Forward-deployed engineering (FDE) is now a core part of enterprise AI deployment. In the first half of 2026, FDE job postings grew 729% on Indeed, while the share of companies planning to hire FDEs for AI enablement went from 5-10% to above 70% in a survey of 180 companies. So what does that mean for AI deployments?

The FDE model originated at Palantir, which defines the role as an engineer embedded directly with a customer . The FDE adapts a general software package to that customer’s specific systems and workflows.

Over the last six months, AI labs and hyperscalers accelerated the rollout of FDE organizations largely because customers were buying AI model access faster than they were putting it into production. A model is a general-purpose input rather than a finished application. Getting from one to the other requires work specific to the buyer: connecting the model to the organization’s systems of record, encoding rules that staff apply from precedent rather than documentation, and establishing what a correct output looks like for that buyer. In many cases, little of this work transfers to the next customer, limiting how much can be recovered through productization and leaving the balance to be delivered by FDEs.

I discussed this with Asli Gulcur , a former Fortune Global 500 executive who built, managed and stress-tested agentic systems while earning an MS degree in AI Systems Management at Carnegie Mellon. Previously, Asli was at Salesforce, working on the company’s critical Agentforce platform at incubation. Asli’s view is that agentic deployment work is recurring rather than a one-time integration, which is why companies are building FDE capacity they own rather than contracting hourly.

The Four Paths To FDE Capacity

In 2026, four structures have emerged for acquiring FDE capacity, summarized below:

Equity joint venture. Frontier AI labs formed independent operating companies capitalized by outside investors. The sponsors in these partnerships hold positions in thousands of companies across their funds and shape technology policy for the ones they control, which Asli characterized as “distribution into thousands of portfolio companies.”

On May 4, Anthropic established a joint venture and in July officially launched Ode with Anthropic alongside Blackstone, Hellman & Friedman, Goldman Sachs and a consortium of investors targeting mid-sized companies. On May 11, OpenAI formed the Deployment Company at a $10 billion initial valuation with 19 investors led by TPG and including Bain & Company and McKinsey. Notably, OpenAI added terms to its FDE arm that Anthropic did not – its investors hold preferred equity and are guaranteed a minimum 17.5% annual return over five years.

Balance-sheet build. In this case, the FDE function is added directly into an existing enterprise and funded straight from its balance sheet. Meta, AWS and Microsoft recently announced their FDE teams within six weeks of each other. Meta established an enterprise unit in late May, and AWS committed $1 billion for FDEs on June 30. Microsoft then committed $2.5 billion and 6,000 headcount on July 2, positioned as model-agnostic with intellectual property retained by the client. Salesforce also has committed to a team of 1,000 FDEs .

Channel agreement. Other AI vendors have established contracts to place their own FDEs into a sponsor’s portfolio companies on multi-year terms, without forming an entity or exchanging equity. Google Cloud signed agreements with Thoma Bravo on April 15, covering companies valued at more than $300 billion; with Vista Equity Partners on April 22, covering more than 90 companies; and with EQT on May 28, supporting more than 300 companies.

Capability acquisition. Anthropic’s Ode acquired Fractional AI in May and Casper Studios in August . OpenAI acquired Tomoro and its 150 engineers to launch the Deployment Company, then Northslope in July. For both AI labs, acquisition remains a viable and likely option for continuing to build out their FDE AI teams.

AI deployment creates two categories of FDE work that a standard product cannot supply on its own.

1. AI agents require company-specific guardrails. Enterprise software is primarily deterministic: programs execute defined rules to produce defined outputs. AI models, by contrast, are probabilistic, so the same instruction can produce different outcomes.

FDEs therefore perform what the industry calls harness engineering: building the environment around the model, including deterministic controls that specify which systems and data an agent may access, which actions it may take and when human approval is required. Asli describes this as a spectrum: “Some firms may allow an agent to complete a purchase autonomously, while others may reserve decision authority for human managers. The design reflects each company’s regulatory obligations and operating protocols.” Encoding those policies as enforceable technical controls is bespoke work, which is why FDEs are embedded within enterprises.

2. AI agents require continuous system maintenance because their behavior can drift after deployment. An agent continues to change its behavior once it goes live, because both its underlying model and operating environment continue to change.

Asli notes, “Providers upgrade and retire models on published schedules . Enterprises route work across models to control usage-based costs, sending routine tasks to lower-cost or open models, and to frontier models for more complex work.” Even when the model remains unchanged, updates to data, tools and company policies can cause its performance to drift from what the company initially tested and approved. FDEs therefore retest agents after model changes, monitor them between changes, and revise their instructions, tools or controls when behavior shifts. Deployment becomes a continuous cycle of testing, evaluation and re-engineering, rather than a one-time implementation.

Why AI Labs Now Employ FDEs

Engineering consultancies could do this work by the hour, and the labs already route it to them in some cases. OpenAI certified McKinsey, BCG, Accenture and Capgemini to deploy its platform through Frontier Alliances in February. Anthropic put $100 million behind a certification network for consultancies deploying Claude in March. Both channels supply engineers the labs do not employ. But then both raised billions to hire their own.

We can think of three reasons why:

1. They move up the AI stack. Whoever employs the FDE teams owns what gets built. Each deployment lowers the cost of the next one in that industry, so the customer funds work the vendor keeps. “Every deployment teaches the vendor a workflow it did not previously own,” Asli says.

2. They build an enterprise channel ahead of their expected IPO. Both AI labs have filed confidentially to go public. Neither venture is likely to show material revenue from these channels yet, and revenue is not the constraint at this stage. But the eventual topline acceleration that is expected to accompany the buildout of FDE capacity is a key element of the AI labs’ making a strong case for future growth when marketing their IPOs. The growth case is stronger where it can be verified externally through the labs’ direct engagements with enterprise clients.

3. They fill compute they have already purchased. OpenAI is contractually committed to purchase 750 megawatts of inference capacity through 2028, and that fixed cost needs durable demand. Consumer subscriptions can be cancelled monthly; however, an enterprise process, once embedded, is harder to remove. FDEs also influence which model powers each workflow. Neither lab has specifically framed its FDE venture in this exact way, but their structures are consistent with this strategic imperative.

The Question For Vertical AI

Through FDEs, AI labs and hyperscalers accumulate industry expertise one customer at a time.

An FDE team that has built deal workflows at twenty investment banks learns how those processes actually run and which exceptions recur. That accumulated knowledge is an AI product specification applicable to many other deal teams, developed through customer-funded deployments.

Repeat the pattern across multiple other verticals, and model providers can build vertically focused AI agent systems alongside their models. The competitive question is whether recurring workflows can be turned into real AI products that can hold up against what vertical AI vendors already sell today.

Asli sees this as well, but notes the limit. “Not every vertical will experience this the same way. Turning one firm’s build into a product that works at the next firm is hard.”

The 729% rise in FDE hiring is the market pricing that work: turning a model into a custom-built solution. The labs, the hyperscalers and the sponsors have all decided to own it.