AWS Just Committed $1B to a Talent Pool of 2,740. The Math Doesn't Work.

On June 30, 2026, AWS announced a $1B Forward Deployed Engineering org that will embed "thousands" of engineers in customer AI deployments. That announcement joined Anthropic's $1.5B FDE joint venture with Blackstone and Goldman (May) and OpenAI's $4B Deployment Co with TPG and Bain (also May). Three of the largest AI buyers on earth are now competing, with $6.5B in committed capital, for the same archetype Palantir has been quietly training since 2005.

The problem is not that FDEs are expensive. The problem is that they barely exist.

The three biggest AI vendors are bidding on the same 2,740 people

There are roughly 2,740 people in the United States who currently or previously hold a "Forward Deployed Engineer" title, and about 5,870 globally. AWS alone says it wants "thousands." The arithmetic does not close.

In Refolk's index of professional profiles, "Forward Deployed Engineer" and its close variants surface as one of the rarest technical titles in circulation. Here is the shape of what all three vendors are fighting over:

AWS' pod model runs 45-day cycles with five or six engineers per client, per Francessca Vasquez, AWS' VP of Frontier AI Engineering and Services. Divide "thousands" by six and the entire AWS org can service, at most, a few hundred concurrent enterprise deployments. If OpenAI and Anthropic both hit their stated plans as well, the three combined would need to hire roughly the entire global FDE-titled workforce, and then some.

2,740

At least one of the three big spenders will structurally miss its plan. That is not a forecast, it is a subtraction problem.

Why Palantir alumni became the moat (and why they're already spoken for)

Palantir is the only company that has trained FDEs at industrial scale, and its alumni are the single most-recruited profile in enterprise AI right now. Every FDE recruiter is emailing roughly the same 500 people.

Until 2016, Palantir employed more FDEs (internally called "Deltas") than software engineers. That is a decade-plus of on-the-job training in the exact motion the AI labs are now scrambling to reproduce: land at a Fortune 500, scope a messy data problem in a week, ship a working prototype in three, and stay embedded until it runs in production.

In Refolk's index, Palantir accounts for 20% of the top US FDE company cluster. No other company appears more than twice. New York (8 of the top 25) and San Francisco (5) dominate, DC is a distant third at 2, and the entire top-region European footprint across Paris, London, and Hamburg totals about four people. Those NYC and DC concentrations map cleanly onto Palantir's Gotham and Foundry deployments in finance and defense.

Which is why Anthropic's job description does not pretend otherwise. It explicitly names three feeder backgrounds:

• Senior solutions engineering at infrastructure companies (Snowflake, Databricks, MongoDB)
• Palantir forward deployed engineering
• Full-stack engineering with heavy customer-facing experience at high-growth startups

Anjor Kanekar, a seven-year Palantir FDE now advising on FDE org design, is the profile every AI lab wants: someone who has run the 45-day cycle across dozens of accounts and can now teach a pod to do it. There are not many Anjors.

Every FDE recruiter in enterprise AI is emailing the same 500 people. The reply rates reflect it.

The comp war: Anthropic is paying 2.5x Palantir base to poach the same profile

Anthropic Applied AI Engineers earn $300K to $600K total comp, with senior ICs at $450K to $550K all-in. Palantir's own FDSE listing is $135,000 to $200,000 base. AI labs are paying roughly 2.5x cash to move the same person across the street.

That gap explains the last 18 months of resignations on the Palantir engineering roster more than any single manager change. It also explains why the labs cannot solve this with money alone. There are only so many senior FDEs; once every one of them has a $500K offer, the marginal dollar buys nothing.

The comp curve also disqualifies most 2023-era "AI engineers." Anthropic requires 3+ years in a technical customer-facing role plus production LLM experience: advanced prompt engineering, agent development, evaluation frameworks, and deployment at scale. Someone who spent 2024 wrapping a RAG demo in Streamlit does not clear the bar, no matter what their title says.

The Tomoro acquisition is the tell

OpenAI's May 2026 acquisition of Tomoro, a 150-person applied AI consulting firm with prior deployment work at Tesco, Virgin Atlantic, and Supercell, is the loudest signal in this market. It says the highest-comp employer in AI could not hire the pipeline organically and had to buy a company to instantly claim 150 FDEs.

Expect Anthropic and AWS to make similar acqui-hires by Q4 2026. The candidates are already visible: mid-sized applied AI consultancies, boutique data-science shops with named-account logos, and the customer engineering arms of Series C infra startups. When the org chart says you need "thousands" and the market has 2,740, M&A becomes a sourcing channel.

Where the real supply lives (hint: it is not in the "FDE" title)

The adjacent titles are 5 to 10 times more populous than "Forward Deployed Engineer," and the good ones have most of the same skills. Recruiters who limit searches to the literal title are working the smallest possible pool.

