There’s an engineer paid $785,000 a year. But this person doesn’t build models. Doesn’t research new architectures, doesn’t spin up bigger GPU clusters. The job is to embed directly in a customer’s office, figure out how that company actually works, and turn it into code that runs. This is the story of the fastest-growing role in tech in 2026: the Forward Deployed Engineer (FDE).
The hiring data speaks first
What matters is that this is numbers, not vibes. FDE postings grew from 643 in April 2025 to 5,330 in April 2026 — a 729% jump in a single year. It ranks among the fastest-growing roles in all of tech. Palantir, OpenAI, Anthropic, Mistral, and Cohere are hiring at once, and Salesforce has committed to fill 1,000 FDE seats.
The price matches. Palantir’s median total comp runs about $215K; senior FDEs at Anthropic and OpenAI clear $785K. A seat that means “sitting next to the customer writing code” now pays as much as building models, or more.
Why now: the 95% isn’t the model’s fault
The reason is uncomfortably simple. 95% of enterprise AI pilots fail. And the cause isn’t that the model is dumb. It’s that deployment is broken. The model runs fine in a demo. The trouble is that the moment you bolt it onto the company’s real data, legacy systems, regulations, exception handling, and the inertia of “this is how we’ve always done it,” everything comes apart.
The FDE goes in to close exactly that gap. They ship working code, not slides or docs; they debug integration on site, configure the model against the customer’s real data, and train internal teams to operate it on their own. So the industry’s conclusion has settled like this — the model isn’t the product; deployment is. The company that wins the enterprise contract isn’t the one with the best model, but the one that can actually plug that model in.
Palantir knew the answer 18 years ago
The funny part is that this isn’t a new invention. Palantir created the role in 2008. The idea was to send engineers to the customer’s front line instead of the clean abstraction of headquarters. OpenAI and Anthropic pulled the model back out in 2023 for the LLM era, and by 2026 any Series A startup with six-figure ACV keeps at least one FDE.
Why it suddenly became everyone’s model 18 years later is the crux. As model capability became common, the contest moved from “who has the better model” to “who plugs it into the customer’s reality better.” And that work can’t be done sitting at HQ. You have to go inside the customer’s domain.
What an FDE actually has: a T-shape
So who is an FDE? The definition is consistent. Not a salesperson or consultant, but a hands-on-keyboard builder. At the same time, someone who can explain to a non-technical executive why inference latency matters. A T-shaped profile: deep technical ability to write production code, with the breadth to understand the customer’s work and language layered on top.
Why that combination is expensive is clear. Coding ability alone is now largely absorbed by agents. Understanding the customer’s domain alone — a consultant covers that. The FDE’s value is that the two meet in one person: someone who knows the problem and can turn it into a running system themselves.
The moat comes from being embedded
This is also why the FDE model is strong as a business. Once code is woven deeply into how a customer’s organization operates, the cost of ripping it out to switch vendors becomes enormous. So FDE-centric companies have high acquisition cost but very high retention and large contracts. The embedded system itself becomes the moat.
This is the exact structure of the vertical-AI moat from the last post. Just as Harvey’s or Sierra’s strength came not from the model but from the integrations and procedures embedded into the work over tens of thousands of hours, an FDE’s strength comes from being embedded in the customer’s workflow. At the company level it’s vertical AI; at the person level it’s the FDE. The same moat, cut from a different angle.
The same thesis, a third proof
Get here and the picture lines up. In AI coding, what separated success was domain expertise, not coding skill (#26), and a company’s moat was domain data and workflow, not model weights (#27). The FDE is that thesis come down to the labor market. The market has put a $785K price tag on domain understanding and on-the-ground deployment.
Three layers point at the same spot. In the user data, in the company moats, and now in the job postings and salaries — what becomes common (coding skill, model capability) loses value, and what doesn’t (knowing the domain and plugging it into reality) gains it. For a while we asked which model is smarter. The market is already betting on a different question: who can plug that model into reality.