The most sought-after new job position has emerged in Silicon Valley.
In the past three years, the most expensive people in the AI industry were model scientists.
Today, the people that OpenAI, Anthropic, and Google most want to recruit have changed.
They are not researchers, not algorithm engineers, and not even large - model experts.
Instead, they are a group of people who need to travel on business, stay on - site, attend meetings, and change processes.
They have a new name: Forward Deployment Engineer (FDE for short), Front - line Deployment Engineer.
This is a seemingly unremarkable position, but it may represent the biggest shift in the AI industry in the past three years: The myth of models is officially over, and the battle for implementation has fully begun.
The large - model giants in Silicon Valley have finally discovered that the model is no longer the problem. The most difficult part is that enterprises don't know how to use it. As a result, a position that was previously ignored has skyrocketed in value overnight.
The starting base salary for an entry - level FDE at Anthropic is $170,000 - $200,000, and the total package is $300,000 - $500,000.
The base salary for an FDE position at OpenAI starts at $210,000, and both types of positions come with four - year equity.
For the FDE positions at ByteDance's Doubao/Flyte in China, the monthly salary is 35,000 - 70,000 yuan with a 15 - month pay scale, and the maximum annual salary is 1.05 million yuan.
Some headhunters in the market have offered a special offer of a $400,000 annual salary and a fully remote position for a senior FDE with two years of implementation experience.
The LinkedIn 2026 Labor Report shows that from 2023 to 2025, the number of global FDE recruitment positions increased by 42 times, while the number of AI engineer positions increased by 13 times during the same period. The growth rate of the former is about three times that of the latter.
This unconventional frenzy of talent - hunting has torn off the most tacit veil in the entire AI industry in the past three years.
I. The model is ready, but the organization lags behind
Since the birth of ChatGPT, the main line in the AI industry has been clear. It has extended from who can build a stronger model to who can build the best Agent.
By 2026, the question has changed. Corporate customers are starting to ask another question: We've bought AI, but why hasn't there been much change?
This is the biggest illusion in the entire industry, thinking that the model equals productivity.
In reality, many enterprises spent a lot of money to purchase AI/Agents. Employees registered accounts, and the IT department created a demo of an internal knowledge base. They were excited for a month.
Then... half a year passed, and no one used it. The way of working remained exactly the same as before.
It's not that employees are uncooperative, nor that management lacks determination, and it's not that the model is not good enough. The real bottleneck for enterprises in the production environment has never been how to chat, but where the historical data is, whether the format is correct, and what the quality is; which approval process to follow and who has the leading power; how to import customer data, how to integrate with the ERP system, and how to be compatible with the existing compliance and security systems?
These are not technical problems, but organizational problems.
It's like installing a rocket engine on a carriage. The engine is real, and the thrust is real, but the horse is still a horse, the track is still a dirt road, and the driver has never learned how to step on the accelerator, let alone where the emergency brake is.
Model companies have always sold their products as tools, giving users the most powerful digital brain and letting users figure out how to integrate it into their operations.
As a result, most enterprises have tried for two years, but the "brain" is still sitting on the table, and the operations remain unchanged.
II. The legacy of Palantir
It was not OpenAI that truly established FDE as a profession, but Palantir Technologies.
This mysterious big - data unicorn, founded by Silicon Valley godfather Peter Thiel and once helped the US military kill Osama bin Laden, was ridiculed in Silicon Valley for fifteen years.
The reason is that its business model is too heavy. It doesn't sell standardized software but sends engineers to the customer's site for on - site support, sometimes staying for half a year or more. VCs labeled it as a consulting firm in the guise of a software company.
In the Silicon Valley hierarchy, SaaS is considered high - end, while projects relying on manpower are considered low - end. Palantir is at the bottom of this hierarchy.
In 2011, when Palantir was selling data software to government and defense agencies, it discovered a recurring problem: customers didn't know how to use the software after purchasing it.
