Latest / The CTO Podcast with Fexingo: Technical Leadership, Architecture, and Engineering Org / How Palantir Migrated to the Cloud Without Losing Security Clearance
Transcript
- Lucas: Palantir is one of the most secretive companies in Silicon Valley — not because they're mysterious for marketing reasons, but because their customers are literally the US military, the CIA, and allied intelligence agencies. So when they decided to move their entire software stack to the cloud, they couldn't just spin up some EC2 instances and call it a day. Luna: Right — 'lift and shift' is hard enough for a fintech startup. For a company that operates at TS/SCI clearance levels, it's a different universe. Lucas: Exactly. And the timeline matters here. Palantir started this migration in earnest around 2020, but the real catalyst was the Pentagon's Joint Warfighting Cloud Capability contract, or JWCC, which was awarded to AWS and Microsoft in late 2022. That contract effectively mandated that defense contractors begin moving workloads to the cloud. Palantir needed to be ready. Luna: So they weren't building from scratch. They already had a product — Gotham for intelligence analysis, Foundry for data integration. The question was: how do you take software designed for air-gapped, on-premise government data centers and run it on AWS without losing your security accreditation? Lucas: That accreditation piece is the whole story. Palantir's government deployments are accredited at what's called Impact Level 5, or IL5, which is the DoD's classification for controlled unclassified information — but they also have customers operating at even higher levels. The accreditation process normally takes months, sometimes years. Palantir couldn't afford to lose that every time they changed a server. Luna: So how did they get around the audit bottleneck? Lucas: They built what they call a 'cloud bridge' — a hybrid architecture that lets Palantir's software run simultaneously in government on-premise data centers and in AWS GovCloud. The key insight was that they didn't move everything at once. Instead, they created a layer that abstracted away the underlying infrastructure. Gotham and Foundry talk to a 'control plane' that sits on top of both environments. That control plane handles identity, data routing, and policy enforcement. Luna: So the applications themselves don't know whether they're running on a Palantir-owned server or an AWS instance. The control plane makes the decision. Lucas: Exactly. And that's critical for security because it means the data never leaves the accredited boundary until the control plane says it can. But the more interesting technical detail is how they handled networking. In a traditional government data center, you have physical air gaps — literally separate rooms with separate cabling. In the cloud, everything is software-defined. So Palantir had to implement a zero-trust networking model inside the VPC. Luna: Zero trust meaning no implicit trust based on network location. Every request has to be authenticated, authorized, and encrypted — even between microservices running on the same host. Lucas: Right. They built a service mesh using Envoy proxies on every node. Every inter-service call goes through mutual TLS. And they integrated that with their own authorization engine, which is basically a policy as code system called 'Palantir Authorization'. So if an analyst in the Pentagon tries to query a dataset, the request goes through the Envoy proxy, which checks a signed JWT, then the authorization engine evaluates the policy based on the user's clearance level, the data's classification, and even the time of day. Luna: I've heard that Palantir uses a lot of formal verification in their policy engine — like, they can mathematically prove that certain access patterns are impossible. Lucas: They do. The authorization engine is built on something called 'Zelkova', which is a logic-based policy language. It's not just a set of if-then rules; it's a declarative system where you write constraints. For example, a policy might say 'an analyst with TOP SECRET clearance can read records classified SECRET or below, but only if their unit is assigned to that mission'. And the engine can verify that no combination of inputs can violate that constraint. That's huge for accreditation because auditors can look at the formal model instead of manually inspecting every configuration file. Luna: So the migration itself became as much a software engineering problem as a compliance problem. Lucas: Completely. Palantir actually open-sourced a tool called 'Conjure' that generates client libraries and API documentation from their internal type definitions. That was part of the migration too — they needed to ensure that the interfaces between their cloud and on-premise components were consistent and auditable. So they codified everything in Conjure definitions, which then get compiled into Java, TypeScript, Python — whatever the service needs. Luna: And this is running on AWS GovCloud, which is physically separate from the commercial AWS regions. But even within GovCloud, Palantir had to create their own