NewSovereign AI: private AI infrastructure on hardware you rent or own, fully managed.
NewSovereign AI: private AI infrastructure on hardware you rent or own, fully managed.
Nov 10, 2026
Written by: Kelsey Krapf
If your AI can’t run with the network cable pulled, it isn’t sovereign. Here’s how defense and aerospace teams get AI inside the fence, and how to pick the right platform.
Sovereign AI means you own the whole stack: the model, the data, the compute and the rules it runs under. Nobody outside your walls can see it, change it or switch it off. For defense, aerospace and regulated industries, that’s the difference between using AI on real work and keeping AI away from it.
Here’s why. The data that would make AI most useful on a program is usually the data that can’t leave. Think export-controlled drawings, test reports and classified files. Yet a public AI service can’t touch any of it. So teams end up with AI for email and meeting notes, and nothing for the mission.
That’s also why the money is moving. Gartner expects worldwide spending on sovereign cloud infrastructure to reach about $80 billion in 2026, with $16 billion of it in North America (Gartner).
We’ve spent our careers on the other side of that line. Our founders came out of Sandia National Laboratories and Cray Research, where the wrong answer is worse than no answer. So this guide is what we’d tell a program lead who asked how to get AI inside the fence.
A quick note: we sell one of the platforms compared below. That’s why every competitor fact links to the vendor’s own page, checked in November 2026.
The term started as a national idea. NVIDIA, for example, describes a country’s models as “hosted and run on local infrastructure, subject only to local laws” (NVIDIA). That’s the right instinct, but it needs to shrink to the size of a program.
Our co-founder Connor Brown breaks sovereign AI into three legs: data, iron and rules. A deployment is only sovereign when all three hold. On a defense or aerospace program, they look like this:
Yet most vendors only talk about the first leg. Where the data sits matters, but it’s one-third of the problem. If the software has to phone home for a license check, you don’t control the iron. And if someone outside your program can administer the system, you don’t control the rules either.
Regulated teams have plenty of ideas for AI. The hard part is finding a place to run it, on the data that matters. That gap shows up in three ways.
Much of the technical data on a defense or aerospace program is export-controlled. Export rules limit who can access that data and where it can go. In fact, sharing it with a foreign person can count as an export even inside the United States.
That’s where hosted AI gets complicated. A model can’t reason about a prompt it can’t read, so a typical hosted service decrypts your data on someone else’s servers, run by someone else’s staff. As a result, many export-control policies keep controlled data away from public AI tools entirely.
Sovereign AI changes the question. When the model runs inside your boundary, on hardware you control, controlled data never has to leave it. Hoonify’s Sovereign AI and Sovereign HPC are built to support export control requirements, with role-based access by user and project and audit trails for every job. Your export-control officer still has the final say on what data goes where. But now the system makes that decision easy to enforce.
Closed AI tools need the internet, so on a classified or controlled network, they don’t exist. As a result, engineers on those networks either work without AI or start wondering whether a paragraph or two really counts. A ban without an alternative creates the very risk it was meant to prevent.
Instead, the fix is to put the model where the data already lives. When open models like GLM and Qwen run inside the network, engineers stop reaching for public tools, because they no longer need them. And if you’re wondering whether open models are good enough, Connor has a short answer. In his words, for data that can’t leave your boundary, “the real comparison was never open versus frontier. It’s open versus nothing.”
Anything that needs an outside connection stops when the connection does. A system that runs offline, however, keeps going when public clouds or outside networks fail. That’s also why we design for uptime inside the building. For DeliverFund, a nonprofit intelligence team you’ll meet below, we built a multi-site, high-availability setup, so if one site goes down, the work moves and keeps running.
The same logic holds well beyond defense. Hospitals, banks and law firms also hold data that can’t cross a trust boundary. For them, too, AI data sovereignty decides whether AI ever touches their most valuable work.
In our experience, most programs don’t fail on hardware. They fail in the long stretch between the purchase order and the first useful job, when nothing can come in from the internet and every hidden dependency breaks.
So we use a simple check we call the unplug test. Pull the network cable, then see what still works. If a platform passes these seven points, it’s ready for a real air gap:
Want to go deeper before you talk to anyone? Download our Sovereign AI Guide.
Not every workload needs an air gap. Public documentation, internal tools and code that carries no controlled data can run in a cloud with zero data retention. In fact, that’s step one of how we work with every team.
They prove the use case in our cloud first, with work that isn’t export-controlled. Then, once they know which models fit which jobs, they move the sensitive work inside their own walls. So let the data decide where each workload runs. If it can’t leave, it needs sovereign AI. If it can, start simpler.
Most sovereign AI platforms follow one of three approaches. The right one depends on how fast you need to move, and on who will run it once it’s live. Here’s how they compare.
| What matters | Public cloud AI | Build it yourself | Fully managed on your hardware |
|---|---|---|---|
| Time to running | Minutes, on their servers. | 6 to 12 months. | Same day in our cloud, weeks on yours. |
| Where your data lives | On their servers. | Inside your walls. | Our secure cloud, or inside your walls. |
| Air-gapped option | No. | Only after you build it. | Yes, from cloud to air-gapped. |
| Runs AI and simulation | AI only. | Two separate builds, two teams. | Yes, so one team runs both. |
| A model you own | Not an option. | Yes, but you tune and maintain it. | Yes, tuned on your data. |
| Expertise you need in-house | Some, plus a support ticket queue. | An HPC and MLOps team. | None, because it’s our team. |
Public cloud AI is the fastest way to start, and it’s fine for work that carries no controlled data. Some services also offer government editions for agencies and qualified contractors. Even so, the model, the servers and the roadmap belong to someone else. So for export-controlled or classified work, this is usually where the road ends.
