The Virtuality and Reality of FDE

In 2026, FDE has become one of the most high-profile keywords in the AI industry. Entering the third quarter, the activity and discussion level on the recruitment side continue to rise, the training market is heating up simultaneously, and more participants are joining the track.

Statistics from US recruitment platform Indeed show that the number of FDE-related positions increased from 643 in April 2025 to 5330 in April 2026, a nearly 7.3-fold increase in one year.

The 2026 Global Labor Market Trend Insight Report released by LinkedIn in January 2026 shows that the number of new FDE positions increased 42 times from 2023 to 2025.

The domestic market is equally hot: ByteDance offers the position of "Doubao AI Large Model FDE" with a monthly salary of 35,000 to 70,000 RMB and 15 months' salary per year; Ant Digital Technology offers the B-end FDE position with a monthly salary of 40,000 to 60,000 RMB. Alibaba, Tencent, major cloud vendors and AI companies are all making layouts in this field. From the second to the third quarter of 2026, large tech giants have concentrated their FDE recruitment expansion.

FDE service providers are also expanding at the same time. Serialized "FDE courses" have appeared on Xiaohongshu, Zhihu has released the 2026 China FDE Talent White Paper, and various training camps and certification programs have spread from first-tier cities to county-level regions.

Shanghai has taken the lead in incorporating FDE talent training into the government plan and implemented the "Hundreds, Thousands, Tens of Thousands" program, while the human resources and social security department is studying the feasibility of linking FDE training with intermediate professional titles. It is worth noting that large manufacturers in the traditional information security field are also accelerating the construction of their own FDE communities, with the core goal of supporting the delivery and implementation of their own agent products.

I. Disputes over the Definition of the FDE Concept

FDE, short for Forward Deployed Engineer. This position was first scaled up by Palantir, and its core mode is to station engineers at customer sites to embed standardized data products into enterprises' existing systems and workflows.

In the AI era, the functions of FDE have changed. It is no longer just "deploying software", but going deep into customer sites to connect large models to business processes. FDE practitioners need to write code and build systems, communicate scenarios with business departments, and discuss ROI with management teams.

No matter how powerful the large model is, there is still a large gap between its implementation in specific industries and scenarios. Enterprises need personnel to convert AI capabilities into business solutions, and then implement the solutions into operable systems.

Controversies have always existed: Is this a new profession spawned by the AI era, or a conceptual hype of "old wine in a new bottle"? "A large part of FDE's popularity in 2026 comes from conceptual arbitrage." The judgment of Zaniel, a PhD student at Stanford University and independent consultant, points out the core controversy in the industry.

The essence of FDE's popularity reflects a deeper issue: what kind of capability structure and talent profile do enterprises need to support the implementation of AI? From practical experience, neither single technical capability nor single business capability is sufficient, and composite capability is the answer. However, the market lacks consensus on this, leading to the emergence of chaos.

II. Concept Expansion and Data Bubble

According to industry observations, a considerable number of domestic recruitment positions labeled as FDE are just renamed traditional technical implementation posts. Implementation engineers, pre-sales technical engineers, operation and maintenance engineers — enterprises have not added new positions, but only renamed the original positions as FDE. The work content remains unchanged, and the salary and evaluation system still follow the original framework. In this context, the so-called "FDE" is just a new job title.

Although the position name has been changed, the work content has not changed substantially; while the popularity is rising, capability building has not kept pace simultaneously.

This is not the expansion of the FDE market, but the expansion of the concept of FDE. A real FDE requires capabilities in large model invocation, business understanding, solution design, and on-site delivery — none of which can be achieved by simply changing the job title.

The rise of a new concept often goes through the process of "naming — packaging — dilution". FDE is currently in the transition stage from packaging to dilution: the renaming of traditional posts, crash courses of training institutions, and the wild growth of solution service providers are all accelerating the dilution of the essence of this concept.

III. Structural Mismatch of Service Supply

FDE service providers are entering the market even faster than enterprise recruitment. Some institutions even use the JD of enterprise recruitment to reverse-sell solutions. If they do not truly understand customer needs, they are not serving customers, but disturbing them.

The original logic of the FDE market is clear: enterprises have AI implementation needs and require FDE talents, and service providers deliver such talents to enterprises. This is a B-end service chain: enterprises pay, institutions deliver talents, and talents serve enterprises.

However, the reality has deviated from this track. A large number of service providers have shifted their targets to individual users — enrolling students, offering courses, charging fees and issuing certificates, and the students will find jobs on their own after graduation. Whether the content they learn is recognized by enterprises and whether the students are qualified for the positions are not within the scope of the institutions' concerns. The individual FDEs trained are only paying users, not the final customers — because it is enterprises that actually pay for FDE.

