# Why OpenAI and Anthropic are both building FDE teams

> The two labs leading the model race are hiring more and more people who do not do research. The reason is a gap that customers cannot close on their own.

Original: https://fdetimes.net/en/analysis/why-openai-anthropic-build-fde-teams/

At Anthropic, the share of job postings devoted to adoption has risen from 5% to 11%. At OpenAI, it has gone from 11% to 17%. The two labs are still seen as competing on model quality, but Epoch AI's job-posting data show them committing a growing share of their workforce to helping other people actually use what they build.

Epoch reads the figures in a way the industry may not enjoy: customers are struggling to get full value from OpenAI's and Anthropic's products. The models have outrun companies' ability to absorb them. The gap between the two is exactly where the FDE appears.

If you are a developer thinking about moving into FDE work, these numbers deserve more of your attention than any new model release. They show that demand does not come from a single company but from the structure of the market, and the way the labs are responding is sketching out a fairly clear career path.

## The adoption gap is not the customer's fault

It is tempting to assume companies are slow because they lack talent. Joe Schmidt of a16z offers a different view: enterprise AI products require even more implementation effort than the previous generation of enterprise software. That is a property of the product, not a failing of the buyer. A language model is only worth something once it is wired into real data, real processes and constraints nobody wrote down.

That is why a16z concludes that it sometimes takes labour-intensive services to produce software that genuinely changes how an organisation runs. It is blunt about the rest: at the best complex AI application companies, this role is often rebranded as forward deployed engineer or implementation specialist. In other words, the FDE title is largely a new label on familiar professional services work.

Decrypt captures the job with an apt image: FDEs parachute into a customer's organisation and work in the middle of its mess. The model is borrowed from Palantir. The easiest way to picture it is a bank that has signed an API contract but, six months later, still has no use case in production, because its data is scattered across three legacy systems and the compliance team has not signed off.

The person sitting with them untangling each knot is neither an account manager nor a researcher.

## Money follows deployment speed

The Pragmatic Engineer points to a more pragmatic reason: for the labs, the faster an AI solution is deployed, the more revenue it generates. A signed contract earns nothing if the tokens are not consumed. Every month a customer spends floundering is a month of deferred revenue, while the cost of training the model has already been paid. Seen this way, the FDE is the person who shortens the distance between signature and invoice.

But if FDEs matter that much, why are neither OpenAI nor Anthropic keeping them inside the lab? According to The Pragmatic Engineer, OpenAI has put this work into an independent legal entity backed by outside private equity, and Anthropic has also set up a separate company with outside funding for its FDE consulting arm. For OpenAI, the stated reason is to free the lab to focus on building better models.

For Anthropic, The Pragmatic Engineer records only the structure: a separate company, funded from outside. Read through OpenAI's logic, the fact that both labs chose the same structure suggests they are solving the same problem: keeping research insulated from the rhythm of a services business.

The detail worth noticing is speed. Decrypt reports that OpenAI acquired Tomoro to have around 150 FDEs from day one for the new unit. Nobody buys a company to get 150 people if they think they can hire them gradually over a year. When a lab chooses to buy rather than build, it is saying the skill is scarce and needed urgently.

## Four reasons, one conclusion

Set the four drivers side by side and each demands a different skill from an FDE.

| Driver | Who points it out | What it demands of an FDE |
|---|---|---|
| Customers are not getting full value from the models | Epoch AI, from hiring data | Understand the model well enough to see what it can do that the customer has not spotted |
| AI products need heavier implementation than older software | a16z | Integrating with messy data and processes |
| Faster deployment brings revenue sooner | The Pragmatic Engineer | Measure outcomes and prioritise the use cases that make money first |
| FDEs are the first source of R&D ideas | Palantir, 10-K filing | Capture patterns from the field and carry them back to the product team |

The last row is the easiest to overlook. In its 10-K for fiscal year 2020, Palantir wrote that its FDEs are the front line in identifying research and development opportunities for the platform. The model, in other words, does not just consume an existing product; it feeds back into it. For a lab, the person sitting inside a bank or a factory is the earliest sensor of where the model fails in the real world.

**Key point:** The labs are building FDE teams not because their models are weak, but because the models have become more capable than customers can use, and that gap is both stuck revenue and wasted product data.

## What spinning out a separate company says about the job

Both labs placing FDEs in independent legal entities carries two contrasting implications for you. On one hand, it shows FDE work being run as a services business, separate from the rhythm of the research team, which fits a16z's description of professional services under a new name. On the other, outside capital and 150 people from day one show this is not a support function but a business built on purpose.

If you are aiming for a lab's deployment unit, the most valuable thing to invest in is evidence that you have got a system running in an environment that was not designed for it. That advice follows directly from the image Decrypt uses to describe the job: parachuting into a customer's organisation and living in the middle of its complexity.

Vietnamese developers working in outsourcing or building solutions for domestic businesses have an unexpected advantage here: they are used to dirty data, manual processes and clients who do not know what they want until they see a demo.

Your CV should therefore change its axis. Instead of listing frameworks, tell the story of a deployment: what system, what obstacles on the customer's side, how you cleared them, and what measurable result followed. When reading job descriptions, look for phrases such as "customer environment", "production deployment" and "work on-site with clients" rather than technology lists. That is the language of people looking for FDEs.

As long as customers struggle to get full value from the models, as Epoch reads from the hiring data, the people standing at the junction between lab and enterprise will be needed.

When a lab chooses to buy 150 engineers rather than wait, the question for you is no longer whether this job is real, but whether you already have a deployment story to tell.

**Try this week:**

- Pick an organisation you know well (a former employer, a current client) and write down three specific reasons it still cannot use an LLM API despite having an account: data, process, or the people running it.
- Rewrite one line of your CV using this formula: what system, running in which customer environment, with what measurable result, instead of listing technologies.
- Read three FDE or Deployment Engineer job descriptions closely and underline the words they share, to see what recruiters are looking for.

## Sources

- [AI lab job postings (Epoch AI Gradient Updates)](https://epoch.ai/gradient-updates/ai-lab-job-postings)

- [services led growth (a16z, Joe Schmidt)](https://a16z.com/services-led-growth/)

- [The Pulse: Forward deployed engineering heats up again](https://blog.pragmaticengineer.com/the-pulse-forward-deployed-engineering-heats-up-again/)

- [OpenAI Just Launched a Consulting Arm to Help Companies Deploy AI (Decrypt)](https://decrypt.co/367403/openai-launched-consulting-arm-help-companies-deploy-ai)

- [Palantir Technologies Inc. - Form 10-K - FY2020](https://www.sec.gov/Archives/edgar/data/1321655/000119312521060650/d65934d10k.htm)