Building your own stack gives you full control. You buy the GPUs, choose and serve the models, wire in security and run it all. But you also staff all of it, which usually means hiring HPC and MLOps engineers the market barely has. By our estimate, that’s 6 to 12 months before the first useful job, while your program waits.
This is the approach we built Hoonify around. You get the control of building it yourself, without the build. Our national lab engineers deploy, run and support the system on hardware you rent or own, on-prem or fully air-gapped. And because you can start in our cloud the same day, the team proves the use case before any hardware arrives.
Want vendor-by-vendor detail? Our comparisons of managed on-prem LLM companies and air-gapped private AI infrastructure name names.
We built Hoonify for the gap between buying hardware and getting work done. Everything we deliver runs in classified or disconnected environments, with no cloud dependencies. Our national lab engineers deploy, configure and run it (Hoonify), so your team never has to hire specialists to keep it going.
Hoonify sells two products, Sovereign AI and Sovereign HPC. You can buy either one or both, and the same team runs whatever you choose.
Sovereign AI is private inference. Open-weight models like GLM, Qwen, Gemma and Inkling Small run on your network, behind a standard OpenAI-compatible API. Engineers use it as a private assistant, to search program documents and to fine-tune a model on their own data. Because it keeps the same models and API wherever it runs, moving from our cloud to your facility costs no rework.
Sovereign HPC, on the other hand, runs on TurbOS, our HPC software. TurbOS stands up a full HPC environment in about an hour, tying CPUs and GPUs into one system for modeling and simulation. Every deployment also uses the same validated image, so a simulation in one facility matches a simulation in another. TurbOS has run in production for five years.
They’re separate products, but defense programs rarely need just one. The same teams that want a private AI assistant also run:
When one team runs both, new workflows open up. For example, an assistant can search simulation results right next to the test reports they predicted.
DeliverFund’s story is a Sovereign HPC one, but it shows how we deliver either product on a customer’s own hardware. DeliverFund is a nonprofit whose analysts build the intelligence that more than 500 law enforcement agencies use against human trafficking. Its data is evidence of crimes against real people, so it belongs on hardware DeliverFund controls. That’s why TurbOS now runs its collection on machines it already owned, fully on-prem.
Collection runs five times faster, and DeliverFund went from 400 intelligence reports a year to more than 600. Most of all, the analysts got their time back (case study):
“We can reduce our initial analyst load by more than 90%. Our future is all the brighter thanks to Hoonify’s involvement.”
Sean Fennema, President, DeliverFund
Defense and aerospace teams work under the same constraint: more data than people, and data too sensitive for someone else’s cloud. Today, our customers include Anduril, Los Alamos National Laboratory and Sandia National Laboratories.
Whichever product you start with, Connor puts it best: the air gap changes your operations, not your architecture.
Every month spent building is a month your program waits. That’s why teams can start in our cloud the same day, prove the use case, then move the same models and code onto their own hardware in weeks. By our estimate, building it in-house takes 6 to 12 months. Since only the location changes, nothing has to be rebuilt.
Behind all of it is a team with more than 100 years of combined national lab experience.
Sovereign AI means you own the whole stack: the model, the data, the compute and the rules it runs under. Nobody outside your walls can see it, change it or switch it off.
Regulated teams hold data that can’t leave their control, such as export-controlled technical data, classified files and patient records. Without sovereign AI, that data stays away from AI entirely. With it, AI can work on the data that matters most.
It depends on how fast you need to move and who will run it. Public cloud AI is fastest, but your data lives on someone else’s servers. Building it yourself gives you full control, but it takes 6 to 12 months and a dedicated team. A fully managed platform like Hoonify, by contrast, runs on your own hardware, operated by our engineers, and you can start in our cloud the same day.
Yes, if the platform is built for it. Hoonify runs fully air-gapped, with no external calls or telemetry. Before you buy any platform, though, run the unplug test: pull the network cable and see what still works.
Not always. If you build it yourself, you’ll need HPC and MLOps engineers. With a fully managed service like Hoonify, however, our engineers deploy and run the system on your hardware.
Ready to bring AI inside the fence? Talk to an expert about sovereign AI for your program. Still researching? Get the Sovereign AI Guide first.
You bring the mission. We handle the rest.
Hoonify runs AI and HPC for organizations that can’t afford to get it wrong. Our founders come from Cray and Sandia National Laboratories, and our team brings more than 100 years of national lab engineering to every system we build. We learned our standards in places where a wrong answer is worse than no answer: national labs, defense programs, energy research, and life sciences. For five years, we’ve run TurbOS, our HPC software, in production for customers including Los Alamos and Sandia National Laboratories. Now we bring that same discipline to AI. Customers choose Sovereign AI, Sovereign HPC, or both, fully managed by our engineers. They can start in our cloud and move on-prem or air-gapped when they’re ready, with open models, standard APIs, and no lock-in along the way. You keep the data, the hardware, and the roadmap. We keep it running.