The FDE market is essentially a vertical track in the training market, and its business logic is not complicated: it can operate as long as large-scale supply connects with large-scale demand, but there are problems on both the supply and demand sides at present.

On the supply side, service providers are pouring in sharply, but the quality standards of their outputs are uneven. Most institutions only have the ability to offer courses, but cannot grant authoritative professional qualification certificates.

Where do real enterprise projects come from? How to build a practical training system? How to evaluate and track students' abilities? Most emerging FDE service providers are unable to solve these key problems. Most of them rely on external forces to obtain customers, start operations hastily without professional personnel leading the construction of the practical training system, cannot prove their own evaluation standards, lack professional, scientific and efficient human resource industry management and operation modes, and have no professional team to support key account management.

Thus a paradox arises: the more prosperous FDE training is, the more "semi-finished" FDEs there will be. This group of students has learned the concepts and done exercises, but they have not been exposed to real enterprise needs, do not understand the complexity of business scenarios, and do not have independent delivery capabilities. When they pour into the market, they will drive down prices and blur standards — bad money drives out good money.

Taking open public courses as the only entry channel, lacking systematic and global thinking, and blindly competing with peers on prices will only lead to vicious competition. Without following up business leads after the courses end and no conversion, the revenue will be meager.

The trained FDEs form their own teams or flow to peers — this is creating competitors for themselves, which is not only digging traps for their own business, but also squeezing the demand market that has not yet been fully developed.

On the demand side, enterprises' real needs exist and are urgent. Traditional industries such as manufacturing, retail, energy and logistics — they have seen the AI trend and realized that they must keep up, but do not know where to start.

Such enterprises almost follow the principle of "whatever you say is right" for FDE — which indicates that the market is still in the initial stage driven by cognitive gap. The problem is that these enterprises are scattered and not connected with each other, as if they are in their own information cocoons, lacking awareness of FDE service standards and prices. It is difficult for service providers to reach them, and it is also difficult for them to find reliable service providers. On the way to mutual cooperation, there are full of obstacles, challenges and uncertainties.

Large leading manufacturers have clear FDE needs and sufficient budgets, but they tend to build their own teams or cooperate for a long time with organized and systematic institutions (such as technical consulting companies and software outsourcers). Enterprise customers trust such formal service providers more than individual FDEs who claim to be OPC (One Person Company) — the latter's delivery capability and stability are difficult to verify.

The contradiction of the whole market lies in: service providers are expanding student enrollment, and enterprises are expanding talent recruitment, but the two sides cannot connect effectively all the time.

Behind this contradiction is a cognitive deviation: the market positions FDE as a standardized technical type of work, while what enterprises need are highly personalized compound talents. Cognitive mismatch is the deep reason why supply and demand cannot match.

IV. Decision-Making Logic of Foreign Enterprises

If domestic enterprises' attitude towards FDE is "enthusiastic", foreign enterprises show a different rhythm. Although the FDE model has been verified and successfully applied by leading foreign tech companies, foreign-funded enterprises in China in traditional industries present a different pace.

The FDE model was originally pioneered and verified on a large scale by Palantir, a US big data analysis company. In April 2023, Palantir launched the AIP platform, and then launched the FDE-led "bootcamp" — stationing engineers at customer sites to complete the whole process from data modeling to LLM application delivery within 3 to 5 days.

By the end of 2024, Palantir had held more than 1000 AIP bootcamps, with a high proportion converted into million-dollar contracts. In May 2026, OpenAI went a step further and announced the establishment of OpenAI Deployment Company (internal code name DeployCo) with an initial investment of over 4 billion US dollars, and acquired British AI consulting firm Tomoro (with about 150 FDEs and deployment experts), directly stationing engineers inside customer organizations to restructure workflows.

In the same year, Anthropic jointly established an AI-native enterprise service joint venture with Blackstone, Goldman Sachs and other institutions, investing 1.5 billion US dollars, and also adopted the FDE embedding model.

However, these leading cases do not represent the overall rhythm of foreign enterprises. Especially for foreign-funded enterprises in China in traditional industries, they do not deny the value of FDE, but their cost accounting is stricter and their decision-making cycle is longer.

These foreign enterprises have a characteristic in technology adoption: one-step behind. They do not chase the latest technologies, but only the most stable ones. After new concepts emerge, they will observe and study, but will not follow up immediately. They will decide whether to enter the market after the market completes a round of verification and the problems are fully exposed. This is not conservatism, but the logic of cost control.

FDE also follows this logic. The core concern of foreign enterprises is: can this position reduce costs? If yes, they will further ask: can efficiency be improved under the condition of unchanged existing resources? If these two questions cannot be answered, the budget will not be approved.