But this problem changed everything. The traditional model of sales collecting requirements and engineers developing remotely completely failed in the face of highly confidential and extremely complex customers. Customers themselves didn't know what they wanted; they only knew that the existing products were not useful.
Palantir's approach was not to provide better instructions but to directly send its engineers to the customer's site. They entered the CIA, energy companies, and banks. Engineers sat beside customers, observed how they worked, studied data processes, understood organizational structures, and then modified software, processes, and even work methods.
This model was never replicated on a large scale in the era of standardized software. In the past, products defined processes, and if customers were not satisfied, it was considered a lack of training.
The era of large models has completely broken this logic. There is no standard way to use AI, and its ceiling depends entirely on how to access private data, design workflows, and implement them within the organization. Each enterprise's chimney systems are completely different, and general products cannot solve the deep - seated customization problems.
So, the methodology that Palantir has developed over more than a decade has suddenly become a textbook for the entire industry.
Today, OpenAI is starting to replicate this model, which essentially means admitting that AI has changed from a software development problem to an organizational evolution problem.
III. In one month, three giants, the same judgment
If Palantir just set an example for the industry, then in May 2026, the three top giants in the global AI field simultaneously made a collective move towards application implementation with real money.
On May 4th, Anthropic, in collaboration with Blackstone, Goldman Sachs, Hellman & Friedman, and several global asset management institutions, launched a joint venture with a total committed capital of $1.5 billion. Its core business is to deploy the Claude large model for enterprises.
Immediately afterwards, on May 11th, OpenAI officially announced the establishment of an independent deployment subsidiary, Deployment Company (DeployCo). The total initial investment in the cooperation exceeded $4 billion, and the cooperation camp included 19 institutions in total, including private equity investors such as TPG and Bain Capital, as well as consulting integrators such as McKinsey and Accenture.
OpenAI also acquired the on - site AI consulting firm Tomoro. After the acquisition, Tomoro will provide about 150 front - line deployment engineers to DeployCo. Tomoro's existing customers include Tesco, Virgin Atlantic, Red Bull, and Supercell.
Less than two weeks later, Google Cloud CEO Thomas Kurian publicly posted on LinkedIn to recruit FDEs on a large scale. Google Cloud has opened more than 1,500 AI implementation - related positions internally, with FDE being the core recruitment category.
The three top AI companies in the world did the same thing at the same time, not releasing a stronger model, but establishing entities specifically to help enterprises implement AI.
This is a signal that deserves more attention than any model release.
OpenAI COO Brad Lightcap even said the following:
Today, the capabilities of AI systems for individuals are already very powerful, but we haven't really seen AI penetrate into enterprise business processes. Enterprises are complex organizations with fragmented systems, many compliance constraints, and complicated legacy processes. The biggest challenge currently is to integrate AI into the core business processes that enterprises rely on to operate.
In short, the model is good enough. The problem lies within the company and the organization.
It is precisely because they understand this that OpenAI and others are willing to pay any price to hire the disciples of Accenture and McKinsey and upgrade them in batches to front - line FDEs.
This talent - hunting battle worth billions of dollars has directly taken away the underlying assets of the traditional consulting and IT implementation industries and has also kicked off a revolution in the large - model delivery model.
IV. The end of selling tools is selling results
Many people thought that AI would eliminate the consulting industry. McKinsey would be finished, Accenture would be finished, and large - scale IT implementers would be finished.
On the contrary, AI has revitalized the consulting industry.
But there is a deeper change behind this. The business model of the entire software industry is undergoing the biggest transformation in the past twenty years.
License era: Customers buy software and solve problems on their own.
Subscription era (SaaS): Customers buy services and solve problems on their own.
AI era: Customers buy results, and suppliers are responsible for solving problems.
This is exactly the survival rule that Palantir developed more than a decade ago: Don’t sell software. Deploy outcomes. (Don't sell software, sell results).
