Enterprises in the Middle East and Asia-Pacific have been telling AI vendors the same thing for the past year: they want agentic AI, but they don't want to hand their data, prompts, or operations to somebody else's cloud to get it.
Anurag Gurtu has heard that message enough times to build a company around it. He's the co-founder and CEO of Airrived, a two-year-old infrastructure startup that just launched a Sovereign AI Platform aimed at organizations that want the full agentic AI stack — models, orchestration, GPUs, and data — running entirely inside their own walls. The company is showcasing the platform this week at GISEC Global in Dubai. Gurtu says Airrived already has this running in production at large U.S. and international companies, several major banks, and government agencies — though specific customer names are being withheld pending client legal review.
When I asked him why sovereignty is such a strong pull in that region specifically, he broke it down to three things. First, trust — a lot of large organizations are wary of sending sensitive data to Anthropic or OpenAI. Second, cost — companies running pilots are watching token spend climb unpredictably as they move agentic apps into production, with no ceiling in sight. Third, talent — the skills needed to build and run agentic systems are scarce, and a lot of ambitious AI projects stall out because of it.
Airrived's answer is what Gurtu calls a "box." Customers can order a fully loaded 10U GPU rack and drop it into their own data center, air-gapped if they want it that way. "Nothing leaves. The models are there. The agents are there. The apps are there. The data is there," he said. Once it's installed, token costs effectively disappear, because the compute is owned, not rented by the call. "It doesn't mean they don't have to invest — they do have to buy GPUs," Gurtu said. "But they have to buy it once, and they can use it forever."
That capital outlay is real, and GPU supply is part of the calculation. Gurtu told me the newest chips — Nvidia's B-series and RTX Pro line — are backlogged one to two months right now. H100s and H200s, by contrast, are available immediately, even though he considers them "old" less than two years after launch. A typical order-to-deployment timeline runs about a month to six weeks.
For developers, the more interesting question isn't the hardware. It's what's actually running on top of it. Airrived's platform exposes a range of AI capabilities on a single stack: retrieval-augmented generation, fine-tuning via SFT and LoRA, agent orchestration, reinforcement learning, and deep-reasoning pipeline composition. On top of that sits an app store — Airrived's term for a library of prebuilt, production-grade applications covering security operations, identity management, IT operations, sales, finance, and vulnerability management, among others.
I pushed Gurtu on how this is actually different from a team standing up Llama or Mistral on-prem and wiring it together with LangChain or CrewAI for orchestration and LangGraph for reasoning. He didn't dispute that path exists — he's arguing against the effort it takes. "You can DIY yourself," he said. "You would have to install Llama or Mistral on-prem, get LangChain or CrewAI deployed. If you want deep reasoning, you'll have to deploy LangGraph and code on top of it. And if you want orchestration automation, you'll have to use n8n. Instead of using multiple point-product solutions and stitching them together, you just get a single operating system and you're ready to go." His comparison: installing Linux and discovering you don't even have a GUI yet, versus buying a laptop that's ready to use the moment you open the lid.
To prove the "ready to use" part, Gurtu built an app live on our call. He picked a vulnerability-management agent from Airrived's library, connected it to a sample data source, and named it. A few minutes later, it had ingested 80 vulnerabilities and produced a working dashboard sorted by exploitability and CVE, no code involved.
Airrived's own explanation for how it produces so many of these apps so fast is a Lego analogy. Gurtu estimates 60 to 70 percent of the underlying components — the AI tools and agents — are shared across different apps. "Think of it like we're the factory behind IKEA," he said. "We're producing the pieces. IKEA comes and assembles them into SKUs, which are the apps. We want customers to be able to operate behind IKEA too, so they can create their own pieces." Customers who need something that doesn't exist yet build it themselves inside the same platform, using the same reusable components, rather than reaching for a separate point product.
Sales tend to start with the CISO, a reflection of Gurtu's own two decades in cybersecurity and the security-heavy makeup of the app library. From there, he said, conversations expand toward the CTO, CIO, or chief AI officer as the platform starts running apps for other teams. Airrived's pitch at that stage echoes how enterprises already standardize on infrastructure: one ticketing system, one firewall vendor, one routing vendor. Gurtu's bet is that agentic AI ends up the same way — one platform running every team's agents, rather than a pile of disconnected point products with no shared governance or context between them.
A marketplace where customers can publish and share agents across organizations is planned but not yet live. For now, Airrived's argument to developers is a narrower one: everything you'd otherwise assemble yourself is already built, already integrated, and already running behind your own firewall.
