A million-yuan salary and a 42x surge in demand: What exactly does FDE, the hottest job of the AI era, do?

By the end, you'll realize your previous understanding may need updating.

In September 2025, Bob McGrew attended an AI startup conference. He is a Palantir veteran who later went to OpenAI as Chief Research Officer. To his surprise, no one at the venue was talking about how powerful GPT was or how big its parameters were—everyone was chasing him with the same question:

"How exactly does Palantir's FDE model work?"

FDE stands for Forward Deployed Engineer, usually translated into Chinese as "前线部署工程师". Hiring demand for this role has surged 42x over the past two years (LinkedIn's January 2026 Global Labor Market Report), far outpacing the 13x growth for ordinary AI engineers. OpenAI poured $4 billion into setting up a deployment company, and Anthropic put $1.5 billion into a joint venture—all to grab these people.

ByteDance offers monthly salaries of 35K-70K RMB, Zhipu offers 60K-80K, and senior FDEs in Silicon Valley can earn total annual packages of over $500,000.

Why is this role so valuable? What does it actually do? How is it different from an ordinary software engineer?

Article Outline

• FDE's battlefield origins: from the CIA to the killing of bin Laden
• "Dirt roads" and "highways": FDE's working methodology
• Why did the AI era suddenly need FDEs?
• The salary truth: a global roundup of FDE earnings
• What kind of person can become an FDE?
• FDE vs SWE vs Sales Engineer vs consulting: what exactly are the differences?

FDE's battlefield origins: from the CIA to the killing of bin Laden

The term FDE was invented by Palantir.

In 2003, Peter Thiel, Stephen Cohen, and Alex Karp founded Palantir. The company's name comes from the "seeing stone" in The Lord of the Rings—a magic stone that lets you see distant things. In-Q-Tel, the CIA's venture arm, gave them a first investment of $2 million.

But things didn't go smoothly at first. Founder Stephen Cohen brought a demo to the CIA, and after watching it, they said flatly: "This has nothing to do with what we do."

Cohen didn't give up. He shot back with a question: "Then what do you want it to do?"

That question changed everything. He realized that the traditional model of "sales collects requirements → engineers develop remotely" completely fails with highly classified, extremely complex clients. Engineers had to go on-site in person, see the real problems, before they could write useful code.

That was the birth of the FDE. CTO Shyam Sankar systematized this practice into a formal methodology.

The first two real-world FDE cases alone demonstrate its value:

The battlefields of Iraq and Afghanistan. Palantir stationed FDEs with the troops. They found that soldiers didn't need fancy intelligence analysis dashboards—they just wanted to mark on a map that "this stretch of road is dangerous." An FDE wrote a simple map tool on the spot, letting soldiers mark suspicious road segments with a single click. This feature later significantly reduced casualties from roadside bombs (IEDs) and became a standard feature of the Palantir platform.

The operation to kill bin Laden. Palantir's Gotham platform integrated roughly 3 million pieces of intelligence, 170,000 satellite images, and 8,000 hours of phone recordings, compressing money-flow tracking from 72 hours to 12 minutes. They also built a 1:64 scale simulation model of the Abbottabad compound for SEAL Team Six to rehearse before the operation.

Palantir's market cap grew from about $17 billion at its 2020 IPO to $400 billion by the end of 2025—640% in 5 years. FDEs are widely recognized as its deepest moat.

"Dirt roads" and "highways": FDE's working methodology

Palantir's FDEs work in two-person teams, like a miniature startup:

The two-person team goes deep into unfamiliar environments with limited resources and must deliver fast. Palantir is called "Silicon Valley's startup boot camp" precisely because so many entrepreneurs have come out of its FDE system.

The core methodology is "dirt road to highway":

• The FDE first lays a "dirt road" at the client's site—rough, but it immediately solves the urgent problem
• The headquarters team evaluates all the dirt roads, picks out the most valuable ones, and upgrades them into "highways"—standardized platform features
• As the platform gets stronger and stronger, new clients need less and less customization, freeing FDEs to tackle deeper problems

Three iron rules of survival:

• Must solve the CEO's Top 5 problems—without executive support, nothing moves internally
• Dare to lose money first—promise the client "a refund if it doesn't work" and build trust first
• Beware of becoming an outsourcer—clients will ask you to do simple, repetitive things, but your mission is to find the key problem that can transform their business

Palantir's core business philosophy can be summed up in one sentence: "Scale the things that don't scale." This seemingly clunky approach, which keeps compounding advantages, is the real moat.

Why did the AI era suddenly need FDEs?

In the traditional SaaS era, products were highly standardized and clients could use them right out of the box. But AI is completely different—it's a brand-new "species" with no ready-made usage paradigm.

36Kr put it precisely: For every enterprise, adopting AI is like connecting to the power grid for the first time—and every plug has a different shape.