This is an essential transformation. In the past, Microsoft sold Office, Salesforce sold CRM, and Adobe sold software suites. They all delivered tools, and it was up to you to use them well. What OpenAI and Anthropic are doing today is to send their people into the customer's company to deliver results.
FDEs are result deliverers. They study organizations, processes, and data, and finally output a system that truly runs in the production environment, rather than a beautiful demo.
In the past, consultants outputted PPTs, and FDEs output Agents. In the past, consultants gave advice, and FDEs give code. The essence is the same: to help enterprises solve the problem of how to work more efficiently, only the deliverables have changed.
This is also why there is a strange requirement in Anthropic's FDE recruitment: Maintain a low sense of self and a collaborative attitude.
This is the most difficult requirement in the engineer culture. FDEs need to have sufficient technical depth to solve any problem on - site, and at the same time, they need to put aside the attitude of knowing more than the customer and patiently understand why the customer doesn't trust the AI output.
The annual salary of $300,000 - $500,000 is not because FDEs have stronger technology, but because a qualified FDE can replace a product manager, a technical architect, a project manager, and an AI engineer.
On the front line of delivery, an FDE is an army.
V. The biggest obstacle to AI implementation has never been technology
Most of the failures of enterprise AI projects today are not due to technical failures but organizational failures.
Even the world's top financial empires and retail giants are not immune to this.
Goldman Sachs encountered a classic middle - level compliance defense when promoting AI migration. The technology department developed an AI audit system that could automatically generate analyst reports and conduct preliminary reviews of IPO compliance documents.
But when the system was ready to be integrated into the production environment, the middle - level executives in the risk control and compliance departments jointly pressed the pause button. They submitted a thick query report to the management, asking who would be responsible for the potential billions of dollars in fines if the "hallucinations" of the large model appeared in the listing documents?
No matter how beautiful the technical prototype was, due to the deep - seated culture of avoiding responsibility within the organization, the project was stuck for half a year. It was not until the FDE team intervened to re - define the power and responsibility boundaries of human - machine collaboration that the project barely passed.
If Goldman Sachs was stuck on power and responsibility, then the famous failure of the early cooperation between US retail giant Target and Palantir hit the wall of organizational interests and culture.
At that time, Palantir sent a large FDE team to Target, trying to use data models to reconstruct its supply chain and inventory forecast with annual revenues of tens of billions of dollars.
However, the most powerful senior buyer team within Target was extremely resistant. They believed that their decades of fashion acumen should not bow to an algorithm. Middle - level managers deliberately delayed the data interface, and front - line employees deliberately did not execute the system's replenishment instructions. This multi - million - dollar technological transformation ended in a disastrous way when Target unilaterally tore up the contract due to the power struggle between humans and machines within the organization.
The code was correct, but the project couldn't move forward. This is the most real implementation scenario. Technology only accounts for 20%, and the remaining 80% is all about the internal interest structure, power and responsibility allocation, and historical burdens of the organization.
For example, a bank's loan approval process is based on decades of power and responsibility allocation and regulatory requirements. A hospital's scheduling system is related to the interest structure of all departments. A factory's quality inspection process is connected to supplier contracts and quality insurance.
These will not automatically change because of a GPT account.
These obstacles cannot be solved by an engineer who only understands technology. What is needed is someone who can think from both the technical and organizational dimensions.
So, what FDEs really do is not just deploy AI. The core is to help organizations complete AI migration. If the IT department was responsible for digitizing paper processes in the past twenty years, then in the next ten years, FDEs will be responsible for AI - enabling the digitized processes.
This is the next stage of the same thing.
[Beyond the Page] Words:
As models become cheaper, computing power becomes cheaper, and Agents become cheaper.
The truly expensive thing has become another ability: Understanding organizations, transforming processes, and driving change.
This is why FDEs have become popular.
It's not that this position is so important. In essence, the entire AI industry has finally admitted one thing:
The most difficult part of a technological revolution has never been technology.
It's people.
This article is from the WeChat official account “Beyond the Page”, author: Huahua. It is published by 36Kr with permission.