Yu Pi never expected that after accepting an offer with a monthly salary of CNY 50,000, his first day on the job would involve going to a factory to stick QR codes.
The position he applied for is called FDE, or Forward Deployed Engineer. The job description required production-level Python coding, experience with large model systems, and building RAG (Retrieval-Augmented Generation). The interview focused on Agent architecture and Prompt engineering. However, upon arriving at the client site—a manufacturing factory—the real work scenario turned out to be diving into the workshop, recording the "nicknames" of each piece of equipment amidst the roaring sound of machines, and personally sticking QR codes onto the devices for positioning.
There are no office buildings, no floor-to-ceiling windows, and no ready-made data pipelines. Everything about the FDE role seems far removed from the spotlight of AI.
Yet, on social media, golden narratives about FDEs keep emerging. Terms like "the most secure job in the AI era," "monthly income of CNY 100,000," and "composite talent" have never left this position. Reportedly, over the past two years when the entire industry has been laying off employees, the recruitment demand for FDEs has surged by more than 40 times. OpenAI offers starting annual salaries equivalent to over CNY 1 million, while ByteDance offers a maximum monthly salary of CNY 70,000 for its "Doubao AI Large Model FDE" positions.
A position that was almost unknown in the past has become the most sought-after "golden ticket" in the AI industry in just two years. What exactly is it? Why has it risen against the trend amidst the wave of layoffs? And how much of the glamorous recruitment data is merely hype?
The Golden Ticket Amidst the Layoff Wave
In 2026, the tech industry is experiencing an unprecedented divide.
On one side, the wave of layoffs continues to spread. According to the monthly report from the U.S. outplacement consultancy Challenger, in the first half of 2026 alone, the cumulative number of layoffs in the U.S. tech industry approached 140,000. Amazon cut about 30,000 jobs within a year, Meta laid off 8,000 employees, and Oracle's total headcount dropped from 162,000 to 141,000 over the past twelve months, a decrease of 13%. Oracle reported in its annual report: "The application and deployment of AI technology in our operations have resulted in, and may continue to result in, a reduction in employee headcount."
On the other side, a position named FDE is surging against the trend at an astonishing speed. LinkedIn's 2026 Workforce Report shows that between 2023 and 2025, global FDE job postings grew 42-fold, while the growth for AI engineer positions during the same period was only 13-fold. ByteDance offers FDE experts a monthly salary of CNY 30,000 to 50,000 with a 15-month annual salary package. Ant Digital Technologies offers a monthly salary of CNY 40,000 to 60,000, and the FDE lead position at Zhipu AI commands a monthly salary of CNY 60,000 to 80,000. In the United States, OpenAI provides annual salaries ranging from USD 162,000 to 280,000 for FDE roles, while Anthropic offers USD 200,000 to 300,000.
When Yu Pi saw these numbers, he was sitting at his desk in his previous company, scrolling through his phone. It was a mid-sized B2B services company where he had worked as a backend developer for six years, writing Python, debugging APIs, and building some small-scale RAG projects. The company conducted a round of layoffs at the beginning of the year; he stayed, but the colleague at the next desk left, and the workload for those who remained became increasingly heavy. Yu Pi started browsing recruitment apps, and as he scrolled, the term FDE kept appearing repeatedly.
After looking into it further, Yu Pi decided to submit his resume. When he went to Beijing for the interview, the interviewer asked him how he understood "production-level large model systems." He explained the retrieval-augmented generation processes he had built and the clients he had interfaced with, passed the interview, and started working as an FDE at a top-tier tech giant.
Among Yu Pi's former colleagues who left, some were doing algorithm work at major tech companies, and others were building Agents at startups. When they heard he was going to be an FDE, their reactions were divided into two types: some said, "Isn't this just high-end outsourcing?" while others said, "I heard this position is incredibly hot; you're so lucky."
To understand FDE, one must first return to its birthplace. The term was first coined by Palantir in the mid-2000s. At that time, the data environments for its government and defense clients were extremely sensitive, and the architectures were highly customized, making remote delivery completely unfeasible. An architect who flew away right after delivering a solution would leave the client with nothing but a pile of pending tickets, rather than a running system. Therefore, Palantir simply embedded its engineers on-site with clients long-term, writing code while understanding the business. Thus, the FDE was born.
Unlike traditional AI engineers or ML (Machine Learning) engineers, FDEs do not write core model code. Instead, they embed model capabilities into clients' business systems, bridging the "last mile" from model to value. Simply put, AI engineers are responsible for model metrics, while FDEs are responsible for business outcomes.
Technology, business, and trust—all three are indispensable, and precisely for this reason, compensation has been pushed higher and higher. LinkedIn data shows that the median annual salary for FDEs is about USD 199,000, while the total compensation packages for mid-level FDEs at frontier labs like OpenAI and Anthropic generally fall in the USD 350,000 to 550,000 range. It is not uncommon for senior or even Principal-level FDEs to break the USD 1 million mark. Translated to the domestic context, top-tier tech giants like ByteDance and Alibaba frequently offer FDEs annual salaries of CNY 1 million. Yu Pi is one of those swept up in this wave of high-salary recruitment.
