This article comes from the WeChat public account: 版面之外, author: 画画

Over the past three years, the most expensive people in the AI industry were model scientists.

Today, the people OpenAI, Anthropic, and Google most want to hire have changed.

Not researchers, not algorithm engineers, not even large language model experts.

But a group of people who have to travel, work on-site, attend meetings, and overhaul processes.

They have a new name: Forward Deployment Engineer (FDE for short).

This is a seemingly unremarkable role, yet it may represent the biggest shift in the AI industry over the past three years: the model myth has officially exited the stage, and the war over real-world deployment has fully begun.

Silicon Valley's large model giants have finally realized that models are no longer the problem. Enterprises not knowing how to use them is the hardest last mile. As a result, a role that once attracted no attention has seen its value skyrocket overnight.

  • Anthropic's entry-level FDE base salary is $170,000–$200,000, with total packages of $300,000–$500,000;
  • OpenAI FDE roles start at a base salary of $210,000, with both types of positions including four-year equity.
  • Recruiters in the market have already offered exceptional packages of $400,000 annual salary, fully remote, to senior FDEs with two years of deployment experience.

LinkedIn's 2026 Workforce Report shows that from 2023 to 2025, global FDE job postings grew 42-fold over two years, while AI engineer postings grew 13-fold over the same period—the former growing about three times as fast as the latter.

This unconventional talent-grabbing frenzy has torn off the entire AI industry's most unspoken fig leaf of the past three years.

I. The Models Delivered, the Organizations Didn't

Since the birth of ChatGPT, the AI industry's main storyline has been clear: from who can build stronger models, to who can build the best AI agents.

By 2026, the question has changed. Enterprise customers have started asking a different question: We bought AI—why has so little changed?

This is the industry's biggest illusion: believing that models equal productivity.

The reality is that many enterprises have spent big money procuring AI/agents, employees have registered accounts, the IT department has built a demo of an internal knowledge base, and everyone was excited for a month.

Then... six months passed, and nobody uses it. Work happens exactly the way it did before.

It's not that employees won't cooperate, nor that management lacks resolve, and certainly not that the models aren't good enough. The real choke point for enterprises in production has never been how to chat, but rather: where is the historical data, is the format right, what's the quality? Which approval chains and responsibilities apply, who has the lead? How do you import customer data, integrate with ERP systems, and make it compatible with existing compliance and security frameworks?

None of these are technical problems. They are organizational problems.

It's like bolting a rocket engine onto a horse-drawn carriage. The engine is real, the thrust is real, but the horse is still a horse, the track is still a dirt road, the driver has never learned how to press the accelerator, and doesn't even know where the emergency brake is.

Model companies have always sold their products as tools: giving users the most powerful digital brain and letting them figure out how to fit it into their bodies.

The result: after two years, most enterprises still have the brain sitting on the table, and the body hasn't moved an inch.

II. Palantir's Legacy

The company that truly turned the FDE into a profession wasn't OpenAI—it was Palantir Technologies.

This mysterious big data unicorn, founded by Silicon Valley godfather Peter Thiel and once credited with helping the US military take out Osama bin Laden, was mocked in Silicon Valley for fifteen years.

The reason: its business model was too heavy. It didn't sell standardized software; instead, it sent engineers to work on-site at customers, staying for more than half a year at a time. VCs slapped a label on it: a consulting company wearing software's skin.

In Silicon Valley's hierarchy of contempt, SaaS was high-class, while projects built on headcount were low-class. Palantir sat at the very bottom.

In 2011, while selling data software to government and defense agencies, Palantir noticed a recurring problem: customers bought the software and had no idea how to use it.

But it was this very problem that changed everything. The traditional model of sales collecting requirements and engineers developing remotely completely failed with highly classified, extremely complex customers. The customers themselves didn't know what they wanted—they only knew that what they had didn't work.

Palantir's approach wasn't to write a better manual—it was to send its own engineers directly to work on-site at customers. Into the CIA, into energy companies, into banks. Engineers sat next to customers, observed how they worked, studied data flows, understood organizational structures, and then modified the software, the processes, even the way people worked.

This model was never replicated at scale in the era of standardized software, when the product defined the process and customer dissatisfaction simply meant insufficient training.

But the large model era has completely shattered that logic. AI has no standard usage; its ceiling depends entirely on how it's connected to private data, how workflows are designed, and how it's rolled out within the organization. Every company's siloed systems are completely different, and generic products simply cannot solve the deep-water problems of customization.

So Palantir's methodology, accumulated over more than a decade, suddenly became the textbook for the entire industry.

Today, OpenAI has begun copying this model—essentially admitting that AI has shifted from a software development problem to an organizational evolution problem.

III. One Month, Three Giants, the Same Judgment

If Palantir merely set an example for the industry, then in May 2026, the three most elite giants in the global AI race simultaneously sealed a collective pact on real-world deployment with cold, hard cash.

On May 4, Anthropic, together with Blackstone, Goldman Sachs, Hellman & Friedman, and several global asset managers, launched a joint venture with $1.5 billion in committed capital, whose core business is helping enterprises deploy the Claude large model.

Then, on May 11, OpenAI officially announced the creation of an independent deployment subsidiary, the Deployment Company (DeployCo), with total initial investment exceeding $4 billion and a coalition of 19 institutions, including private equity investors such as TPG and Bain Capital, as well as consulting integrators like McKinsey and Accenture.

OpenAI simultaneously acquired Tomoro, an AI on-site consulting firm; after the acquisition, Tomoro will supply DeployCo with about 150 Forward Deployed Engineers. Tomoro's existing clients 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 at scale; Google Cloud has opened more than 1,500 AI deployment-related positions internally, with FDE as the core hiring category.