For the same "AI lawyer" product, a big law firm needs risk flagging of contract clauses, while a small firm needs help drafting first drafts. Same model, completely different use cases. This highly fragmented demand means a purely standardized product simply can't cover it.

Three key changes have made FDEs a necessity:

Change one: contract sizes are exploding. AI contracts with big banks and logistics companies routinely run into the tens of millions of dollars. That scale of investment is enough to support FDE teams stationed on-site long-term.

Change two: the business model is shifting to token-based billing. Moving from per-seat pricing to charging by token consumption—the more AI systems an FDE deploys deep into a client, the higher the daily token consumption, which is high-stickiness recurring revenue.

Change three: structural lock-in. When an FDE spends 6 months helping a client build an AI system deeply integrated with their data, workflows, and compliance architecture, that system becomes a "load-bearing wall"—you can't swap it out.

In May 2026, two major events pushed FDE completely into the spotlight:

OpenAI Deployment Company — over $4 billion invested, 19 investors (TPP leading, with Bain Capital, Goldman Sachs, SoftBank, and others following), plus the acquisition of AI consulting firm Tomoro (bringing in roughly 150 experienced FDEs). OpenAI CRO Denise Dresser said: "The real speed of AI commercialization doesn't depend on model parameters, but on deployment capability."

Anthropic Joint Venture — valued at $1.5 billion, with top institutions like Blackstone, Apollo, and GIC participating. Interestingly, OpenAI and Anthropic have zero investor overlap — completely different institutional networks betting on the same direction.

The salary truth: a global roundup of FDE earnings

First, the US:

Now China:

A practitioner named Yasha spent 8 years as a product manager, switched to SWE, then to FDE, and has been doing it for 2 years. She said: "A fresh college graduate can't do this job... so the starting salary isn't low either — around $200K+ in the first year. In Silicon Valley, there's no upper limit."

On demand growth: LinkedIn data shows FDE job postings grew 42x between 2023 and 2025, while regular AI engineers grew 13x over the same period, and traditional software engineer postings dropped about 70% in Q1 2025.

What kind of person can become an FDE?

FDE is not an entry-level role. That's the most important sentence.

The typical backgrounds are two kinds of people: a product manager who can code, or an engineer with business sense.

Technical skills checklist (combining ByteDance's "Doubao AI Large Model FDE" JD and OpenAI's requirements):

• Full-stack development ability (Python/Java)
• LLM technical principles, performance and quality evaluation methods
• Familiarity with Post-Training techniques
• LLM product API integration
• System architecture and data pipelines
• Agentic Engineering: orchestration, memory management, task handoffs
• Physical-layer deployment ability (able to install and run models on customer hardware, not just call APIs)
• Air-gapped environment deployment experience (for defense/national lab clients)

Soft skills matter just as much:

• Analytical ability + patience: Yasha says "every day you face ambiguous requirements — it's like untangling a ball of yarn"
• Learning ability + the ability to see the essence of problems — what recruiters value most
• Communication skills: you must be able to dig real requirements out of clients who "don't know what they want"
• Technical depth: "You can't be a product manager who only communicates... you must have technical depth yourself, be able to make judgments, and even solve some on-site problems with your own hands"

You need to understand 70% of the product's full picture — not an expert in any single module, but you need to know a bit of everything, and be able to make quick judgments and act on-site.

OpenAI also has a "Tiger Team" model: deployment requires not only technical experts but also people with "institutional knowledge" — the SOPs and experience that live in veteran employees' heads and were never written into documentation. The evaluation standard is "Evals First": success metrics are defined bottom-up by frontline operators, not top-down by management.

FDE vs SWE vs Sales Engineer vs consulting: what exactly are the differences?

FDE ≠ consulting. This is something Bob McGrew particularly emphasizes. Consulting is linear — 10 projects require 10x the headcount, so margins can't fundamentally improve. FDEs also lose money early on, but frontline experience keeps getting absorbed into the platform — costs fall, value rises, margins flip from negative to positive, and a profit inflection point emerges.

Chinese practitioner Lai Juncheng's summary is more down-to-earth: "This role happens to integrate pre-sales, product, and project management into one."

Some compare FDE to the "product manager" of China's mobile internet era — hot for 5-10 years, then gradually cooling down. But in the process, "top-tier 'super individuals' like Zhang Xiaolong and Zhang Yiming will definitely emerge."

If the AI competition of the past was about who could train stronger models, the competition is now shifting to: who can get models into enterprises' real businesses the fastest.

FDEs are the people who do exactly that.

Are you an engineer or a product manager? Would you consider switching to FDE? Share your thoughts in the comments. If you found this useful, give it a like so more people can see this exploding track.

Author: itech001 Source: WeChat Official Account: AI人工智能时代 Website: https://www.theaiera.cn/ Sharing the latest AI news and technical research daily.

This article was first published on AI人工智能时代. Please cite the source when reposting.