Palantir's Forward Deployed Engineer Model Drove 640% Returns — Now Anthropic and OpenAI Are Copying It

Palantir Figured Out Enterprise AI Deployment in 2012. Everyone Else Is Just Catching Up.

Palantir IPO’d at roughly $19 in 2021, slid to $6 by 2022, and then returned 640% over five years. That trajectory is not a story about model quality or benchmark scores. It’s a story about a deployment model — specifically, the forward deployed engineer (FDE) — that nobody else in enterprise software was willing to copy because it looked too expensive and too weird. Now Anthropic and OpenAI are explicitly copying it, and if you’re building AI products for enterprise clients, you need to understand why.

The timing matters. Both Anthropic and OpenAI announced enterprise deployment ventures within weeks of each other. Anthropic’s joint venture — backed by Blackstone, Hellman & Friedman, and Goldman Sachs as founding partners, with Apollo Global Management, General Atlantic, GIC, Leonard Green, and Suko Capital also in — is valued at $1.5 billion with a $300 million commitment from Anthropic, Blackstone, and H&F. OpenAI’s parallel “development company” is raising $4 billion from 19 investors at a $10 billion valuation. There is zero investor overlap between the two. The financial establishment has split into two camps, and both camps are funding the same underlying idea: embed engineers inside client companies and ship real code.

That idea has a name. It’s the Palantir FDE model.

What the FDE Model Actually Is (And Why Normal SaaS Doesn’t Work Here)

The standard enterprise software motion goes like this: build product, hand to sales, sales closes deal, customer success helps with onboarding, customer figures it out. This works fine when the product is well-understood — a CRM, a project management tool, something with a clear interface and a known job to be done.

It breaks completely for AI.

The problem is a knowledge gap that sits in the middle of every enterprise AI deployment. The client’s engineers know everything about their business — the data schemas, the edge cases, the compliance requirements, the internal politics of which team owns which system. The AI lab’s engineers know everything about how to make models actually work — the prompting patterns, the harness architecture, the retrieval strategies, the failure modes. Neither side has the other’s knowledge. And you need both to ship something that actually runs in production.

Palantir’s insight was that you can’t bridge this gap with documentation or a customer success manager. You bridge it by taking your best engineers and physically embedding them inside the client’s organization. These aren’t consultants writing slide decks. FDEs write code. They set up the harness, configure the data pipelines, handle the weird edge cases that only appear when you’re running against real production data. They stay until the thing works.

This is expensive. It doesn’t scale the way a SaaS product scales. And it requires engineers who are comfortable operating in unfamiliar environments, which is a specific personality type that’s not always abundant at AI labs. All of which explains why the rest of the industry looked at Palantir’s model and said “interesting, not for us” — right up until the moment the revenue numbers became impossible to ignore.

Why This Moment Is Different

The reason the FDE model is suddenly the consensus strategy isn’t that anyone had a new insight about enterprise sales. It’s that the capability gap closed fast enough to make the deployment gap the binding constraint.

A year ago, you could reasonably argue that AI wasn’t good enough to justify the deployment investment. The models hallucinated too much, the context windows were too small, the tool use was too unreliable. Enterprise clients who tried to deploy AI and got burned could blame the technology. Now the technology is good enough that the failures are almost entirely deployment failures — wrong harness, wrong data pipeline, wrong integration, wrong prompt architecture. The model isn’t the problem. Getting the model into the client’s actual workflow is the problem.

This is exactly the environment where the FDE model shines. The client has a real problem. The technology can solve it. The gap is purely operational. Send in an engineer who understands both sides, and you close the gap.

Anthropic is targeting financial services first, which makes sense given that Blackstone is the world’s largest alternative asset manager. Finance has the three properties that make FDE economics work: weird and complicated problems, high stakes, and enough money to pay for the deployment cost. A hospital or a bank or a hedge fund has compliance requirements, data governance constraints, and internal systems that no off-the-shelf product will ever handle correctly. You have to go in and build the custom harness. And the value of getting it right is large enough that the client will pay for the time it takes.

