The walls we rebuilt with cloud modernization

The Boeing-built SLS, Lockheed-built Orion capsule, the Airbus-built Service Module, and Northrop-built SRBs formed the core of Artemis II — but throughout the mission, hardly any procedures moved directly between Boeing, Lockheed Martin, Airbus, and Northrop Grumman. Every one of them moved through NASA. 

Procedures came in and went out as controlled materials: red-lined and version-tracked with one organization owning the document and others tolerating the gap, then reconciled in meetings where someone read line numbers aloud. The agency integrating mission-critical hardware worth billions is running a manual-sharing protocol built on PDFs and legacy software because the substrate offers nothing else.

This is what 15 years of federal cloud modernization built, layered on top of a decades-old contractual regime. The U.S. operational cadence across civil space and the defense industrial base will not survive either of them.

The most ambitious attempt came in the wake of Sept. 11, when the failure to connect government knowledge across agency boundaries was identified by the 9/11 Commission as one of several factors that let the plot succeed. The reform agenda that followed was the most sustained federal effort to dismantle institutional fragmentation in a generation, but it also made the fragmentation problem worse.

None of the changes made were wrong. The migration off legacy on-prem systems was overdue. Zero-trust architectures address real and ongoing adversary behavior. The security culture built in the post-9/11 era remains a response to a still-present threat environment. But the architecture that grew up around those legitimate purposes produced a recoil effect — and that shift toward the restriction logic of the security regime is now strangling the missions it was supposed to enable.

How modernization built the walls

Cloud modernization arrived a decade later, sold as the technical layer that would finally dissolve the agency-by-agency data islands that had defined federal IT for 30 years. 

The 2011 Cloud First policy and the FedRAMP authorization model promised shared infrastructure, common platforms, and the end of duplicative procurement. What they delivered was something close to the opposite. 

FedRAMP certifies that a cloud is safe to use; it does not certify that two clouds can talk to each other. The impact-level stratification and the classified enclave architecture above them turned the federal cloud landscape into a marketplace of certified isolation. 

Every program manager’s easiest path to compliance is to stand up a new environment rather than negotiate shared access to someone else’s. We took a human coordination problem and replaced it with a technical and contractual architecture in which sharing is structurally harder than it was before the reform started.

Let’s examine what every serious federal organization built in response. The Department of Defense’s CDAO stood up Advana; Customs and Border Protection stood up RAVEN. Both are  legitimate, well-resourced attempts to solve fragmentation inside one organizational boundary. 

And each is now a new wall — a massive, modern, internally coherent environment that does not interoperate with the ones the agencies next door have built for the same reasons. We are modernizing into a federal landscape of certified isolation, one ambitious enclave at a time.

The wall that survived everything

Underneath all of that sits a regime no reform agenda has been able to untangle. The DFARS data rights clauses were written for a world in which the deliverable was hardware and the IP was in the drawings. They have been carried forward, with marginal updates, into a world where the deliverable is operational capability and the IP is in how the work gets executed. 

Prime contractor A will not put its core procedures into a shared environment where prime contractor B can see them, and vice versa. Every richer form of collaboration creates ambiguity about what was shared, who owns the derivative, and whether a contractor just gave away a trade secret. So the PDF wins, and NASA stays in the middle, hand-carrying the data.

The PDF persists because it is legible. A frozen artifact has an unambiguous boundary: You know what you shared, you know when, and the recipient cannot reach back into your live system. That legibility is the entire reason the substrate still works at all. At the cadence the country is now asking the industrial base to sustain, it cannot.

AI is not going to fix this. A model running against data inside Advana cannot reach data inside RAVEN, and a model trained on procedures inside prime contractor A’s environment cannot see the procedures inside prime contractor B’s. The walls are still walls, even at AI speed, and a model without access to the context on the other side is just a faster way to be wrong inside one boundary.

The Hubble Space Telescope flew in 1990 with a primary mirror ground to the wrong specification because a procedural quality check at the contractor seam never caught that the null corrector had been assembled incorrectly. The fix required three Space Shuttle servicing missions and roughly $1.5 billion in correction costs — almost twice what it should have. 

The country was lucky the failure was on a telescope in low-Earth orbit, where astronauts could reach it. There are no servicing missions to the moon. 

The federal government has spent something on the order of $100 billion over the past 15 years on cloud and IT modernization explicitly intended to dissolve the kinds of fragmentation that produced Hubble’s mirror — and what we have purchased, at that price, is a federal landscape in which the same class of failure is now structurally easier to produce, not harder.

The fix is smaller than the debate

A decade of cloud consolidation, zero-trust architectures, and data-mesh reference models has missed the actual failure point. The failure is one layer up, in the workflow and procedure layer where humans and organizations meet. 

What the fix looks like is actually simple: Let a contractor share the state and interface of a piece of work without sharing its substance — the inputs it needs, the outputs it produces, the holds it has called, the step it is on, with none of the proprietary content underneath. 

Software figured this out in the 2000s when it stopped shipping source code to partners and started shipping endpoints. The procedure equivalent does not yet exist as a default in the government-prime ecosystem. It needs to.

Right now, somewhere in the federal apparatus, an engineer is doing in eight days what should take eight minutes, because the procedure they need to verify is on a system they cannot reach. 

Now multiply that scene across every program office in Huntsville and Seal Beach and Denver and Vandenberg, every contractor seam in every active aerospace and defense program, every day, for the next 20 years. 

And the substrate that made the meeting necessary will still be there the next morning, waiting for Sentinel, waiting for the moon base, waiting for whatever comes after — every program, on both sides of the industrial base, paying the same coordination tax in the same currency, because we built the walls ourselves and we have not yet decided to take them down.

Laura Crabtree is CEO and co-founder of Epsilon3, an enterprise software company providing procedure execution infrastructure for aerospace, defense, and advanced manufacturing. She previously served as a lead operations engineer at SpaceX.