The three most elite AI companies in the world did the same thing at the same time—not releasing stronger models, but establishing entities dedicated to helping enterprises deploy AI.

This is a signal more worth watching than any model release.

OpenAI COO Brad Lightcap went so far as to say:

AI systems for individuals are already extremely capable today, but we haven't yet truly seen AI penetrate enterprise business processes. Enterprises are complex, structured organizations with fragmented systems, heavy compliance constraints, and tangled legacy processes; the biggest challenge right now is integrating AI into the core business processes that enterprises depend on to operate.

Simply put: the models are good enough. The problem lies inside companies and organizations.

It is precisely because they see this clearly that the OpenAIs of the world are willing to pay any price to acquire the disciples of Accenture and McKinsey and mass-upgrade them into FDEs on the front lines.

This multi-billion-dollar talent war has directly drained the core assets of the traditional consulting and IT implementation industries, and has kicked off a revolution in large model delivery models.

IV. The End of Selling Tools Is Selling Outcomes

Many people believed AI would destroy the consulting industry. McKinsey is finished, Accenture is finished, the big IT implementers are finished.

The result is exactly the opposite: AI has made consulting bigger than ever.

But behind this lies a deeper change: the business model of the entire software industry is undergoing its biggest shift in twenty years.

  • The license era: customers buy software and solve problems themselves.
  • The subscription era (SaaS): customers buy services and solve problems themselves.
  • The AI era: customers buy outcomes, and the vendor is responsible for solving the problems.

This is exactly the survival rule Palantir distilled more than a decade ago: Don't sell software. Deploy outcomes.

This is a fundamental transformation. In the past, Microsoft sold Office, Salesforce sold CRM, and Adobe sold suites—all delivering tools, and whether you used them well was your problem. Today, what OpenAI and Anthropic are doing is sending their own people into customer companies to deliver results.

An FDE is an outcome delivery person. They study the organization, the processes, the data, and ultimately output a system that actually runs in production—not a beautiful demo.

In the past, consultants delivered PPTs; FDEs deliver AI agents. In the past, consultants gave advice; FDEs give code. The essence is the same—helping enterprises solve the problem of working more efficiently. Only the deliverable has changed.

This is also why Anthropic's FDE job posting includes a strange requirement: keep a low ego and a collaborative attitude.

This is the hardest part of engineering culture: having enough technical depth to solve any problem on-site, while also dropping the "I know better than you" posture in front of customers and patiently understanding why they don't trust AI outputs.

An annual salary of $300,000 to $500,000—not because FDEs are technically stronger, but because one qualified FDE can replace four people: a product manager, a technical architect, a project manager, and an AI engineer.

On the delivery front line, one FDE is an army.

V. The biggest obstacle to AI adoption has never been technology

Most enterprise AI projects that fail today are not technical failures—they are organizational failures.

Even the world's top financial empires and retail giants are not immune to this.

Goldman Sachs encountered classic middle-management compliance defense when pushing forward its AI migration. At the time, the technology department had developed an AI audit system capable of automatically generating analyst reports and performing initial reviews of IPO compliance documents.

But when the system was about to be connected to the production environment, middle managers in the risk control and compliance departments jointly pressed the pause button. They submitted thick interrogation reports to management: if the large model's "hallucinations" appeared in listing documents, who would bear responsibility for potential fines of billions of dollars?

No matter how beautiful the technical prototype was, the project was stuck for a full six months because it couldn't overcome the deeply entrenched culture of liability avoidance within the organization. It only barely got through after the FDE team stepped in to redraw the boundaries of responsibility between humans and machines.

If Goldman Sachs got stuck on accountability, then the famous early failure of the American retail giant Target with Palantir ran into the wall of organizational interests and culture.

At the time, Palantir dispatched a large FDE team to Target, attempting to use data models to restructure its supply chain and inventory forecasting for a business generating tens of billions of dollars in annual revenue.

However, Target's most powerful veteran buyers were extremely resistant, believing that their decades of fashion acumen should not bow to an algorithm. Middle management stalled endlessly on data interfaces, while frontline employees deliberately refused to execute the system's restocking instructions. This technology overhaul, which cost tens of millions of dollars, ultimately ended tragically with Target unilaterally tearing up the contract, a casualty of the internal power struggle between people and machines within the organization.

Not a single line of code was wrong, yet the project simply couldn't move forward. This is the most real-world picture of implementation: technology accounts for only 20%, while the remaining 80% is entirely about internal power structures, allocation of responsibilities, and historical baggage within the organization.

For example, a bank's loan approval process is backed by decades of responsibility allocation and regulatory requirements. A hospital's scheduling system is tied to the interests of every department. A factory's quality inspection process is connected to supplier contracts and quality insurance.

None of these will automatically change because of a GPT account.

These obstacles cannot be solved by an engineer who only understands technology. What's needed is someone who can think across both the technical and organizational dimensions at the same time.

So what FDEs really do is not just deploy AI—the core is helping organizations complete their AI migration. If the past twenty years were about IT departments digitizing paper-based processes, then the next ten years will be about FDEs making digitized processes AI-powered.

This is the next stage of the same thing.

A word from 【版面之外】:

Models are getting cheaper. Compute is getting cheaper. AI agents are getting cheaper.

What's becoming truly expensive is a different kind of capability: understanding organizations, transforming processes, and driving change.

That's why FDEs are hot right now.

It's not that this role is so important—the essence is that the entire AI industry has finally admitted one thing:

The hardest part of a technological revolution has never been the technology.

It's the people.