OpenAI’s development company is targeting a broader set of verticals — finance, manufacturing, healthcare — which suggests a different theory of scale. Rather than going deep in one sector, they’re betting on a deployment playbook that generalizes across industries. Whether that works depends on whether the FDE model can be productized, or whether it’s inherently artisanal. The strategic divergence between Anthropic, OpenAI, and Google on agent architecture is real, but at the enterprise deployment layer, the two labs have converged on the same answer.

The Evidence That This Is Working

Anthropic’s ARR reportedly went from $9 billion to over $44 billion in 2026, doubling roughly every six weeks. Analyst Ming Li calculated that this implies Anthropic is adding approximately $96 million in ARR per day. For context: AWS took 13 years to reach $35 billion in annual revenue. Salesforce took over 20 years to pass $20 billion. These are not comparable growth rates. Something structurally different is happening.

Part of what’s different is the move from seat-based to token-based pricing. A single developer using Claude Code or Codex through the API isn’t a $20/month subscription. They’re potentially hundreds or thousands of dollars per month in token consumption, and that consumption scales with the value they’re generating. As token-based pricing becomes the norm, the revenue ceiling for a single enterprise customer becomes essentially uncapped — bounded only by how much economically valuable work the AI is doing.

The margin story is equally striking. Anthropic’s inference margins are reportedly at 70%, up from 38% last year. That’s not a company burning money to buy growth. That’s a company that has figured out how to run inference efficiently enough that the unit economics work at scale. The combination of explosive top-line growth and improving margins is what makes the FDE investment rational: you’re spending on deployment to acquire customers who then generate high-margin recurring token consumption.

Palantir’s Q1 2026 earnings showed 85% year-over-year revenue growth — their fastest pace since their public market debut. Government revenue grew 84% in Q1, up from 66% in Q4. CTO Shyam Sankar described Palantir’s position with a line that’s worth sitting with: “Tokens are the new coal. Palantir is the train.” The FDE model isn’t just a sales strategy. It’s infrastructure for the token economy.

It’s also worth noting what’s happening at the model layer underneath all of this. The compute constraints Anthropic is navigating are real, and they shape which clients get prioritized and how deeply FDEs can be deployed. Scarcity at the inference layer makes the deployment layer even more important — you need to make sure the tokens you do have are going toward high-value, well-integrated use cases, not poorly configured pilots that churn.

What This Means If You’re Building Enterprise AI

The FDE model has a specific implication for anyone building AI products for enterprise clients: the deployment layer is now a competitive moat, not a cost center.

If you’re a smaller team trying to sell AI into enterprise accounts, you probably can’t afford to embed engineers at every client site. But you can think about what the FDE model is actually doing and find cheaper ways to accomplish the same thing. The FDE is solving the knowledge gap problem — bridging the client’s domain expertise with the AI lab’s technical expertise. There are other ways to close that gap.

One approach is to build the harness so well that the client’s own engineers can deploy it without help. This is harder than it sounds, because the harness has to handle all the weird edge cases that the FDE would normally handle in person. But if you can do it, you get the scalability of SaaS with the stickiness of a custom deployment. The systems that get installed this way are extremely sticky — the client becomes dependent on your specific harness architecture, your specific integrations, your specific prompt patterns. Switching costs are high.

Another approach is to invest heavily in the onboarding layer — not customer success in the traditional sense, but actual technical depth at the point of deployment. This is where platforms like MindStudio become relevant: if you’re building an AI workflow that needs to connect to a client’s existing tools, having access to 200+ models, 1,000+ pre-built integrations, and a visual builder for orchestrating agents and workflows dramatically reduces the deployment time, even without an embedded engineer on site. The platform handles the integration surface area that would otherwise require custom engineering at every client.