The feeder pools worth mapping:

1. Solutions engineers and solutions architects at Snowflake, Databricks, MongoDB, Confluent, and Elastic. Same customer motion, weaker LLM depth, easy to close on comp.
2. Deployment strategists and customer engineers at Modal, Roboflow, Cresta, Gecko Robotics, and Amp Code. Small clusters in Refolk's index, but every profile in them has shipped real production AI at a named account.
3. Applied ML engineers with a services background at boutique shops like Tomoro (before OpenAI bought it), QuantumBlack, and BCG's GAMMA practice.
4. Ex-consultants who moved into engineering at Ramp, Rippling, and Brex, where the FDE-adjacent motion has existed under different names for years.
5. Second-year Palantir FDEs, not the alumni. The people still inside are the ones every hiring manager is quietly hoping to reach.

Leo Mehr's FDE org at Ramp is the proof point that this profile now matters outside the hyperscalers. Ramp stood up its FDE function about nine months ago and runs around 15 engineers in pods. If Ramp is bidding, every mid-market SaaS company with an AI story will be bidding by mid-2026.

This is the exact gap Refolk closes for sourcers working the FDE market: you describe the person in plain English, including the feeder titles Anthropic actually names, and get a ranked shortlist across GitHub, LinkedIn, and the open web. The Boolean string version of "Palantir OR Snowflake OR Databricks solutions engineer with LangGraph experience in NYC" returns noise. A natural-language query returns people.

The agentic requirement quietly cuts the pool in half

AWS' pod model is not five humans, it is five humans plus purpose-built agents that are meant to keep running after the humans leave. FDEs who cannot ship agent frameworks are effectively unqualified, which shrinks the real pool well below the 2,740 headline.

The concrete stack the top-tier FDE roles now assume:

• Agent orchestration: LangGraph, DSPy, or a custom equivalent
• Tool-use plumbing: MCP servers, function calling patterns, retrieval routers
• Evaluation infrastructure: eval suites that run on every deploy, not spot checks
• Deployment: production LLM apps at scale, not notebooks

Most "FDE" resumes written before mid-2025 do not have these on them, even if the person is otherwise excellent. The candidates who ship agent-first are disproportionately concentrated at a small set of AI-native companies, and they are the exact people the labs and hyperscalers most want to poach from each other.

95%

What to actually do if you're hiring FDEs right now

Stop searching on the title. The 2,740-person US pool is drafted, and the Palantir subset is triple-drafted. The playbook that works in Q4 2026:

1. Search on the motion, not the title. Look for customer-facing engineers who have shipped a named-account deployment in the last 18 months, regardless of what their business card said.
2. Prioritize the second-tier training grounds. Modal, Roboflow, Cresta, Gecko Robotics, Amp Code. Small footprints, mostly un-poached.
3. Read the agent stack, not the resume. LangGraph, DSPy, MCP, and a real eval suite in the GitHub history beat any title.
4. Reach the currently-employed, not the alumni. Everyone is emailing ex-Palantir. Fewer people are emailing current Palantir Y3 FDEs.
5. Consider the acqui-hire. If you need 40 FDEs and the market has 40 FDEs available, buying a 40-person consultancy is not exotic, it is now standard practice.

Refolk was built for exactly this kind of query, where the title filter fails and the real signal lives in what the person actually did last quarter. Ask for the motion in plain English and get the people, across GitHub, LinkedIn, and the open web.

FAQ

What is a Forward Deployed Engineer, exactly?

A Forward Deployed Engineer, or FDE, is a customer-facing engineer who embeds inside a client's team to scope, build, and ship production software against that client's real data and workflows. Palantir invented the modern version of the role in the mid-2000s and called them "Deltas." The 2026 version, as defined by AWS, OpenAI, and Anthropic, adds agentic AI: the FDE ships not just the app but the eval suite and the long-running agents that continue operating after the pod rotates off.

Why can't AWS, OpenAI, and Anthropic just train more FDEs?

They can, and they will, but the timeline does not match the announcement. Palantir spent roughly a decade training a workforce that until 2016 outnumbered its software engineers. The AI labs are trying to compress that into 18 months while also shipping. Training also requires senior FDEs to mentor juniors, and every senior FDE in the market already has three competing offers. The bottleneck is not budget or intent, it is people who can teach the motion.

Is the Palantir alumni pool really tapped out?

Effectively yes for the top of the market. In Refolk's index, Palantir accounts for 20% of the top US FDE cluster, and the New York and DC concentrations trace back to Palantir Gotham and Foundry deployments. The most senior alumni are already at Anthropic, OpenAI, Ramp, or advising. The pool is not literally empty, but the reply rates on cold outreach have collapsed because the same ~500 people are getting every message.

If FDE-titled supply is 2,740 in the US, where should recruiters look instead?

Look at feeder titles. Anthropic explicitly recruits Snowflake, Databricks, and MongoDB solutions engineers, which is a pool five to ten times larger. Add customer engineers at Modal, Roboflow, Cresta, Gecko Robotics, and Amp Code, plus applied ML engineers coming out of BCG GAMMA and QuantumBlack. The people who can do the job outnumber the people with the title by roughly an order of magnitude. The trick is describing the motion instead of filtering on the two words.