young people and the judgment of veterans. Everyone can say this, but the flow of money only favors the former.

The middle layer is being swallowed

All three headhunters mentioned a common change: management layers are shrinking.

“Pure management roles are probably already difficult. Many things are being overturned; the systems you built might be dismantled tomorrow,” Steve said.

Organizations are becoming extremely flat, no longer needing hierarchical pyramids with layers of reporting, but small teams capable of fighting on the front lines. Relying on people to do tasks is less effective than deploying an Agent. Previously, strong management skills and complex teams were essential, but that’s being challenged.

The boundaries between product managers, operations, front-end, and back-end engineers are blurring. One person can now run an MVP of a product using AI.

Chen Lei (pseudonym), a product director at a mid-sized AI company, managed an eight-person team. Earlier this year, after a restructuring, her team was disbanded—four moved to Agent products, two were laid off. Her title changed from “Director” to “Senior Product Manager,” reporting to a technical lead five years her junior.

“I’m not laid off, but I know this is even harder,” she said. “What I built in three years is gone with just an organizational change. And I can’t complain because they’ll say, ‘You’re still here, aren’t you?’”

This is the cruelest part of the liquidity illusion. At the top of the funnel, dozens of geniuses are fiercely contested with sky-high salaries. At the bottom, hundreds of thousands of newcomers can’t even get in the door. And in the middle, those with five, ten, or even fifteen years of experience are being internally sidelined.

The career ladder has lost several middle rungs. Instead of climbing step by step, it’s now a parachute jump—either land at the top or fall freely.

Who is creating this illusion?

Who benefits from this liquidity illusion?

Recruitment platforms generate traffic with headlines like “AI job postings surge 29 times” and “Talent shortage of 4 million,” pushing more anxious job seekers into the funnel with every share.

Companies use AI as a shield. Forrester Research found that 55% of employers regret layoffs driven by AI, often because the AI capabilities they replaced were unprepared. Resume.org’s survey is more direct: 59% of companies admit to disguising layoffs as “AI-driven,” because it sounds better for stakeholders. Saying “AI” sounds like a strategic upgrade; citing poor performance sounds like management failure. AI has become the best excuse.

Klarna laid off 700 people claiming AI replaced customer service, but service quality plummeted, and customers revolted. They quietly rehired some. This isn’t an isolated case. Forrester predicts that half of AI layoffs will be quietly reversed, often with lower salaries or outsourced overseas.

Steve accurately summarized the current mindset of bosses: “Their first question now is whether to hire, then what to hire.”

According to Forrester, only 16% of global employees are highly AI-ready. Companies don’t invest in training; employees rely on self-learning. Generation Z has the highest AI readiness at 22%, but they are the first to be pushed out of entry-level roles—precisely the easiest to be replaced by AI. Mercer’s survey shows anxiety about AI-induced unemployment rising from 28% in 2024 to 40% in 2026.

AI is both a reason to hire and a reason to lay off. Whoever controls the definition wins the game.

The funnel won’t widen

Returning to the initial numbers:

29 times more jobs, 200% surge in applicants, 3 million yuan salaries, 4 million talent shortage. Each number is true, but together they tell a different story: jobs are increasing, but the opening is extremely narrow; applicants are flooding in, but most can’t even pass screening; salaries are soaring, but only for the top few dozen; shortages are widening, but what’s needed and what’s supplied are mismatched.

But this funnel won’t widen. AI technology iterates every six months; today’s hottest direction might become obsolete in half a year. You might think you’re close to the sun, but with a new model release, you could be pushed to the periphery.

Steve said something that can serve as both a eulogy and an entry ticket: “Using tenure to measure experience may no longer be enough. What matters is the density and depth of your interaction with AI. Someone who entered the industry four years ago but only used AI casually—versus someone who just joined last year but is fully immersed—who has deeper experience?”

Even the three headhunters are being reshaped by this industry. Eva is studying algorithm principles, Steve is researching Agent frameworks, and Xiao Mafeng just left a meeting with a post-00 entrepreneur, exclaiming, “Their understanding has already reached the next level.” Even those selling shovels must keep pace with the gold rush.

Chen Lei recently started a small project on GitHub—an automated legal document generator using an Agent framework. No one asked her to do it, and she’s not paid. She said she realized one thing: instead of waiting to be filtered by the funnel, it’s better to carve out your own hole.

This might be the only slightly optimistic part of the entire article, but only just.

Most people are neither among those 60 PhDs earning 3 million nor like Chen Lei, who still has the ability and willingness to carve a path. They are the silent majority in the middle of the funnel—insufficiently top-tier to be fiercely contested, not resolute enough to start over.

This funnel will not widen.