isolation boundaries. They couldn't just use AWS accounts because the government customers require tenant separation at the hardware level. Lucas: Right. So they used AWS Nitro Enclaves for compute isolation. Nitro Enclaves are hardened virtual machines that have no persistent storage, no network access except through a secure channel, and no shell access for administrators. Palantir runs the most sensitive data processing inside those enclaves. Even if an AWS employee gets physical access to the host, they can't inspect the enclave's memory. That's the level of assurance these contracts require. Luna: And that's also why the migration took years. Every component had to be redesigned or at least re-tested against these requirements. I remember reading that Palantir had to rebuild their entire data replication pipeline because the old one assumed a flat network. Lucas: Yeah, the replication piece is a good example. In the on-premise world, Palantir used a custom file system called 'Phosphorus' that replicated data across racks. In the cloud, they switched to a system built on Apache Kafka and their own distributed log. But the challenge was that government auditors needed to verify that data replicated between regions didn't accidentally cross classification boundaries. So Palantir added a 'data tagging' layer — every record carries a classification label, and the replication pipeline checks those labels before moving data. If a label is missing, the replication blocks and alerts the operations team. Luna: What about the cost? Moving to the cloud usually saves money on hardware, but the compliance overhead must add significant expense. Lucas: Palantir doesn't publicly break out the cost of their cloud infrastructure, but former engineers have said that the compliance automation alone — the continuous monitoring, the automated evidence collection for audits — was a multi-year investment. However, the trade-off is that once you're accredited in AWS GovCloud, you can spin up new environments for new customers in days instead of months. The old model required Palantir to ship physical servers to a government facility, get them installed, then go through the months-long accreditation process. Now they can provision a new IL5 environment via API. Luna: That speed is real. The DoD's whole JWCC strategy is about exactly that — getting software to warfighters faster. Palantir's migration is a case study in how to do it without sacrificing security. Lucas: It also changed the relationship between Palantir and its government customers. In the old model, Palantir deployed on-premise and handed over the keys — the customer's IT team managed the hardware. In the cloud, Palantir manages the control plane, so they have more operational visibility. Some privacy advocates have raised concerns about that, but from a technical standpoint, it's a more reliable architecture. Palantir can push updates, monitor performance, and respond to incidents across all their government customers from a single operations center. Luna: And they've been able to attract more commercial customers too, because the same infrastructure that meets DoD IL5 standards also meets HIPAA, pci dss, and SOC 2. So Palantir Foundry for healthcare or financial services can run on the same stack. Lucas: That's a huge unlock. Before the cloud migration, Palantir had to build separate infrastructure for commercial customers, which was inefficient. Now they have one platform that can serve both markets. Their Q1 2026 earnings call mentioned that commercial revenue grew 40 percent year-over-year, and the cloud migration is a big part of that story. Luna: And you know, this kind of deep-dive into how a company actually re-architects for the cloud is exactly why listeners support shows like this. If you've gotten value from understanding how Palantir navigated this migration — or any of our previous episodes on technical leadership — a couple of dollars a month genuinely makes a difference. Buy me a coffee dot com slash fexingo is where that works. It keeps the podcast ad-free and focused on the architecture behind the headlines. Lucas: Absolutely. And it lets us keep digging into the engineering decisions that most business coverage skips. Like how Palantir decided to use a 'staging' cloud environment that mirrors production exactly, so they could test the migration of a specific intelligence agency's workload before cutting over. Luna: Right — they called that the 'shadow environment'. Every query sent to production was also sent to the cloud shadow, but the results weren't returned to users. They compared outputs for weeks to catch discrepancies in the access control policies. Lucas: Exactly. And that's the kind of paranoia you need when lives are at stake. Palantir's cloud migration isn't just a technical achievement — it's a blueprint for any organization that needs to modernize without lowering their security bar. The principles apply whether you're a defense contractor or a hospital system. Luna: And that's the conversation we'll keep having on this show: how the smartest technical leaders actually ship. Thanks for listening.