Operation functions — supply chain, financial sharing, customer service, human resources — have long used mature SaaS tools with stable processes, familiar personnel and predictable costs. Introducing FDE means increasing labor costs and adjusting processes, which is most likely to be rejected in financial calculation.

More critically, these SaaS products themselves are iterating. Salesforce, SAP, Oracle and Workday are embedding AI capabilities into their products. Foreign enterprises generally adopt Workday because of its mature system, stable architecture and high reliability. Functions that previously required manual configuration can now be completed automatically by AI; links that previously required on-site FDE debugging can now be upgraded online directly by global SaaS vendors.

In the human resources scenario, Workday has embedded AI capabilities in salary calculation, organizational structure adjustment, talent inventory and other links. The same is true in supply chain, finance, customer service and other fields — the stronger the SaaS capability, the smaller the space for FDE "deployment".

Another structural constraint: most of the IT infrastructure of foreign enterprises is deployed on global servers, and local enterprise-level deployment requires authorization from the headquarters. Even if the China region recognizes the value of FDE, its implementation still requires decision-making and authorization from the global headquarters. If FDE is not assigned by the global headquarters, the local team usually cannot promote it independently. This is not only a matter of willingness, but also a matter of authority.

However, foreign enterprises have not completely closed the door to FDE. Their way of entering the market is different from the expectation of domestic service providers. Foreign enterprises tend to be in a state of "human-machine collaborative symbiosis" on the operation side.

They are not in a hurry to hand over all positions to AI, nor to introduce a large number of FDEs to transform processes, but gradually explore the embedding links of AI on the basis of existing systems and personnel. This kind of exploration often starts with "do-it-yourself": business department employees build agents and develop lightweight applications to solve specific problems. Such exploration will not form a boom, but will not fade away quickly. It is a slow but continuous curve.

For FDE service providers, the foreign enterprise market is not inaccessible, but the "batch training and batch delivery" mode cannot be adopted.

V. Non-Standard Nature of FDE Capability

A good FDE is not a jack of all trades. They cannot be quickly trained only through courses, but need to grow up in specific businesses — with sufficient and profound industry accumulation, solid professional skills, mature professional cognition and literacy, and practical experience tempered by projects, so as to truly serve enterprises.

Even for suppliers who have already run through the FDE service model, they are facing common challenges: the industry and professional background of FDEs are the foundation to cope with different demands. The difference in customer demands is not a problem in itself — if FDEs have profound industry understanding and professional domain knowledge, the demand difference can be digested. However, the current standards for selecting FDE trainees in the market are mostly focused on technical developers, and few trainees are trained from the business side.

This is a lack of market cognition. Just like the origin of project managers, BAs and product managers in the past — whether to develop business after accumulating technical background or to supplement technology after accumulating business background? Facts show that the sources of talents can be diversified. After supplementary training, compound talents are the final profile.

The essence of technology is a tool, and the essence of business is a problem. The core value of FDE does not lie in how many tools they master, but in their ability to quickly understand, disassemble and solve complex problems.

A qualified FDE's adaptability and judgment cannot be standardized — a good consultant knows what method to use in what scenario, and can distinguish real problems from false propositions. This kind of ability cannot be cultivated through unified textbooks and standardized exams.

The most difficult problem for FDE service providers is not technical, but to make students understand the actual business of customers. Customers cannot clearly state their demands, and students cannot ask for the real demands, which often leads to deviation of the project promotion direction.

FDE is not a position that "you can do it as soon as you learn it", and it requires experience accumulation in real scenarios. If the FDE market continues to operate with low standards, it will damage its own reputation sooner or later. Enterprises will only pay for deliveries that truly solve problems, not for glamorous job titles.

Perhaps we can look at it from another perspective: the heads of HR operation team and HR agile team of a long-established foreign manufacturing enterprise both have technical backgrounds, they require HR teams with engineer thinking and cultivate HR teams with product manager capabilities.

Such an HR team can be transformed into FDEs through lightweight FDE training — becoming high-equipped FDEs based on business and functions, meeting the internal needs of enterprises and supporting business growth. This reveals another path for FDE training: growing from inside the enterprise, rather than being delivered in batches from outside.

VI. Another Path for Physical AI

The implementation of FDE cannot be separated from the support of underlying technologies, and the evolution of underlying technology paradigms is redefining the cognitive boundary that FDE needs to possess.

The speech of Richard Sutton, 2024 Turing Award winner and father of reinforcement learning, at the HICOOL 2026 Global Entrepreneurs Summit provides another dimension of thinking. His core judgment is different from the mainstream large model paradigm: human-labeled data cannot cultivate real intelligence, and machines must achieve autonomous growth through real trial and error in the physical world.