So, why did the FDE role suddenly explode in popularity? What are the undeniable driving forces behind its breakout?
The "Last Mile" of AI Implementation
The fundamental reason for the explosive popularity of FDE lies in a simple judgment: model capabilities are rapidly being leveled, but integrating models into legacy systems, passing security and compliance checks, and getting frontline employees willing to adopt them—these three things cannot be leveled in the short term.
In May 2026, three of the world's top AI companies almost simultaneously made the same move. Anthropic partnered with institutions like BlackRock and Goldman Sachs to launch a joint venture with a total committed capital of USD 1.5 billion, with the core business of helping enterprises deploy Claude. OpenAI established an independent deployment subsidiary, DeployCo, with an initial investment of over USD 4 billion. Less than two weeks later, the CEO of Google Cloud publicly posted about massively recruiting FDEs, opening over 1,500 internal AI implementation-related positions. The judgment of the three companies was highly consistent: just selling APIs is no longer enough. After getting the models, clients often get stuck at "don't know how to use it," "data can't be connected," "unstable performance," and "afraid to use it after going live." They need a role to go on-site and actually get things done.
Liepin's report shows that FDE positions are highly concentrated in four cities: Shanghai, Beijing, Shenzhen, and Hangzhou, accounting for a combined 68.33%. In terms of industry distribution, AI/Internet remains the largest source of demand, accounting for 72.36%.
Newly entering the FDE workforce, Yu Pi naively thought that interfacing with clients might only take up a small portion of his time, and that most of the time he would still need to be in an office building, constantly attending meetings, writing code, and aligning on solutions.
But Yu Pi soon discovered that among the clients he interfaced with, traditional industries such as retail and manufacturing had gradually shifted from the long tail to become the main bulk. FDE is gradually expanding from an internal buzzword in the AI circle to the entire industry and society.
And this process will not be easy. The first client he took on was a manufacturing factory. The client wanted AI predictive maintenance, but the factory hadn't even collected basic data like temperature, vibration, and running time. Yu Pi could only settle for a secondary option: first streamline the repair request process, let AI help workers organize repair information, and simultaneously save each fault description and repair result in a structured way to accumulate data for the future.
It is not just about AI capabilities; the FDE faces issues regarding organizational capability, process maturity, and data infrastructure. What an FDE can do is expose these problems one by one, and then use the minimum viable solution to take the first step forward.
After three months on-site, Yu Pi slowly figured out what the FDE role is actually selling.
Selling outsourcing-style project customization is the most common and exhausting type. Clients pay money for you to enter and deliver customized solutions, earning project fees. They throw the task of "figuring out the problem" to you. Many clients haven't even figured out what they want themselves; as they chat, the requirements just disappear. Over-customization also makes it difficult to precipitate into a product, and the ceiling is visible to the naked eye. The first project Yu Pi took on was exactly this kind. The client initially said they wanted "AI predictive maintenance," but by the third week of discussions, he discovered that the real problem was that the quality inspection standards themselves were not unified.
The second type is selling product implementation, which involves having a mature Agent product and sending people to help clients connect data and systems. The name is FDE, but the essence is implementation. Yu Pi later transferred to this track, and only then did he actually get to use the things written in the recruitment JD.
The third type is selling cognitive transformation, which involves doing consulting and training, avoiding heavy delivery, and helping bosses and employees understand AI and transform their working methods. A peer Yu Pi knows runs a one-person company doing AI infrastructure consulting for traditional industries, with projects fully booked for three months, and is already preparing to expand recruitment.
There are also those who don't do delivery but act as matchmakers, organizing communities and hackathons to connect people who want to transform with enterprises in urgent need of AI integration.
Yu Pi found that, from this perspective, an FDE is more like a translator, translating AI capabilities into solutions that enterprises can understand; it is also a sensor, transmitting the real pain points of enterprises back to the product team; and it is also a catalyst, proving through the results of AI efficiency improvements in certain links that this endeavor is worth continuing.
The Price Behind the Explosive Popularity
Any position that explodes in popularity is inevitably subject to hype and exaggeration, and FDE is no exception.
Narratives of million-yuan annual salaries circulate online, but the real experience of frontline FDE workers is far less glamorous than the recruitment copy suggests.
Some overseas employees even complained: "FDE is simply the worst position I have encountered since becoming an engineer."