The harness question is also where the abstraction level of your tooling matters. When Wes Roth described the Minecraft Voyager agent — GPT-4 navigating a game world through text descriptions of its state — he was describing exactly the kind of harness problem that makes enterprise deployment hard. The model is capable. The harness that connects the model to the real environment is the hard part. This is why Claude Code and Codex are getting so much attention: they’re harnesses that work well enough that individual developers can deploy them without an FDE. For enterprise, the equivalent harness is more complex, but the principle is the same.

For teams building production applications on top of these models, the abstraction question goes even deeper. Remy takes a different approach to this layer: you write your application as an annotated markdown spec — prose carries intent, annotations carry precision — and it compiles into a complete TypeScript backend, SQLite database, frontend, auth, and deployment. The spec is the source of truth; the code is derived output. It’s a different answer to the same underlying question: how do you reduce the distance between “what the client needs” and “working production code”? Where the FDE model solves that gap with human expertise embedded on-site, spec-driven compilation solves it by making the translation from requirements to implementation nearly automatic.

Comparing how the leading models perform in these agentic harness contexts is increasingly important for deployment decisions. The GPT-5.4 Mini vs Claude Haiku 4.5 sub-agent comparison is a useful reference point here — the differences in reliability, tool use, and latency at the sub-agent layer compound significantly when you’re running complex enterprise workflows with many chained steps.

The Stickiness Problem Is the Point

There’s a reason both Anthropic and OpenAI are making this investment now, and it’s not just about current revenue. It’s about lock-in.

Enterprise AI systems, once deployed, are extremely difficult to replace. The harness is custom. The integrations are custom. The prompt architecture is tuned to the client’s specific data and workflows. The client’s engineers have learned how to work with the system. Switching to a different model provider means rebuilding all of that. The FDE model accelerates this lock-in by making the initial deployment so thorough that the switching cost becomes prohibitive.

This is the same dynamic that made Palantir’s stock recover so dramatically after 2022. The clients who had FDEs embedded and systems running weren’t going anywhere. The revenue was sticky in a way that pure SaaS revenue isn’t. As Palantir’s installed base grew, so did the floor on their revenue — and the ceiling on their growth, because each installed client became a reference for the next one.

Anthropic and OpenAI are trying to build the same installed base, but faster and at larger scale. The $1.5 billion JV valuation and the $10 billion development company valuation are bets that the FDE model can be applied to AI at a speed and scale that Palantir never attempted. Given that Anthropic’s ARR is reportedly doubling every six weeks, the early evidence suggests the bet is paying off.

The Deployment Gap Was Always the Real Problem

The AI bubble narrative that dominated 2024 and early 2025 was built on a real observation: AI wasn’t getting deployed at the rate that the technology’s capabilities seemed to warrant. If the models were this good, why wasn’t every enterprise using them?

The answer wasn’t that the models weren’t good enough. The answer was that the deployment gap was real and nobody had solved it at scale. Palantir solved it for government and defense. Now Anthropic and OpenAI are solving it for finance, healthcare, and manufacturing. The FDE model is the solution, and the 640% return Palantir generated over five years is the proof of concept.

What’s interesting about the current moment is that the deployment gap is closing from both directions simultaneously. The FDE model is pushing from the top down — embedding expert engineers to handle the hardest deployments. And better tooling is pushing from the bottom up — making it easier for client engineers to deploy AI themselves without needing an embedded expert. The middle ground, where most enterprise AI actually lives, is getting squeezed from both sides.

For builders, this means the window for building deployment infrastructure is open right now, but it won’t stay open indefinitely. The labs are investing heavily in making their own deployment easier. The clients are getting more sophisticated. The tooling is improving. The FDE model is being productized. In two or three years, deploying enterprise AI will probably be much more like deploying enterprise SaaS — still complex, but with established patterns and tooling.

Right now, though, the knowledge gap is still wide enough that the people who understand both sides — the model capabilities and the enterprise requirements — are generating enormous value. That’s the real lesson from Palantir’s 640% return. It wasn’t about the technology. It was about being the people who knew how to make the technology work in the real world, and being willing to go sit inside the client’s office until it did.

Anthropic and OpenAI have read that lesson clearly. The question is whether you have too.