From Yu Pi's personal experience, being on-site is the norm and constant moving is the price. In the early stages, you are the sales, delivery, and after-sales person all by yourself. At the busiest times, it's nearly 50 hours a week, with three days a week on-site or traveling. In the morning, you go to the client site to attend meetings for a few hours to confirm requirements, and spend the rest of the time writing code. When the schedule is tight, you write code continuously for 7 to 8 hours, even leaving super-long tasks for the AI Agent before bed to execute overnight. A former FDE at Palantir described working on-site 3 to 4 days a week, staying at one client for about a year, with a team of 4 to 5 people.
But exhaustion is not the hardest part. The hardest part is that you never know what role you will have to play tomorrow.
An FDE at a domestic AI company frankly admitted that there is almost no fixed office location, weekend emergency support is the norm, and long-term travel has caused irregular schedules for many. They must write high-quality code and debug like R&D, interface with requirements and present solutions like pre-sales, and handle on-site faults and coordinate multiple teams to advance projects like operations and maintenance. Playing multiple roles alone, the error tolerance rate is extremely low. Customer satisfaction is directly linked to performance, and on-site issues must be resolved immediately.
In Yu Pi's third week at the factory, the client's boss pulled him aside and asked, "Can your AI help me predict which machine will break down tomorrow?" Yu Pi said they needed data first. The boss replied, "Haven't we been recording the data all along?" Yu Pi followed the workers to take a look, and the so-called "records" were handwritten in an oil-stained ledger, with messy handwriting and some incorrect dates. He spent a considerable amount of time typing the ledger into Excel, and then spent another week persuading the workshop director to have the workers fill in two extra columns every day.
Ironically, these preliminary tasks that have absolutely nothing to do with AI directly determine whether the AI work can even begin.
Furthermore, not all company executives who want to use AI to improve quality and efficiency understand the most basic principles of AI. Therefore, frontline FDEs, who juggle multiple roles, also need to practice how to explain businesses like Agents and skills in plain, easy-to-understand language to traditional enterprise bosses. After all, the final closing rate also relies on the FDE employees.
This explains why burnout has become a high-frequency word in the FDE circle. Frequent business trips, compound pressure, and blurred professional boundaries are common pain points for almost all frontline practitioners.
However, even when juggling multiple roles and rushing day and night, the stories on social media about "earning CNY 100,000 a month by taking orders alone" or "switching careers with zero foundation and earning an annual salary of CNY 1 million" do not apply to everyone. The FDEs who can truly earn an annual salary of CNY 1 million are concentrated in top AI companies and premier consulting firms, and usually require over 5 years of engineering experience and client-facing work experience like Yu Pi, and the position is not stable. Therefore, FDE has another nickname: "on-site outsourcing."
Anyone who has done B2B business in China understands a pain point: clients want customization, privatization, and on-site presence. In the end, SaaS turns into projects, and product companies turn into outsourcing companies. Over the past decade, many B2B startups have failed and sunk into the sand on this path.
The reason the FDE narrative in Silicon Valley is so valuable is that two premises hold true: first, there is a reusable platform behind it; second, frontline discoveries can feed back into the product, making the deployment for the next client faster and cheaper.
Customized delivery means that labor costs grow linearly with the number of clients, which contradicts the logic of diminishing marginal costs pursued by the software industry. Palantir can sustain this model because its average customer unit price is high enough, and its clients are concentrated in government and large enterprises. But when the FDE model penetrates downward to small and medium-sized clients who need AI, whether the profit per client can cover the on-site costs remains an unverified question.
Many FDE workers complain that some clients actually just have a "freeloader" mindset. After spending a small amount of money to get AI knowledge, they cancel the contract and build their own workflows. But in an era where AI has not yet fully penetrated, pricing too high in the early stages would result in losing 90% of the client base.
However, it is too simplistic to completely dismiss FDE as just a rebranding. When a technology moves from the lab to the production environment, a role specifically responsible for "making it run" will always be born. The cloud computing era gave birth to SRE (Site Reliability Engineering) and DevOps (Development and Operations), and the large model era is giving birth to FDE. SRE was once considered a variant of operations and maintenance, but later became an independent technical branch. FDE might be walking the same path.
The popularity of FDE has a solid foundation of demand, but it is not a golden ticket suitable for everyone.
On the positive side, the growth speed provided by this position is scarce. An engineer who has soaked on-site with clients for three months might have an understanding of AI implementation that equals three years of iteration by the back-end team. But on the other hand, the boundaries of this position are being diluted. Some analyses point out that with the surge in demand, more and more positions with the FDE title are seeing their actual work content slide towards pre-sales, solution architecture, and even customer success.
AI moving from the lab into business processes indeed requires a group of people who understand both engineering and business, and are also willing to dive into the field. This part of the demand will not disappear as the bubble recedes. But within the 42-fold growth figure, there is indeed a mix of title inflation, role merging, and unverified value expectations.
Yu Pi, who has broken into this field, doesn't know whether he has stepped on the tailwind or walked through a narrowly packaged door. All he knows is that at 8:00 a.m. tomorrow, he still needs to be on time at the workshop.